It was necessary for my Cici series. You may have noticed that many of the articles pinged as partially AI-created with Substack’s new feature. Cici is working behind the scenes, but what it creates is imperfect. The Cici version of yesterday’s article was less than half the size of the finished product. Other articles suffered from pervasive LLM-speak.

While Cici reduces the time required, there is still a significant amount of time required to create each article. Why am I writing so much here? As I mentioned in my last article, I have a business problem. My reach on LinkedIn has dropped, and course sales have come down with it.

I need a new workflow to offset the lost revenue. The best-performing content channel is now Substack, but that new workflow (creating articles vs short-form LinkedIn posts) is more time-consuming. If it weren’t for Cici, the new workflow would be infeasible. This is the way we must position every agent in the enterprise because it aligns technology with value creation and revenue generation.

We need obvious and significant value, especially if we are early in the maturity journey and most parts of the business are still skeptical or unwilling to transform. This is how we build a track record of success, narrative, and the coalitions that can move a business.

When Agents Are Mostly Good, They Are Still Very Bad

This is a good example of the AI Last Mile Problem.

The comment is 80% good, but the 20% bad creates significant loss in a production environment. First, it is obviously AI-generated, which immediately damages the poster’s perception with readers. Second, it is the kind of ‘right’ that LLMs are known for. In another context, this would be an extremely relevant point to raise.

In this context, it makes the poster appear technically inept. The taxonomy describes document formats. Automating document change tracking and formatting updates is a solved problem, especially in finance and construction domains. That last 20% can damage the reputation these comments are meant to enhance.

Content-to-cash is a long-chain workflow that starts with content, and AI generating content is synonymous with slop. The first problem I need to solve is getting my agent to create content that sounds like me, or I’ll sound like the commenter above and incur the same losses. My audience here is far less forgiving than the average LinkedIn user.

I detailed an approach in a prior article that works fairly well in terms of writing style. It gets 80% right, but the problem is that the content it writes still sounds like an LLM wrote it. It is much better, but not quite right. How do we get the other 20%? The AI Last Mile Problem shows that getting the last 20% is harder than getting the first 80%, which is why AI and agents demo so well, then fail to reach production.

Again, we see that business architecture informs, optimizes, or constrains our agentic architecture.

A Bridge That Crosses The AI Last Mile

The LLM needs more granular instructions. I can keep feeding articles into the model and build increasingly granular style guides, but if 700 articles isn’t enough… How many more articles do I need to write? Luckily, there’s another way to approach this challenge.

Read more

In a time of unprecedented opportunity, the definition of full-stack engineering now includes the business. No one who can substantively advance your career wants to talk to you if you don’t understand business architecture. Customers and clients don’t want to talk to you unless you understand value engineering. VCs won’t fund you until you can speak to both.

The Principles Of Business Architecture

I see a lot of businesses that did multiple things right with information, AI, and agents, but revenue is still falling. They built the platform, shipped agents, and customer engagement is trending up and to the right. But when the CFO asks why none of it shows up in top-line growth, my phone rings, and it’s the start of a challenging conversation about business architecture.

I have some version of it with prospective clients multiple times every quarter, going back to long before the ChatGPT moment. I must explain the Orchestration Imperative and the Big Picture. AI, information, and agents (like data and ML before them) are extensions of the business and operating models in ways that prior technologies weren’t.

Every CEO says that technology is a core pillar of strategy, but when revenue doesn’t show up…is it really? If technology or agents or AI is a core pillar of strategy…where is it?

Most leadership teams can show me a business and operating model. When I ask, “Where is your Technology Model?” typically, I get a list of initiatives, a hiring plan, and the budget for technology investments. Those are tactics. What strategy informed those tactics?

AI strategy should help the business see the Big Picture, but most AI strategies are aspirational, not actionable.

AI strategy should help the business make 3 critical decisions that align technical maturity with opportunities, value creation, and business transformation. Most AI strategies provide cover to delay those decisions with a flurry of interesting-sounding activity.

They overfit to technology investment, which creates multiple problems that don’t become obvious for several quarters after the decisions that lead to them. Without the Big Picture and 3 Critical Decisions, that distance makes finding and addressing the root causes challenging.

Asking New Questions So Business Architecture Guides Technical Architecture

• How do opportunities move parts of the business and operating model into the technology model?

• What new parts of the business and operating model is the business building for the first time in the technology model?

• How does the business monetize those transformations?

• How does it keep them aligned with core strategic goals, not technical goals?

These are core tenets of Top-Down Opportunity Discovery. An actionable AI strategy manages a lot of complex orchestration across the enterprise. Handwaving, slide decks, and playbooks have never been enough, but we are finally seeing the weaknesses exposed in quantifiable ways.

Yesterday I explained what happens when you get the architecture right. In this article, I will explain what happens when you get the architecture right, but the business lags behind it. Agentic architecture is business architecture, and vice versa.

Baidu Vs Google: The Story Of 2 AI Strategies, But Only 1 Includes Monetization

Baidu just had this conversation publicly during its most recent earnings call, and it wasn’t pretty. It didn’t have answers to those questions, and doesn’t seem to have the frameworks or methods to get them. Its AI business grew 25% and now accounts for half of Baidu’s core revenue. GPU cloud grew 283%. Total revenue fell 4%, and that’s the fifth straight quarterly decline.

Google faced the same disruption with the same technology and grew 24%. Investors are making comparisons in this market, and this is your opportunity to provide clarity to the C-Suite.

I understand many will be quick to point out that Baidu operates in China while Google is in a very different market. Alibaba’s results today prove that the China variable isn’t a factor. Strategy is. Its revenue is up by 9% for the same quarter that Baidu just reported.

Yes, there are significant differences between the two companies’ business models and where they derive most of their revenue from. As I’ll explain in a minute, that’s my point. It’s not enough to just be ready for the technology. Businesses must also be ready for the new monetization paradigms as well. Most aren’t.

In this article, I’ll explain what Baidu got right, where the monetization strategy broke, and how you can tell whether your own transformation is heading in the same direction. The pattern generalizes beyond search. I’ve watched it repeat in insurance, finance, retail, industrial equipment, and enterprise software. The leading indicators are the same every time. The gap and steps to resolve it also follow a formula.

You can transform your technology model, but leave your business model behind, and it’s all for nothing. That’s what Baidu did. Technology can do everything right, and the business still fails to generate the growth it should.

Baidu Did The Hard Part Right

I want to start here, because the simple analysis of Baidu’s results is wrong, and you’ll take the wrong lesson from it. Baidu climbed the capability maturity model in the right order and executed well.

Their GPU cloud has delivered four consecutive quarters of triple-digit growth. Management said that demand is broadening across gaming, autonomous driving, smartphones, and financial services rather than concentrating in one or two customers. Concentrated demand proves risk, rather than revealing a market. That’s why Microsoft went out of its way to make the same point.

Their model-as-a-service platform grew token revenue from external customers more than 9X. I don’t love the token-based pricing model, but it’s hard not to love that kind of growth story. Obviously, the demand is there.

Their robotaxi business delivered a million driverless rides in the quarter. They were granted the first fully driverless permits in a right-hand-drive market, started open road testing in London, and launched in Dubai. Baidu’s leadership made a solid unit economics case for international expansion. Higher taxi fares overseas produce better margins than in the Chinese market. (This is the one place that business, operating, and technology models align, but it appears accidental vs intentional.)

They’re also managing costs better than Google. Baidu reported a 12% non-GAAP operating margin and positive operating cash flow for a fourth straight quarter, even while the top line shrank. Google’s free cash flow, on the other hand, went negative for the first time in the company’s post IPO history.

This isn’t a case of a company that failed to understand AI, information, and agents. Baidu figured the technology out early and built more of the stack than most of its competitors did. The problem starts one layer above the technology: monetization.

They’re Paying The Innovation Tax With Shrinking Growth

Read more

There are two business roles in the agentic commerce economy. You’re either the agent that reads customer intent, or you’re a service that the agent calls. Even if your ‘service’ makes and delivers the product, to the agent it’s just a service that can serve its user’s intent.

The intent-facing agent has pricing power, while the ‘service’ competes on price and gets swapped out the minute that something cheaper or that ships faster shows up. Most businesses will fall into the service role unintentionally. They’ll wait for the market to settle, and while they wait, someone else will build the customer-facing intent surface.

Once you’re behind that surface, it’s like being on social media. The platform can crush your reach at any time for any reason. You don’t own the connection to your customers unless they come to the intent surface and specifically ask for you. Most don’t.

Your agentic architecture enables or limits your business’s future. It is an instrument of strategy. It is, in large part, determined by the opportunities your business’s leadership chooses to go after. Technology and the business’s growth or decline have never been so interdependent.

In this article, I’ll explain how the companies building the intent or action surface are paying for it, why the smart ones are giving it away, and what you need to do in your category or domain before ‘the road,’ as Alipay calls it, gets finished.

I’ll use Alipay as the example because they published a clear vision for what’s next, and they have tested it for the last couple of months. The approach generalizes, but agentic commerce will be one of the first to see it delivered. Their approach also follows my frameworks for agentic platform monetization, so to offset my biases, I wrap up this post with a list of risks and what could still go wrong.

Han Xinyi, Ant’s CEO, said, “Every technological revolution builds the roads and facilities of its era, and Alipay’s role is to pave that road.” This is a new picks-and-shovels play at the next layer down. We saw it work with compute hardware, then models, and now we’re seeing it emerge at what used to be the app layer. But what Alipay is building is very different from the traditional app layer.

Someone Is Building ‘The Road’ In Your Category & It’s More Than A Surface

Read more

‘Everyone’ (my overly conservative strawman) on social media says it’s impossible to know what’s next for AI. SAP’s CEO opened 2026 with that message in Fortune, in a market where the largest supplier of GPUs has already reserved its manufacturing capacity through 2027. The market is providing strong signals, but they aren’t strong enough for everyone.

In November 2024, Reuters and TechCrunch ran the same story. Scaling had hit a wall, the labs were out of ideas, and the buildout was about to stall. Ilya Sutskever said everyone was hunting for the next thing. Marc Andreessen said models were converging on a ceiling.

Then the labs pushed ahead with better agentic architecture combined with models that were trained to use it. The technology is providing strong signals, but they aren’t strong enough for everyone.

Evidence becomes proof, but not everyone has the same threshold. I am at mine even though there’s still plenty of risk, and there will be several unexpected developments over the next two years. This article is as close to an evidentiary case as we have right now for what comes next.

As the joke goes, there are two types of people in this world. Those who can work with incomplete information. Which one are you?

The next 10 years belong to people who don’t wait for certainty to form conviction, decide, and act. They pet the cat and hope for the best.

Credible Voices Provide Answers

Jensen Huang steps in and explains what’s next. NVIDIA has transparency into the full AI stack, so it makes sense that his insights would be worth listening to. He has a strong track record of success to back it up.

Transparency is not metaphorical. NVIDIA holds more than half of TSMC’s CoWoS advanced packaging capacity for 2026, with an estimated 800,000 to 850,000 wafers reserved through 2027. It sees hyperscaler purchase orders, neocloud contracts, sovereign programs, and model lab roadmaps long before any of it reaches the general public. When Huang gives a number, he is reading it from a highly accurate model of demand.

In May 2023, he guided NVIDIA’s next quarter to $11 billion against a $7.2 billion consensus. The stock rose 25% the following day, and Susquehanna said that the new gold rush was on.

In August 2024, reports said Blackwell was slipping. By November 2025, Huang told investors that Blackwell sales were off the charts and cloud GPUs were sold out.

In January 2025, NVIDIA fell 17% and lost $589 billion of market value in a day, the largest single-day loss in US market history, on the thesis that smaller models would lead to less compute demand. Huang said the opposite. Reasoning models would need 100X more compute. NVIDIA closed fiscal 2026 with $215.9 billion in revenue, up 65% YoY.

Huang raised NVIDIA’s guidance. At GTC in October 2025, he said he had visibility into $500 billion of Blackwell and Rubin bookings through 2026. Five months later, he said, “right here where I stand, I see through 2027, at least $1 trillion.” The first quarter of fiscal 2027 came in at $81.6 billion, up 85%, with data center revenue delivering $75.2 billion, up 92%. He guided the following quarter to $91 billion.

The Google View Of Ads & Agentic Commerce

Read more

I get called in to clean up the mess, so it is a line item on V Squared’s balance sheet too. I won’t throw stones without getting hit myself. I profit from business dysfunction, but I’d rather this get addressed internally. It is the one engagement category that frustrates me the most.

Weaponized incompetence should end an executive’s career and destroy the reputation of any consulting company that enables it. In reality, both often profit from it. Parts of my field are shady, but they don’t need to be. We should monetize frameworks that accelerate revenue, so every business gets the most efficient path to growth. But as I have said for over a decade, there’s just as much money in perpetuating problems, and it’s much easier to deliver.

In this article, I will reveal the business cycle of weaponized incompetence and explain how to defuse people and consultants who are trying to run the game on you.

How The Game Is Played: The Crisis & Speech

The speech. “These AI initiatives and agentic spending are out of control. I’m stepping in to ensure we start seeing significant ROI for all the money we’re investing in these programs.”

OR

“The future of this company depends on agents and AI. I’m stepping in to ensure we get there.”

The speech is always given by a pseudo-technical executive leader with credentials that are typically bigger than the outcomes and expertise they promise. They say things that sound amazing and make promises to end the chaos or crisis that everyone is really tired of.

• Spending is up, but value hasn’t materialized, so margins are compressing.

• The core business is suffering from all the contradictory priorities, and revenue is flat.

• The business is tired of the endless pivots and new directives. People are leaving.

• Infighting has gotten out of hand, and business politics have halted progress.

Something happened that investors or the board got wind of. The CEO decided, or was convinced, that something must be done about the something happening. Enter the executive giving the speech. The root cause of the crisis is often manufactured to support a narrative. The speech is the version of the narrative that puts the executive in control of AI and its sizeable budget.

In just 12 short months, the business will be in a better place.

The Listening Tour & The Plan Becomes Your Problem

They spend the next few weeks pulling everyone who has ever said the letters ‘AI’ into a meeting to get their ideas. Thoughts are shared. Grievances are aired, and promises are made to address them. Many questions are asked. Someone takes a lot of notes.

Unless you have a C-level title, everything from that meeting is thrown out. The direction has already been decided, but the listening tour is critical so the executive can say, “We listened and incorporated what we heard into the tentative plan.” Except nothing you or your team asked for ends up in the plan. None of the risks you called out are identified and accounted for.

“Don’t worry. This is just the tentative plan, and we’re still making changes. Those things will be in the final version.”

Three months later, a consultant or small team shows up. The consultant is there to review everything and ensure nothing has been missed. Three months later, the plan gets their seal of approval. When you see the plan, it is exactly the same as the tentative plan, with more words and charts. There are also several case studies supporting it and multiple decks ready to be distributed to every part of the business.

Then you get to the end and read the first round of initiatives. To your horror, you realize that none of it can be built, and your name or your team’s name is in bold letters. You’re responsible for building it all.

Weaponized Incompetence & The Blame Game

The technology isn’t ready to support what they’ve outlined. Agents don’t work that way. The timelines are all too short. There’s no budget for infrastructure. Your people are all fully committed to other critical projects, and there’s no budget for hiring. This will fail for a dozen different reasons.

You go to the executive to raise your concerns, and they schedule a meeting to discuss it. Then they move the meeting because of a scheduling conflict. Then they move the meeting again due to someone critical being on vacation. They keep moving it until the day your team starts working on it is less than a quarter away. Then the meeting happens.

“Why didn’t you bring this up during our initial meeting 9 months ago?” Well, I didn’t realize you were incompetent.

“Why didn’t you bring this up sooner? The project is about to start!? How are we going to justify pushing it out?” What’s this ‘we’ thing you have suddenly added to your vocabulary?

“Don’t worry. I’ll figure out how to smooth this over. Thank you so much for bringing it all to our attention.” They push the timelines so more meetings can be held to discuss your concerns.

But at the C-level, they are saying something different. “This team had 9 months to prepare or raise these concerns and didn’t. I am worried that they might become a risk, but we will do everything we can to get them up to speed.”

The consultant comes back into the meetings. When you raise a risk, they respond with, “That hasn’t been an issue in any of our other deployments.”

When you ask for more people, they say, “None of our other deployments have required additional headcount.”

When you explain the technical limitations, they say, “Our technical teams were able to work around that.”

The executive stays out of the discussion and lets the consultant do the conflict management and deflection. The executive sides with the consultant, and the roadmap is finalized with minor changes.

At the C-level, they say, “We listened to and addressed every one of their concerns. It turned out to be nothing. I hope this is the end of this roadblock.”

The Endgame: You’re The Problem

Now, if you call out the roadmap as technically infeasible, you’re the problem. If you try to deliver what they asked for and it doesn’t work, even though you warned them it wouldn’t, you’re the problem. After they have figured out a way to make you the problem, the consultant recommends bringing in their team to clean things up.

CFO: “Where will we get the money to do that?”

Consultant: “If the external team is doing the work, we can do a RIF on your existing team to offset most of the cost.”

CEO, who believes your team is holding things up…”That sounds like a good plan. Draw it up.”

Executive: “I can tuck the smaller team into my org and manage the external team as well. That will save even more and speed things up to make up for lost time.”

CEO: “Great. Make it happen.”

Weaponized incompetence is focused on controlling the narrative, not delivering solutions. That’s why it is so effective. C-level leaders in mid-sized to large businesses don’t get enough direct feedback from the front line. They are disconnected from reality on the ground, and most surveys prove that their perception of what’s happening is often very different from what people on the ground see happening.

That’s the power of narrative control. Reality doesn’t matter when you control the narrative that reaches C-level leaders. Reality is whatever you want it to be when you control the context around the evidence that C-level leaders see.

Why They Play The Game

Money and promotions. The consulting firm gets to widen its engagement by taking over projects that used to be done by internal teams. If they play the game right, the business will become dependent on the consulting firm. It will hand over its roadmap and most of its highest-value R&D initiatives along with it.

The executive takes on a critical mandate to deploy AI and agents…promotion and raise 1. They remove all the barriers and blockers, and streamline the organization to reduce the cost of delivering all those agents…promotion and raise 2. They need a larger mandate (read: promotion and raise 3) and more headcount to deliver what the business needs. They run a much bigger organization now, so they need a promotion and raise (4) to reflect it. That’s 4 significant promotions in less than 5 years.

Based on their rapid rise through the ranks, they receive an opportunity they cannot pass up at a new company. This becomes raise 5 when your company counteroffers because they have invested so much in this AI journey, and the executive is the only one who seems to be making any progress.

Or the executive realizes they can’t keep running the game much longer, and they take the promotion and raise at the new company. And you’ll never guess who goes with them…the consultant. I have heard of people running this game for multiple decades until someone gives them a C-level role and they get exposed.

And you’ll never guess where they end up after that…the consulting firm. “After a successful run as CxO of {Impressive Company}, we were able to poach them and bring their expertise in-house.”

How To Beat Them At Their Own Game

I have been brought in to fix 3 of these in just the last 9 months. It makes me mad when I realize that all the people the business needs to turn this thing around were laid off by the executive. The budget the business needed to get back in the race went to the consulting firm. Here they are, having to pay twice to get the result they were promised last time.

This game does so much damage to careers and businesses. There are other ways to get ahead and land clients. So let me explain how to beat this racket. It all comes down to narrative control and avoiding ‘Superhero Syndrome’.

The game relies on really talented people seeing the problems coming and trying to prevent the worst from coming to pass. It depends on everyone being a team player who is willing to jump in and help out. That is ‘Superhero Syndrome’. You put on the red cape and fly in to save the day because that’s the right thing to do.

It’s not just happening to you or engineering teams in general. It’s happening to sales, marketing, operations, finance, HR, and most other parts of the business. Every organization has people who try to be good team players and be proactive about filling gaps in the plan.

That’s what the executive is counting on, so that’s where we begin to defuse their plan. Don’t save people who aren’t worth saving. Let incompetent people fail.

Narrative Control

I teach narrative control in my course about executive presence and C-level influence. (Like I said, it’s a line item on my balance sheet too.) The executive and consultant depend on no one else understanding how narratives are built and how to disrupt them.

When you hear the speech, get with your executive or C-level leader and plant the seed. “It doesn’t look like {the executive} has ever done anything like this before. I hope they have the right people backing them up, not a consultant looking to make their quarter.”

Planting the seed will either tell your executive that you are also wise to the game and they can loop you into their response, or it might be the early warning your executive needs to get in front of this. They are going to try to paint you, your team, or your organization as the incompetent one at some point in the near future. Get there first. Control the flow of evidence to your executive or C-level leader and provide the context that diffuses the narrative they will try to run.

Step 2 starts during those early listening sessions. Ask for a list of questions and discussion points upfront. Prepare your response and deliver it to your executive before the meeting. Compare notes across your organization and align your messaging from the top down. Have everyone provide their responses to your executive or C-level leader, so there’s only one meeting necessary…the one with your C-level leader.

The narrative depends on you contradicting other teams or your executive/C-level leader. Even the smallest gaps will be used to show that your team isn’t on the same page as everyone else. Alignment and consensus give a capable executive or C-level leader everything they need to give you cover.

What If You Don’t Have Cover?

But not everyone has that type of executive or C-level leader. Be prepared to follow up if you don’t. Create an email chain where you reshare your written responses and ask the executive to confirm that each point and recommendation has been added to the plan.

“If these concerns are not addressed, we cannot guarantee delivery, but we are willing to revisit this and help you revise your plan at each phase of implementation. Please share the plan as soon as possible, even if it’s just a draft.”

They will probably go silent and not respond, but it doesn’t matter. When they spring the impossible plan with your name all over it, reply to the earlier email chain with extremely brief feedback.

“Why were our concerns not addressed in the plan prior to assigning us the deliverables? If you had scheduled a meeting with us early on, we could have avoided all this rework.”

Provide no detail on what the plan should be or how to fix things. Talk with other teams to see if the executive is playing the same game with multiple teams and orgs. Build a coalition of teams who are all in the same boat. 3 is good enough, but hopefully you can get 5+ teams to join.

If they try to schedule a meeting with you or any of the others, ask for the questions and agenda upfront. As soon as you get it, decline the meeting. Get everyone going through the same thing to create a document responding to the questions. Have everyone share it with their executive or C-level leader. One of them will know what to do, even if yours doesn’t.

If the executive tries to schedule a meeting…

”Since so many of us are going through the same challenges with your plan, it would be more efficient for you to aggregate the feedback we have shared with leadership and come back to us with an updated plan for review.”

The Endgame: A Seat At The Table

This is where weaponized incompetence falls apart. When 5 teams are isolated, they are easy to dismiss individually. When 5 teams form a coalition, they are difficult to dismiss.

The executive doesn’t know what to do or how to fix the plan. They never expected to do anything other than narrative control, but the coalition has an equally credible narrative with a paper trail to back it up. The executive has promised big results, but the only path available is to “manage expectations in light of new information.” The executive must come to the table and negotiate a viable plan.

But resist the urge to give them the plan. Give them feedback about risks, gaps, and reality with 0 context about what to do about it. The second you make a prescription about what to do, they win. It either becomes their idea when it works, or it’s your fault when it doesn’t. If they press you for specifics, “We would be happy to take this over from you. That seems like the most efficient way forward.”

You’ll never get full control, but the executive will realize that the only way forward is to give you a seat at the table and a voice in the process. That’s the goal. Only after you get that seat do you save the day. You end up with a few new meetings each quarter, but you and your team keep your jobs. If you play your hand well, you might even get some credit for doing all this.

The coalition you build is the bigger win. You are no longer a tactical operator working anonymously in tech. The coalition proves you can operate across teams and have business-level impacts. That will open a lot of doors for you, and the higher visibility makes you harder to replace. Narrative building and control are C-level capabilities, and you’ll be seen in a new light. That’s when new opportunities begin to come to you, rather than you chasing every promotion and new role.

• Marketing → marketing analytics → marketing engineering

• Growth → growth analytics → growth engineering

• Revenue operations → revenue analytics → revenue engineering

• Finance → financial analysis → financial engineering

• Operations → operations analysis → operations engineering

• Product → product analytics → product engineering

• Research → research analysis → research engineering

• Content → content analytics → content engineering

• Sales operations → sales analytics → revenue/sales systems engineering

• Customer success → customer analytics → customer success engineering

In my seminar, I have a slide with 30 roles that have gone from nontechnical to analyst to engineer. The highest demand and salaries are given to people in the third category who have domain expertise, the ability to work with data, and engineering capabilities.

Strategy engineers are coming soon. If you think about it, that’s what C-level leaders do too, and they’re about to get support from emerging strategy engineers. It’s becoming the entry point into C-level leadership and becoming a founder.

The Emergence Of Outcome Engineering Roles

A strategy engineer is a business scientist who helps build an algorithmic business through experimentation. Experiments transform the operating model one causal workflow at a time. They reveal entirely new business models with higher growth ceilings and faster growth rates.

Outcomes force a more rigorous methodology. That’s why F1 racing has some of the most reliable models. They see the results every race. It either works or it doesn’t, and those feedback loops build highly reliable models. It is the same in algorithmic trading, healthcare, and pharmaceuticals. You see a focus on outcomes in any industry that can’t spin the results, and that clear feedback is fuel for reliable models.

• The patient got better, or they didn’t.

• The trade made money, or it didn’t.

• The drug was effective and made it to market, or it didn’t.

AI scales access to information, which makes complex outcomes easier to track. The app made money, or it didn’t. The ROI for code is easier to see. The business’s growth accelerated, or it didn’t. The ROI of strategy used to be opaque, but not anymore.

We’re all becoming outcomes engineers helping businesses compete in an outcomes economy. People with domain expertise, the ability to work with data, and engineering capabilities can run experiments that accelerate information flywheels. In part 3 of this series, I will provide an example of what those extreme experiments look like.

But more importantly, I’ll explain the vanilla experiments that pave the way.

Our Big Ugly ROI Problem: Value Streams Are Big

I’ll start with value and work my way into architecture from here. Every engineering practice, design pattern, and architectural tenet must be financially viable. ROI is just as critical a constraint as any technical one.

As I explained in the last articles of the Causal Workflow Series, the recruiting and hiring workflow doesn’t create any value for the business until the employee delivers value to the business. You’d be surprised by how many jobs and workflows are like this. They don’t create value until after the end of the workflow.

Software engineering has a similar ugly ROI problem. No value is created by code, finished apps, or platforms until a customer pays for it.

No value is created by building and launching an ad campaign until a customer buys.

A hire, app, ad campaign, and even an AI strategy have potential value. The value is only realized after the workflow ends and the roles that own the workflow hand their artifact over. The workflow can succeed locally, and yet the global outcome will not be delivered.

This is a reality that most business units face. Their workflow ends before value creation begins, but organizational leaders don’t want to admit it because that structure makes them look like a cost center. Step 1 in fixing our big, ugly ROI problem is admitting we have one. That’s where the KPI game starts and why most businesses have low KPI maturity.

Most workflows only support a segment of the total value stream, so let’s start there and use software quality as an example. What is the ROI of a defect? We know that if all the defects shipped, no one would buy the product. We don’t really know that, so it’s a hypothesis with strong expert sentiment backing it up.

Early Enterprise Experiments & Laughable Causal Models

Read more

A $12 billion valuation for a holding company full of accounting firms sounds ridiculous. Most will explain it as more evidence that private equity is excited, maybe a little overexcited, about AI. That view is reductive. I work with a few PE firms (though not any I mention in this article…NDAs), and they are running my frameworks to transform their portfolio of companies.

In this article, I’ll use what many PE firms have made public to dance around my NDAs just a little. Won’t my clients mind? Not really. They want it to leak that their thesis from about 3 years ago is what everyone else is rushing into and calling innovative today. It’s good for attracting smart capital.

Thrive is trying to prove a very specific acquisition model, and many other PE firms are putting money to work behind the same thesis. But how do you discover a thesis early?

It’s Always A Question Of Scale

Every technology paradigm scales access to something. Define that, and you find multiple pots of gold.

Read more

The goal is to transform the operating model, but the work starts at a much earlier maturity phase: transforming the workflow. Technology enables the transformation. It makes it feasible. However, the business is the thing that’s transforming, so technology-enabled transformation must be driven by them and their needs.

This is the SaaS conundrum with AI and agents. Neither produces enough value to justify their costs without reorchestrating the workflow. However, SaaS vendors don’t control their customers’ workflows, so they have to build solutions that support the current workflows, or customers won’t keep signing contracts.

SaaS vendors are hoping that FDEs are the solution, but few have built an FDE force that fully understands the assignment. It’s not enough to get buy-in for point solutions and pilots. The ROI just isn’t high enough to hold the C-suite’s interest. FDEs must get multiple business units to buy in on workflow reorchestration, and that is a very different skill set.

Meet The Business Where It Is Or Technology Has No Chance

The purpose of a workflow is what it does or produces, not what people think it should do or produce. The workflow’s true function is defined by its actual observed outcomes. Stated intentions, goals, and mission statements are meaningless. Meet the business and its customers where they are, not where you think they should be or even where they claim to be.

But that doesn’t mean you must settle for workflows to stay the way they are. The power and value of strategists is our ability to transform workflows or reorchestrate them to produce more value and even new types of value. But we don’t own the process, which means we must lead without authority and wield influence to deliver results.

I have taken this approach across a wide range of internal and customer-facing workflows. In this article, I will use an internal workflow (recruiting) as the example, but just keep in mind that the approach generalizes broadly. I have applied it to supply chain, accounting, recruiting, marketing, sales, customer service, research, strategy, engineering, and many more workflow domains.

It works because it starts with systems thinking and supports itself with systems theory. Building agents to support workflows reliably begins with treating the workflow like a complex, dynamical system. The operating model is deterministic. Show me a stochastic workflow, and I’ll show you people doing work with incomplete information. We have yet to implement truly stochastic workflows, but that’s a topic for a different article.

The approach I’ll explain in this article overcomes two of the largest barriers to getting started with reliable agents. First, getting buy-in for everything required to deliver agents, not just the technical pieces. Second, delivering value quickly even when data, information, infrastructure, and capabilities are in early maturity phases.

The mistake I see many businesses make is trying to push this through as a technology-driven change. It must be value-centric and must be driven by the business itself, not the technology team or product management. Here’s how to make that work.

The First Thing I Do

Read more

Just as nuclear weapons changed the calculus of warfare, AI and agents have changed the calculus of business.

***Dario, skip down a few sections. This article answers your question, but everyone else needs the context you have to understand why the answer is relevant to them.***

The Operating Model As A Flywheel

Since most business leaders don’t know the outcome or business and operating model end states, I have information asymmetry, and that is the basis of my business model. But how do I know the end states before they actually happen? I turned futurism into an evidentiary workflow that delivers a reliable model of what’s next. The thesis is built on our best available evidence and continuously improved as new evidence emerges.

We monetize that thesis and its implementation, but it is not a sustainable competitive advantage. Other consulting companies are figuring it out and catching up. If V-Squared were standing still, that would be a huge problem.

V-Squared’s operating model engineers access to evidence or information no one has by building pieces of what’s next for clients and seeing if it does what my model predicts it will. That increases my information asymmetry, creating an advantage for my business that competitors don’t have access to. They would have had to have known what I knew, when I knew it, and built a flywheel that generates information as efficiently as mine does to be in the same place I am now.

Building what’s next and measuring business impact is my information flywheel and advantage accelerator. By the time they know what I know, I have had time to learn more. It’s not enough to build an operating model that delivers value. The operating model must also improve the business’s capabilities and create or sustain its advantages.

The Perfect Setup Vs Your Setup

I just explained how to build a competitor to my business, and only a handful of people will be able to do it. Even if you’re one of the people who can, in the time it takes you to catch up to where I am now, the flywheel will advance V-Squared even further. Your flywheel would need to be faster than mine, but that’s a point for a different post.

This article is about the constraints to implementing a flywheel in the first place and how to overcome them. There’s no point in explaining how to compete against a company that’s built like this when only a few businesses are. It’s more valuable to explain all the barriers to building an operating model like mine and how to overcome them.

I’m a CEO, sole investor, and the only shareholder, so I have fewer constraints than you do. This is why I teach everything twice. Once how it should be if you have the perfect setup. Then again for how it probably is based on your constraints, optimizations, and objectives.

That’s why questions are such rich ground for me to teach with. A question is a statement of, ‘Here are my business realities. How do your frameworks work under those constraints, optimizations, and objectives?’ Students tell me that the live questions and answers during my instructor-led courses are the most valuable parts, followed closely by the office hours, which allow for long-term support through Q&A after the course ends.

My curriculum is structured around this core tenet. Essentially, the frameworks are just a cool story about cause and effect until I explain how they flex to fit your realities. Systems, models, and frameworks are the starting point. Questions give me context about your reality so I can explain how the frameworks flex to fit it.

It’s not enough to understand the structure. You must also understand the mechanics for any of this to be actionable. But for mechanics to inform actions that are relevant and personalized to your reality, they must be framed by the structures you operate in.

Hold that paragraph in your context window or short-term memory throughout this article. It’s a critical agentic architectural pattern that few have built enough to standardize.

Cool Story Bro. I Ran Into Problems Trying To Build It.

Dario Morelli asked a question last week about the challenges of global outcomes because there’s a lag time between a completed workflow and the global outcome. For many workflows, that lag time is months or years. I don’t know if he realized it or not, but that’s a complex, dynamical systems question. I am expanding his original question a bit, but I promise not to answer 5 questions that weren’t asked…no matter how badly I want to.

I just need you to understand that transitioning the business from local to global outcomes introduces complexity and dynamism, so our solution must address that.

Complex because measuring the ‘causal’ (remember causal in the enterprise is the best available definition measured in terms of outcomes, not the scientific standard) chain from the workflow to the outcome has many steps and branches. In the recruiting example, the candidate gets hired, but the global outcome of employee value creation is only filled in after multiple years have passed from the time the hiring workflow finished.

Dynamical because the local and global outcomes and those steps are changing. In the recruiting example, the definition of employee value creation is changing as well. The way the business scores it changes. Business needs change. Employees upskill. Some are promoted into entirely new roles. As we have seen, a lot can happen in two years.

A couple of months ago, I said that AI scales access to information and that we are too early to be measuring intelligence. AI doesn’t know what to do with the information it has, so its agency is very limited. That is where this article and Dario’s question begin. This is novel ground because before agents, this was a difficult, expensive transformation.

AI and agents scale access to information about outcomes. The act of building them using my frameworks scales transparency into local and global outcomes, workflows, and so much more. That allows us to manage the business in a way that was never feasible before now.

AI and agents alone do not transform a business or operating model. However, if you align the technology’s architecture and implementation with the business and operating model transformation, they become massive enablers and accelerators. That is the goal of my frameworks and a much deeper rabbit hole that this article will step past.

Causal workflows were not feasible until recently. We needed AI and agents to scale access to information, so it became feasible to measure workflows and outcomes this way.

KPI Maturity: The Journey To Causal Workflows

Read more

Unfortunately, most of us lack the $100 billion hardware budget required for us to be lazy. Even Microsoft is telling its engineers to tighten their belts and optimize everything. We need graphs managing our agents, and building graphs without an infinite token budget requires us to solve the context and verifier problems.

Graphs aren’t new, but they deliver new capabilities, more control, and improved reliability. They also introduce a new problem: the learning problem. Loops take on a new role to support early iterative learning. Building, improving, scaling, and maintaining a graph manually is more expensive than infinite tokens.

Again, we need a bridge.

The Graph: What Is A Graph?

Graph engineering wires multiple specialized agents or steps into a graph data model. Nodes are the agents or steps. Edges are the routing between them, including branches, splits, merges, and loops. Shared state flows along the edges. Surprise, part three of your agentic harness is a data model and the tipping point into information models. Those two are different, and I will explain the differences in detail in just a bit.

With graphs, instead of one agent cycling through an entire bug fix, you build five steps. One reads the bug report and identifies which files are involved. Three, each examining one file, run in parallel. One collects their findings and writes the fix. Each step is its own model call with its own instructions, and the wiring between them is ‘coded’, so the model doesn’t have to make a decision with incomplete context.

A single loop is the simplest possible graph. It’s one node with an edge pointing back to itself. Workflows are more complex graphs.

You probably noticed that I put coded in quotes. It’s the wrong approach here, but it’s the default approach for software engineering. Logic lives in code, not your data model. Many are rediscovering the ‘code is logic and logic is symbology’ approach to agents.

However, information models store logic more efficiently, and graphs are the bridge from data models to information models. That means software is no longer our primary logic store. This is an entry point into one flavor of neuro-symbolic AI. I explained the concept further in a previous article, but don’t go there until you read the rest of this article.

You need to see a core concept upfront to understand how graphs for agents create the bridge we need.

Read more

Look at Figma and Datadog’s share prices today for the most recent evidence of the storm damage. Airtable’s recent acquisition price hit earlier this week like thunder that rattled VC and PE windows.

Most are focused on Google’s leadership shuffle, but Salesforce’s salespeople are a bigger story. Google’s legendary chief scientist, Jeff Dean, left to start his own lab. Researchers are motivated by the work. Dean wanted to research something that Google didn’t want to devote significant resources to internally. His departure is a difference of opinions on advanced R&D, and Google is hedging its bet by investing heavily in Dean’s new startup.

Salesforce’s salespeople have been leaving for roles at OpenAI. Salespeople are motivated by commission or product sales-related incentives. They want to sell the products with the highest demand and ASP. Leaving Salesforce means most of them believe OpenAI’s enterprise platform and products will perform better than Salesforce’s.

That’s a huge demand signal. The people best positioned to gauge customer sentiment and future demand for Salesforce’s platform are leaving Salesforce. Every light on the board is flashing red for most legacy tech companies. They seem to rotate into and out of the news cycle on a weekly basis. It’s easy to become numb to it.

A growing body of evidence is taking the SaaSpocalypse from narrative to thesis. It has nothing to do with vibe-coding SaaS replacements. That’s what the narrative gets wrong, and there’s no evidence to support that happening. Agents fundamentally reorchestrate workflows to create value, but the SaaS version of agents bolts them on to existing workflows, creating additional costs but no incremental value.

Bolt-on agents carry the lowest risk. Workflow reorchestration carries much higher risks.

Bolt-on agents deliver no value. Workflow reorchestration creates massive value.

There are two types of CEOs running companies today. Type 1 has or is about to get fired. Type 2 is on the rise, but doesn’t have a clear path to meet the investor expectations that got them hired. Align yourself with type 1, and you’ll be laid off with them. Align yourself as an enabler for type 2, and you’ll be best positioned to be promoted as they succeed.

We’re in the endgame for legacy tech. They don’t all disappear, but there is a massive shuffle coming for companies led by type 1 CEOs. Act while your future is still under your control.

The Ugly Truth Of Agents & Why Type 1 CEOs Exist

Every time I talk with C-level leaders about AI, information, and agents, I must make a concession. Companies don’t need them. As soon as the 80% to 95% AI initiative failure rates were published, the lightbulb went on. Companies are investing in AI and delivering solutions, but those solutions are not differentiators.

Almost all AI initiatives fail outright, and the few that succeed don’t give one business a competitive advantage over the rest. There are quantifiable benefits, but they don’t rise to the level of existential threat. CEOs figured it out.

Investors demand that they do something with AI, so they can’t do 0. However, if they do too much and the ROI doesn’t materialize, investors will punish them. The risk of going all in (and failing, which is the most common outcome) is much higher than the risk of doing the bare minimum and claiming victory.

Is it any wonder that most firms are doing the bare minimum, laying people off due to AI efficiencies, and claiming victory? The bare minimum is usually buying Claude or GPT licenses, advancing a few agentic pilots with no incremental budget behind them, and investing in a new agent builder platform from their favorite vendor. They commission an AI strategy and document the roadmap.

Boxes get checked. Jobs get eliminated. Victory laps are taken on the earnings call.

Jensen Huang convinced the hyperscalers that not ramping up their compute capacity was a huge risk. Competitors who did would reap the revenue growth and investor love while those who didn’t would stagnate because they couldn’t keep up with demand. He convinced them that GPUs were a competitive advantage, so they went all in.

The four biggest are on pace to spend roughly $725 billion on capital expenditures in 2026, up 77% over 2025. AI and agents are the killer apps for GPUs, and the hyperscalers’ growth acceleration has proven that out in the last 12 months.

Risk Or Opportunity?

There is no killer AI app or agent. There is no killer information product or use case. There are massive opportunities, but there’s nothing forcing CEOs to take them and no Huang making the case effectively. The risk of chasing opportunities and failing is higher than the risk of doing the bare minimum.

Most CEOs are risk managers. Increased spending increases the risk of lower margins. Approving AI initiatives with real budgets risks losing focus on the core business. The risks keep popping up, and for risk-manager CEOs, that’s enough to keep them from going all in.

BCG found only 52% of US CEOs and 44% of UK CEOs are confident AI will pay off, and Western CEOs are far more likely than their counterparts in India and China to say they are investing because they feel pressure, not because they see an opportunity. About half of US CEOs are risk managers (type 1), not opportunists (type 2).

CEOs must be opportunistic to go all in on AI, information, and agents because the opportunities are massive, but risky. Few CEOs are opportunistic unless they are forced to be by competitors, customers, or investors. OpenAI and Perplexity forced Google to become a risk taker. AWS forced Microsoft to be a risk taker in the last major tech cycle.

But in most industries, competitors aren’t gaining an advantage that forces CEOs to become risk takers. Customers aren’t clamoring for AI either.

BUT investors are demanding accelerating growth. Today that’s most obvious in the technology sector. The risk-manager CEOs of legacy tech are watching their share prices fall. Opportunist CEOs who deliver are riding a wave of enthusiasm. The opportunistic investors who are willing to take risks on AI accelerating growth have plenty of places to pile in. Those who choose risk management have plenty of options as well.

Nothing changes significantly until we see a new dynamic take hold, and that’s the storm that’s raging right now. In early 2027, we will look back to Q2/Q3 of this year as the tipping point for a new dynamic.

Incumbents Must Fall, Both CEOs & Businesses

Intel chose to be a risk manager, and NVIDIA took over the market by being opportunistic. Intel didn’t see the competitive threat. It didn’t feel customer demand shifting. Vocal investors were largely ignored. Then in about 12 months, Intel fell from industry leader to laggard, dropped out of the Dow in favor of NVIDIA, watched its market cap fall below $100 billion for the first time in 30 years, and its CEO was forced out.

Intel is the highest-profile case, but not the only one.

Chegg is the first public company AI took out. The CEO named ChatGPT as a threat on an earnings call in May 2023. They managed the risk for two years and never took the opportunities they had.

Teleperformance is further along the same road. The stock trades roughly 70% below its peak. BPO annual contract value fell 14% in 2025, the lowest since 2020. Nothing broke operationally. Buyers simply stopped signing long contracts for human labor that they expect AI to take over.

SaaS companies are under similar pressure. Buyers are reconsidering whether they should sign long contracts for SaaS platforms and apps they expect AI and agents to take over. It’s not about replacing SaaS with internally built SaaS. Salesforce’s salespeople explained what customers really think. Buyers believe SaaS will be replaced by agents, and their Salesforce contracts will soon be OpenAI contracts.

Investors aren’t waiting this time. They look at the SaaS AI strategy and see risk management, where they believe opportunistic behaviors are necessary. Share prices are falling, even while these companies are growing at the same rates they used to be. About $2 trillion in software market cap has evaporated in 12 months.

Software trades at a discount to the S&P 500 for the first time. Multiples have compressed from 9x sales to 6x. Salesforce is off about 30%, and Workday is off about 40%. The market decided that growth priced per seat has a ceiling. Investors aren’t fully convinced that seats + consumption pricing models will succeed in the long run. When share prices keep falling, CEOs lose their jobs.

US companies changed CEOs 209 times in January 2026 alone, the third-highest January on record. Boards are not waiting for their own Intel.

Investor expectations have shifted from constant growth to accelerating growth. They know that constant growth is a sign that the end of the product cycle is close. Companies like OpenAI and Anthropic are run by opportunistic CEOs who are willing to take risks, not simply manage them. The people who sold Salesforce’s platform are risking their paychecks on opportunistic CEOs.

The proof is in revenue acceleration that runs in the face of rising prices. It’s happening for NVIDIA, memory makers, hyperscalers, and everywhere else there is a lack of supply and a glut of demand. Anthropic’s annualized revenue went from $1 billion at the end of 2024 to $14 billion by February 2026, compounding at roughly 10x a year while OpenAI compounded at 3.4x. And they just keep accelerating. That is an imbalance, and opportunistic CEOs are looking for their own imbalances to monetize.

This is the environment where opportunistic CEOs thrive, but they need help. Step 1 is siding with an opportunistic CEO or distancing yourself from a risk manager, and preparing for their replacement.

Opportunity Isn’t Visible From The Top

We’re early in the transition from risk managers to opportunistic CEOs. In BCG’s January 2026 survey, 72% of CEOs said they are the main decision maker on AI, roughly double the year before. Half said their job depends on it. Half of CEOs are waiting to be shown opportunities worth betting on. The rest either don’t care enough to be the main decision-maker or are still in risk management mode.

94% said they will keep investing even if this year’s spending doesn’t pay off. The constraint is not will or even conviction for the first half. CEOs are the worst-positioned people in their own companies to find specific opportunities. A CEO’s time is often consumed by enabling the business that currently exists. They need people thinking about what’s next who can turn opportunities into revenue growth.

The opportunities in this cycle show up as workflows (customer or internal) that are ready to transform. MIT found that deployments built with outside partners reached production 67% of the time versus 33% for internal builds, and that the gap came down to domain specificity and workflow fit. Domain specificity and workflow fit live four levels below the CEO. And remember, agents create value by reorchestrating the workflow, so the as-is workflow can often hide that value.

That information does not travel upward on its own. 91% of organizations say they use AI while only about 21% of workers actually do. Enterprises are filled with these kinds of disconnects.

61% of CEOs told BCG their boards are moving faster than the strategy warrants. A CEO who cannot describe the opportunity in terms of revenue growth on a timeline cannot defend a pace, a budget, or anything else. An opportunistic CEO without an opportunity pipeline with ROI estimated upfront is just a spender.

What Opportunity Discovery Requires

Opportunity discovery is a job. It needs a short list of hard capabilities, and most companies don’t have them. Risk managers don’t prioritize opportunity discovery, just incremental improvement. Position yourself as the owner of AI opportunity discovery and the person who builds the opportunity pipeline.

Information fluency. Most of the value in this cycle is locked in the information layer and information flywheels, not the model layer. Knowing what information the company holds, how to extract it, and what is missing are critical. It’s even more important to understand how to iteratively grow that information to support new value creation.

Current technical footing. Capability moves faster than legacy planning cycles. Anyone whose picture of what AI can do is six months old is making decisions based on a picture of the market that is no longer accurate, and that describes most executive teams.

Imbalance thinking. Efficiency cases produce efficiency returns and no advantage. NVIDIA’s bet was a read on supply and demand vs. cost and risk. Discovery has to size a future imbalance and quantify its worth before the market does.

Standing to be wrong. Opportunity discovery delivers a portfolio and a pipeline, but some opportunities fail to thrive. Whoever runs it has to kill their own ideas without losing standing. Their track record of success, coalition building, and risk evangelization must create space and tolerance for failure.

I have a complete self-paced course on opportunity discovery. Check out the latest AI redesign of my courses website. If you haven’t seen it in a few weeks, there have been massive updates to the entire website ecosystem. It’s a good example of how information flywheels iteratively improve outcomes.

Version 1 was undeniably AI slop. I left HighROIAI.com at an earlier state so you can compare it with the updated DataScience.vin. Now we’re on version 7 of the ecosystem, and many of the rough edges are smoothing out. Follow their progress:

Outcomes Economy

Endgame Engineering

Strategy

The Frontier Lab Strategy & Its Gaps

Frontier labs are built on the premise that more capable models are the most efficient path to more capable agents. The model will do 90% of the work with minimal support from components like harnesses, loops, and graphs. People from OpenAI and Anthropic have taken to social media urging users to delete their MD file repository with each new release. They claim that most of that context is no longer necessary.

So far, that strategy hasn’t panned out for multiple reasons. The models are undoubtedly more capable, but a model alone hasn’t achieved the consistency, reliability, or customization required to support enterprise workflows. Frontier models, even open-weight options, are still prohibitively expensive for our current workflow.

Let me start out by providing evidence for those claims, so you can see this is more than just personal experience talking. I will explain my approach to model selection, but I’m going to revisit this claim. It is only true if the business has solved enough of the context and verification problems.

As it turns out, the model selection you build today is also a bridge to how the business will manage model selection in the future. That transformation has multiple drivers.

1. How much of the context, verification, and (by extension) learning problems you have solved for the workflow. The better defined the workflow, context, and verification, the smaller the model that’s required. As it turns out, information flywheels iteratively bring down the cost of agents.

2. How frontier AI helps solve the context and validation problems. I will explain this in greater depth in just a minute, but frontier AI has a place in your overall AI strategy. Just not the places most people have it in today.

3. How rapidly small model capabilities are improving.

The ‘One-Shot’ Approach To Completing A Workflow Isn’t Viable, Even With Frontier AI

The math here is extremely straightforward. If an agent’s per-step reliability is p, a workflow of H dependent steps succeeds at roughly p^H. At 98% per-step accuracy, a 100-step workflow completes 13% of the time. At 85% per-step, which is generous for an unassisted model on a real enterprise task, 10 steps get you to about 20%. This is the math of compounding errors and cascading systems failure.

METR’s time-horizon work shows frontier models near 100% success on tasks that take a human under 4 minutes, and below 10% on tasks that take a human more than 4 hours. More steps without context and verification to keep things on track lead to failure in even frontier AI. However, models are improving in ways that most benchmarks fail to capture.

METR’s 50%-success time horizon has been doubling every 7 months, but that appears to be dropping to every 4 months. By this time next year, the doubling rate might be even faster. That progress rate curve tells us where on the workflow complexity scale that frontier AI’s one-shot stops working. This isn’t enough to support a complete enterprise workflow, but it is enough to enable individual steps. That’s really important, and I’ll explain why in a bit.

The paper ‘The Illusion of Diminishing Returns’ isolated execution from reasoning by handing models the full plan and the required knowledge/information/context, then measuring how many steps they could execute. They found that per-step accuracy degrades as the context fills with the model’s own prior mistakes. This is the context problem from my article on harnesses.

You cannot buy your way out of the context or verifier problem on long workflows with frontier AI, and small models are not up to the job over many steps, even with context, unless you solve the verifier problem as well. The verifier problem prevents the context window from filling because you only pass forward what the model needs to fix for the next iteration in the loop.

Spinning Up Multiple Agents In Unconstrained Loops Has Not Worked Either

The best evidence here is MAST, Berkeley’s multi-agent failure taxonomy. They annotated more than 1,600 execution traces across 7 popular multi-agent frameworks and sorted the failures into 14 modes in 3 families:

1. Specification issues caused 41.8% of failures.

2. Inter-agent misalignment caused 36.9% of failures.

3. Task verification caused 21.3% of failures.

Roughly a third of observed failures are inter-agent misalignment. You can’t solve the context problem by adding agents.

Cognition found that actions carry implicit decisions, and conflicting decisions carry bad results. Two sub-agents each make reasonable local choices, and the merge produces something incoherent. The authors’ conclusion is that most of these failures stem from system design, not model capability. Better orchestration fixes them. Bigger models do not.

‘Multi-agent doesn’t work’ is too broad, and I want to clarify a key distinction so you don’t think I’m saying that multi-agent systems don’t work across the board. Anthropic published a multi-agent research system that outperformed a single agent by 90.2% on their internal research evaluation. It also burned about 15x the tokens of a normal chat interaction, and they found token usage alone caused 80% of the performance improvement.

Their guidance is that it works for problems that split into genuinely parallel strands, and works poorly for tightly interdependent work like coding. The failure mode is not agents working together or the number of agents. It is the unconstrained collection of agents thrown together without solving the context and verifier problems.

The strongest proof of the ‘multi-agent + context and verifier’ argument comes from Cognizant AI Lab. They completed a 1,048,575-step task with zero errors by decomposing (defining) the workflow to the point of absurdity. They provided context and created verifier mechanisms to support the agents at every micro-step.

Reliability comes from information and architecture, not the model’s complexity or the number of agents that are deployed. Multi-agent systems work, but adding agents still doesn’t overcome the context or verifier problems.

The Cost Gap Is A Multiple That Undermines Or Enables Margins

As of July 2026, Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens. GPT-5.4 costs $2.50/$15. Claude Sonnet 4.6 is $3/$15.

Llama 4 Scout on Together AI costs $0.18/$0.59, DeepSeek V3.2 costs $0.14/$0.28, and Mistral Nemo costs $0.15/$0.15. That is a 25x to 40x spread on output tokens. Smaller models make agents’ unit economics work for more workflows.

Microsoft’s Azure numbers give us a workflow-level comparison. On a workload processing 50 million tokens per month, the difference between routing to a frontier-class model and routing to Phi-4-mini is roughly $5,000 versus $125. That is the difference between an economically viable workflow and a demo that the feasibility assessment recommends shelving.

This is especially true for workflows that only convert to value in a low percentage of cases. Agentic commerce is a good example. Only a small percentage of searches convert to sales. At $5K per day, the unit economics of those searches is questionable for small to medium-sized retailers. At $125 per day, returns scale faster than costs on much lower volumes and average margins per transaction.

Agentic workflows are not one call, so either the costs or the savings compound with each iteration and loop. Most of the model calls are for doing something with low complexity (for an SLM or LLM) like parsing, routing, formatting, checking, and extracting. Paying frontier rates for simple steps is the single most common cost sink I see in client architectures.

Here comes the “but.” Low complexity calls are only low complexity if we have solved most of the workflow’s context and verification problems. If we have a good understanding of what to parse and how, a very small model can get the job done. If we have a good understanding of what a high-quality parsing result looks like, a very small model can loop its way to success.

If neither of those is true, we must look at frontier models as a potential bridge. Their capabilities can help fill in some of the gaps to make an agent reliable enough for early adopters to use. As I explained in the other articles in this series, usage kicks off information flywheels. Essentially, frontier AI can give us the foot in the door to begin solving the context and verifier problems. As we build both answer keys, we can reduce the size and cost of the model used.

A New Approach: Minimum Viable Model

There’s a new approach to model selection that rejects the frontier AI strategy unless it is absolutely necessary. Companies like Microsoft and Thinking Machines have proven it works, so again, this isn’t just my opinion and experience. Microsoft saw significant optimization in its engineering workflows using what I call the Minimum Viable Model (MVM). It starts with a question.

Read more

The largest tech companies are shifting their business models to provide platforms that support an outcomes economy and outcomes-based operating models. Microsoft ramped up a massive organization of FDEs to help businesses make the transformation. AI, information, and agents are transformation drivers, but the technologies alone will not transform the business.

Businesses don’t have all the pieces to make the transformation in a single step. Harnesses, loops, and graphs are a bridge that helps the business and operating model take advantage of AI, information, and agents by generating value while building the pieces required to transform. The architecture and technical capabilities mature in parallel with the business and operating models’ transformations.

Mark Zuckerberg explained information flywheels as the moat Meta is building on. The flywheel you build by watching a community (either internal users or customers) use a product is the advantage. Every time someone uses the product, they generate contextual data that can be used to improve the product.

That is the purpose of loops and graphs. They are the mechanisms that build the information flywheels that enable the business to progress on its business, operating, and technology model transformations. These are the bridges from an artifacts-based business to an outcomes-based business.

Realize that once the business crosses the bridge, it is never going back. Businesses that don’t build and cross the bridge will not survive. This architecture is a technical and strategic necessity.

The Loop: What Is A Loop?

Read more

Satya Nadella: The goal is empowering every organization to build its own continuous learning loop and ensuring they don’t outsource their core IP. He described the enterprise as a learning machine that needs its own learning machine.

Mark Zuckerberg: When asked what sustainable advantage his lab was building, he said, people are obsessed with intelligence, but the data and the knowledge part is what actually serves people, and the flywheel you get from watching how a community uses the product is how you build sustainable advantage over time.

Sundar Pichai: When asked what the moat is now that everyone has the same models, he said, the model is just an ingredient in the solution. Customers need their data and their trajectories confidential to them, with nothing flowing back to the models.

That is the same thesis, and it’s one you have heard here for over four years. Your data and information strategy is your AI strategy. That is the only source of competitive advantage you control and can own. Last week our frameworks became the script that three multi-trillion-dollar companies gave investors, and every other enterprise on earth was listening.

It Will Soon Be Universal: Data Without Context Has Little Value

Congratulations to us! That reaction feels good. Let’s take a victory lap and tell the world we said it first! The problem is, and everyone who has done this long enough knows, you don’t get credit for being right last year or even 5 years ago. Only amateurs take victory laps. In the enterprise, it’s all about ‘What do you have for me now?’

An idea is worth money in the window between ‘you understand it’ and ‘everyone understands it’. Inside that window, acting on it is an advantage. We know something that the rest of the market doesn’t.

Once the window closes, acting on it is competitive maintenance. The window on information flywheels and advantages is closing. You need what’s next to maintain your internal thought leadership or to position yourself as an AI thought leader in your company.

That’s how you become the C-Suite’s default choice for all things AI. It’s how you get your foot in the door and build executive presence. This is the key to getting invited to the meetings where you can influence decision-making.

Data alone isn’t enough for AI and agents. AI is an information product, and information requires contextual data. C-level leaders are hearing about context extraction and information flywheels for the first time. This is a massive opportunity for you, but it’s a balancing act.

The Opportunity & The Pitfall To Avoid

C-level leaders thought they understood AI. Frontier models, intelligence, and agents were the path to AI success, but the evidence of failed attempts proved that thesis wrong. They are waking up to a new thesis that they don’t fully understand. AI is uncertain again, and they need someone to explain the new reality.

What we must all do now is a balancing act. Do it well, and you’ll put yourself on the fast track for career advancement or new client engagements. Do it badly, and you’ll tell the C-Suite that you’re not ready.

Read more

They get discussed as rival approaches, but in reality they are different parts of the same system, and there are still parts missing. In this series, I will explain all 3 emerging agentic engineering concepts and how I fill in the gaps in my implementations. What you will quickly realize is that adopting agents, integrating them into the business, operating, and technology models, is an act of creative destruction.

But that destruction must happen as a migration vs. an explosive demolition. Harnesses, loops, and graphs create a migration path, but we aren’t explaining that very well. In this article, I will cover the architecture in line with the migration it enables. I will discuss the destination throughout.

What Does An Agent Actually Do? No One Knows…But We Can Figure It Out

Let’s give an AI agent a task, fixing a bug, to provide a high-level explanation of what each component does for the agent.

The SLM or LLM only produces text. It cannot act and do things like open a file, run a program, or save a change. Something else must sit between the SLM or LLM and the rest of the world to translate the model’s text into real actions. It must translate the results back into text the model can read. That is the harness.

One attempt is rarely enough to deliver a high-quality artifact. This is true at the task and workflow levels. The agent reads some code, makes a change, runs the tests, and finds it broke something else. It needs to try again with what it just learned. The layer that keeps an agent cycling until the job is done is the loop.

Some tasks are too big for one agent cycling on its own. The context fills up with dead ends, and independent pieces of work end up waiting in line for no reason. Splitting the job across several specialized steps with defined handoffs between them is the graph.

Those components are an attempt to cover 3 questions:

1. What can the agent do, and see? The harness answers this.

2. How does it decide what to do next, and when to stop? The loop answers this.

3. Does one agent handle everything, or do several split the work? The graph answers this.

There is a fourth question that these approaches fail to answer:

How does the agent know what its actions will cause? That question is the subject of the last part of this series.

Failing to answer the causal questions (predictive, prescriptive, and diagnostic) is why agents remain unreliable at anything without a built-in answer key. I will argue that an answer key is exactly what agents need to do work reliably. However, I can’t ignore the fact that most businesses don’t have that answer key yet. We must start somewhere, and it is best to meet the business where it is. That’s what harnesses, loops, and graphs are doing.

However, each layer is architected as if you will never have an answer key, and that is where I want to call your attention to a massive gap. My WITA cycles (Work, Information, Transparency, Augmentation) reveal a missed opportunity. When work gets done, just the act of using technology to do it creates information about how people do work. That information, one WITA cycle at a time, creates the answer key, but only if the system can learn and has mechanisms to capture that learning.

The ability to learn is implied in harnesses, loops, and graphs, but the mechanisms are unspecified. Andrej Karpathy’s Wiki-LLM is one attempt to address this gap. Keeping everything in text files, even when they are arranged in a taxonomy, is still inefficient. There are more suitable information structures we can use to build and iteratively improve the answer key.

The Harness Takes Agents Beyond The LLM

Read more

Agents pushed the value of a platform or app to the edges where intents and outcomes live. Everything else is getting squeezed into the middle and commoditized. Now we’re seeing the impacts of that shift play out in quarterly earnings.

On Wednesday, Alphabet and ServiceNow both beat their numbers, and both stocks moved. ServiceNow went up about 5% after hours, but Alphabet went down.

That is worth exploring further, because the market’s reading is backwards. Alphabet grew revenue 24% to $119.8 billion, tripled Google Cloud operating income to $8.81 billion, and posted a Cloud backlog of $514 billion. ServiceNow grew subscription revenue 24.5% to $3.877 billion, roughly one twentieth of Alphabet’s quarter, and the market liked ServiceNow more.

The reason is not that one company executed better. The two companies are standing at opposite ends of the same cost and value structures. They have made opposing bets about which end pays the most, and this quarter gave us the first clean read on what each bet costs.

For anyone building or pricing an AI product, this is the most instructive pair of earnings we have had all year. Let me clarify both companies’ AI strategies before getting to the pricing lessons learned.

The Structure Both Companies Are Monetizing

Value in a platform market concentrates at two ends. At one end sits the intent surface. At the other end sits the outcome layer. The middle, the part that translates intent into execution, commoditizes. It always has, in every platform generation. AI has accelerated the schedule.

If you touch neither end, your app or platform is the middle, and you need to figure out how to connect with an intent or outcome to survive.

This moment is fundamentally different from prior platform shifts. For most of the SaaS era, enterprise application vendors held both ends, so no one really thought to explain the barbell shape. The user interface was the intent surface. The workflow engine underneath was the outcome layer. One seat license paid for both and the cost structure of serving both led to massive margins, so no one asked which half the customer was buying. There was no need to define the barbell.

Agents broke the neat package, because intent moved into Claude, Copilot, Gemini, or whatever window the employee already had open. Once intent leaves your interface, the interface becomes a cost center.

Both companies that reported Wednesday have accepted this. They just picked different ends of the barbell to monetize.

Strategy 1: Own Intent, Sell The Layer Underneath It

Google’s position is the one everyone thinks they understand. The part people usually get wrong is where the money comes from. The intent numbers are massive. The Gemini app has 950 million monthly active users, with daily active users tripling over the last year. AI Mode passed 1 billion monthly users after going global last October. Search usage hit an all-time high during the World Cup. More than 140 million people used Ask YouTube on the watch page in June alone.

However, none of that is revenue. Google does not charge those 950 million people for expressing intent. It monetizes the layer underneath. A big part of that is ad revenue. Google doesn’t deliver the sale (outcome); it monetizes the distribution surface (intent to buy). It is working with multiple retailers to build the future of agentic commerce, but again, it monetizes the intent but does not follow the customer through to delivering the outcome.

That approach means every time someone uses Google’s AI, there is an opportunity to monetize that inference. Returns scale in line with costs.

Model APIs are running at roughly 22 billion tokens per minute, up from 16 billion a quarter ago. More than 9 million developers build with those models monthly. On the enterprise side, nearly 500 Cloud customers each processed more than a trillion tokens over the last year, and more than 2,000 enterprises crossed 100 billion. Pichai said existing Cloud customers are exceeding their commitments by more than 50%, an acceleration over last quarter, and that new customer acquisition velocity more than doubled year over year.

Gemini bills by the token. Tokens increasingly run on Google’s chips, which means the margin on each one goes to Google rather than to a chip vendor. Google has the hardware capabilities to make the unit economics of inference work. But again, Gemini supports intents and charges for every intent it consumes and responds to, regardless of the outcome.

That is why Cloud operating income tripled while Cloud revenue only grew 82%. Operating leverage at that ratio is the signature of a business whose unit costs are falling faster than its prices. That’s another reason Google is willing to spend so much to increase its data center capacity. The business prints money and they have more demand than they can service.

The Gemini Flash series, per Pichai, hits the sweet spot of performance and cost. AI Mode response costs reached their lowest level since launching this quarter, even as the underlying capability improved. And the price is $44.9 billion of capital expenditure in a single quarter, roughly double the year ago figure.

Alphabet raised full year CapEx guidance to a range of $195 billion to $205 billion. Anat Ashkenazi told analysts the increase was primarily an acceleration in the delivery of capacity to meet demand, and added that the company is still supply constrained.

Remember the supply constrained piece. I will come back to it soon.

Strategy 2: Give Up Intent, Sell The Governed Action

ServiceNow cannot win the intent surface. There is no version of the enterprise where 950 million people open ServiceNow to express what they want, and to the company’s credit, it stopped pretending otherwise. Salesforce is still chasing intent with its rearchitected AI Slack agentic interface. We will soon see how that’s going.

At Knowledge in May, ServiceNow shipped Action Fabric, which opens the platform’s full system of action to any AI agent through an MCP server. Anthropic is the first design partner. Claude Cowork is wired directly into governed ServiceNow execution. Microsoft Copilot is another connecting agent, but a full implementation is pending. The company’s explanation, from Nenshad Bardoliwalla, is that other platforms let agents read and write data while ServiceNow lets agents execute governed work.

What does that give away and what does it keep? It gives away the intent surface. A password reset that used to require the ServiceNow portal now runs from Claude. ServiceNow still delivers the outcome. An onboarding workflow that HR triggered through a form now runs because an agent submitted the request from a chat window. Again, ServiceNow delivers the outcome. Employees may go months without seeing a ServiceNow screen, but they will still get the outcomes they expect.

ServiceNow’s AI strategy is to keep the execution and outcome. Actions still run through ServiceNow’s workflow engine, hit ServiceNow’s data model, and pass through AI Control Tower. The decade of workflows that customers built on the platform, which were previously reachable only by logging in, are now reachable by any agent with credentials. They remain ServiceNow’s to govern and deliver.

The three-year agreement ServiceNow signed with OpenAI in January, which made OpenAI the only natively embedded model on the platform, makes sense. One model is locked into the experience ServiceNow controls. The execution and outcome layers are open to everyone else. ServiceNow has decided which end of the barbell it can defend and has stopped spending on the other.

McDermott said it on the call. Whichever chip, lab, or price-per-token regime wins, the enterprise needs one governed layer of record for work. He described the platform as optionality on all AI outcomes rather than a bet on any one.

Comparing The Results, Side By Side

Both strategies worked this quarter. ServiceNow’s numbers are strong. Subscription revenue of $3.877 billion beat the high end of guidance. cRPO of $13.20 billion came in above guidance as well. Non-GAAP operating margin was 29.5%, again 3 full points above guide. There were 123 deals over $1 million in net new ACV, up 40% year over year, and 658 customers now spend more than $5 million a year. ServiceNow AI ACV crossed $1 billion. The percentage of renewing customers buying agentic AI for the first time doubled quarter over quarter. Customers with agentic AI in production are up ninefold in nine months.

The pricing is holding too, which matters more than the adoption count. Pro Plus products are carrying uplifts above 30%. The new AI-native products are landing at 20% to 30%. Customers are paying materially more for the AI layer.

BUT put the two companies’ margin lines next to each other and things change.

Google Cloud: Revenue up 82%, operating income up roughly 200%. Margin expanding as AI adoption grows.

ServiceNow: GAAP subscription gross margin at 73.5%, down from 80% a year ago. Non-GAAP at 80.5%, down from 83%. Full year gross margin guided to 81%, and the company attributes the compression to more customers using hyperscaler deployments and to an acceleration of customer AI adoption.

The more AI ServiceNow’s customers use, the less each subscription dollar is worth to ServiceNow. Adoption and margin are moving in opposite directions, and the company somehow disclosed that without attracting much attention.

Why The Same Input Produces Opposite Results

Google is not better run. The two companies occupy different positions relative to the cost curve of the input they both depend on. Google produces inference for intents, making it an AI-first company that owns or influences most of its own input costs. It designs the TPUs, operates the data centers, builds the serving stack, and trains the models. When Google makes inference cheaper, the savings land in Google’s income statement as margin. Cost improvements are captured.

The intent-based product strategy means every time Google serves inference, it has an opportunity to monetize it.

ServiceNow purchases inference, putting it at a very different end of the cost curve. Every AI-driven resolution it delivers has a token bill attached that’s paid to a hyperscaler. When Google makes inference cheaper, ServiceNow’s costs fall only if Google chooses to lower prices.

ServiceNow must also deliver outcomes that meet customers’s expectations of quality. That means more model calls are required. It is not paid per output. It only gets paid if the customer is satisfied with the outcome.

The pricing models and cost structures are where the two strategies compound in very different ways. Google sells inference by the unit consumed. When usage rises, monetization opportunities and revenue rises with it. Pichai can cite customers exceeding commitments by 50% as unambiguously good news.

That would be a cost nightmare for ServiceNow. It sells a mostly fixed-price subscription but has a variable-cost delivery structure. On the call, Mastantuono said that customers are not paying for tokens. They are paying for resolutions, aka outcomes. Amit Zavery said the same thing from the product side, that pricing is based on the solution rather than individual tokens.

Resolution or outcome is the right thing to charge for. It is what the customer actually wants, and the 20% to 30% uplifts prove the market accepts that pricing model. Each outcome is priced in advance, but delivered with a variable number of tokens. Every incremental token spent on a resolution already sold at a fixed price comes straight out of gross margin.

That is the test I run with clients. Ask whether your pricing metric tracks the value the customer receives. Then ask whether it tracks the resources you consume to deliver that value. ServiceNow passes the first test. It cannot yet pass the second, because the resource is bought from a hyperscaler whose entire business model monetizes intent.

Value creation, pricing model, and resource consumption must line up. ServiceNow has aligned 2 of the 3, and the missing one is a misalignment that’s hurting gross margins. It and every other business that monetizes outcomes must solve the same problem or their AI ambitions will bankrupt the company one customer at a time.

A Hedge Against Runaway Token Costs

McDermott’s most repeated message from the call was that most customers are now completely allergic to anything that looks like a project. They only want deterministic. ServiceNow only does deterministic.

Read as sales positioning, that is a trust argument against competitors selling open-ended agent platforms that don’t deliver outcomes. Read it as engineering, and it looks more like a cost hedge.

Reliability in an agentic system is a function of how tightly you define the workflow, context, action space, and state space. Loose definitions require a large model to figure out what you meant, which is expensive and unreliable. Tight definitions let a smaller, cheaper model execute correctly, which is inexpensive and reliable. Reliability and cost are not a tradeoff. They come from the same discipline.

Zavery said the level 1 ITSM specialist is live with more than 40 customers and handles 80% to 85% of service requests without human interaction. That is a narrow, heavily bounded problem with a well defined action space, which is why it can run cheap enough to sell profitably at a fixed price.

When ServiceNow says determinism, the customer hears reliability and margins benefit from fewer tokens per resolution. Workflow definition granularity is the one input cost lever a company in ServiceNow’s position fully controls, and it is available to every business. If you are pricing an outcome on purchased inference, tightening your action space is not an architectural choice. It’s a financial necessity and another example of the kind of technical-strategic crossovers we are seeing more often.

Technical decisions impact P&L more than ever. The people making them must have both sides of the equation (technical and strategic acumen) to succeed.

The Hard Part

Mastantuono was asked directly about the margin pressure. She said, “Near term pressure, mid term tailwind.” Hyperscaler adoption is ramping faster than planned, which she framed as a good thing. As it ramps the cost per unit comes down. Add smart token optimization and the pressure eases even more.

That is a real bet that takes strong conviction. The operating margin discipline behind it is not in doubt. ServiceNow held operating margin flat despite the gross margin hit, came in 3 points above guidance, and McDermott committed to entering 2027 with the same headcount the company had before acquiring Moveworks, Veza, and Armis.

But notice who owns a lot of the pricing action required for that bet to pay off: hyperscalers like Google. ServiceNow’s margin recovery requires hyperscaler prices to fall. Hyperscaler costs are indeed falling, and Google confirmed it. Yet Google also raised capex guidance by roughly $15 billion, told analysts it remains supply constrained, and converted its cost improvements into a tripling of cloud operating income rather than into lower prices.

Falling unit costs plus constrained supply do not produce falling prices. They produce expanding margins for the seller. Prices fall when supply loosens or when competition forces the issue, and neither condition is present in the numbers we saw Wednesday.

None of this makes ServiceNow’s strategy wrong. Owning the governed action layer may well be the more sustainable position over a decade, because governance requirements expand while model advantages decay and hardware commoditizes.

However, the timing of the margin recovery is not in ServiceNow’s hands, and that is a different kind of risk than the one the guidance language implies. Every business that depends on inference running on a hyperscaler’s hardware has the same risk.

What To Take Into Your Own AI Platform Roadmap

Name your end of the barbell. Most product strategies I review assume they hold both ends, because that was true for a decade and no one realizes that it is an assumption. If intent has already moved to somebody else’s model, you hold one end, and the roadmap should stop funding the other.

Inventory your actions, not your screens or UIs. ServiceNow’s real asset turned out to be tens of thousands of customer-built workflows. The question for your product is whether those workflows are individually addressable, permission-scoped, and auditable, because that is the difference between an asset an agent can call and a UI it can’t use.

Run the alignment test twice. Once on the value the customer receives, and once on the resources you consume to deliver it. Passing the first and failing the second is the same failure mode on display this quarter, and it does not show up in ACV, adoption, or win rates. Watch gross margins like a hawk and optimize relentlessly.

Treat workflow definition as a pricing decision. Determinism is a smarter early strategy. Bounded action spaces route to smaller models, and smaller models are the lever on input cost that unchains you from your compute supplier.

Watch what your compute supplier does with its cost savings. If you buy inference and sell outcomes, your margin preservation plan might be a forecast of someone else’s pricing behavior. Their earnings call and investor expectations for growth are part of your financial model unless you get model costs under control.

In these early innings, serving intent is a better business model than serving outcomes. However, the big question is, ‘How long until customers demand that AI and agent platforms deliver outcomes?’ How long until agentic commerce demands sales instead of continuing to pay for ads priced on intent? How long until engineering teams demand working, secure artifacts, not tokens and code they must fix themselves?

We see AI in the loops we are currently executing, but that doesn’t mean it will remain trapped there. The shift to outcomes is inevitable, and it will break every loop it touches.

I was one of the first to propose an alternative. What if we introduced technology intentionally into workflows as an intervention? This is the framework that put me and V Squared on the map. Reframing technology as a workflow-level intervention created an experimental approach that allows you to estimate ROI upfront and quantify it post-deployment. It aligns technology with the way work gets done and how it creates value.

The research caught up to that framework a decade later. McKinsey tested 25 organizational attributes against a company’s ability to show EBIT impact from generative AI, and workflow redesign came back as the single largest factor.

BCG reached the same conclusion from a different direction with its 10-20-70 rule. Roughly 10% of the effort belongs to algorithms, 20% to technology and data, and 70% to people and process. Every serious study of AI value creation now describes a workflow-level intervention.

That changes everything, and it’s the core thesis behind why forward-deployed engineers are so effective. But I did not immediately realize the implications for individual business units, and I am watching FDE teams struggle due to the same blind spot I had to address over a decade ago.

The AI Revenue Bottleneck: Deployment, Capabilities, Or Something Else?

The market has voted on this with its wallet. In May 2026, OpenAI stood up a dedicated deployment company backed by more than $4 billion in committed capital and acquired Tomoro to bring 150 forward-deployed engineers in-house.

Weeks earlier, Anthropic launched a $1.5 billion enterprise services venture with Blackstone, Hellman & Friedman, and Goldman Sachs, built entirely around embedded engineering teams. Accenture launched an FDE practice with Microsoft. All of them are running a play Palantir invented even before I founded V Squared. The bet they are all making is that deployment, not capability, is the bottleneck.

I found that many executive leaders resisted intentional technology usage in irrational ways that blunt deployment efforts. The ROI of technology that no one uses is 0, so this was something I had to fix, and they will too.

As AI and agents attempt to gain traction in the enterprise, they are confronting the same irrational resistance. I studied it, and what I found was shocking. Every company from Anthropic to yours will need to confront the root cause of irrational resistance, or all their AI products will fail, and their AI investments won’t amount to much.

This is something that neither FDEs nor technology alone can address.

Resisting Technology Is A Matter Of Survival

We know many business units resist technology because they are afraid it will replace them, and that’s rational. I already had frameworks to address that type of resistance for survival. Through opportunity discovery, we ask what technology should do FOR each business unit instead of dictating what it will do TO them. What I was seeing was different.

This resistance made no sense. Nothing we proposed would replace people. Roadmaps led to increasing their organization’s headcount and business impact over time. What was there to resist?

Everyone from Microsoft to OpenAI is building FDE teams to streamline customer adoption for the tools they have sunk billions into. The more people involved in scaling AI that I talk to, the more I hear the irrational resistance barrier come up.

Introducing technology intentionally as a workflow-level intervention has a second-order effect: transparency. I must measure the ROI of the technical intervention, which requires getting a benchmark of how much value the current workflow creates. And that’s where irrational resistance begins.

I found that executive leaders weren’t resisting the technology itself or the work of integrating it into their workflows. They were resisting the increased transparency into how their organizations create value and how much value they create. Some executive leaders resist transparency and the technologies that bring it because they are afraid of an ugly truth.

The same instinct shows up further down the org chart, which is how we know it is structural rather than a few bad executives. Microsoft’s research found 53% of employees worry that admitting they use AI makes them look replaceable, and Ivanti found 32% hide their AI usage for that reason.

A 2026 shadow AI survey of office professionals at companies above $500 million in revenue found 39% would rather use AI without telling anyone, rising to 47% at companies with over $1 billion in revenue. Most enterprises read those numbers as a governance problem and write a policy.

They are actually a resistance signal. The bigger the organization, the more there is to keep opaque, and the more clearly people understand that visibility is the thing that costs them.

Most business units don’t create value for the business very efficiently. Hidden cost centers are scattered across the business. Agent and AI adoption threatens to reveal them all. Executive leaders who oversee these business units have a vested interest in maintaining their opacity. It’s a matter of survival.

The 6-2-2 Rule: Consulting’s Secret Ratio

In many business units, for every team of 10 people, 6 create no value. Some actively sabotage it. That means 60% of the team is a cost center or worse. 2 high performers create over 70% of the team’s total output, and 2 mid-level performers step up when needed to deliver the rest.

The ratio is not my own personal folklore. O’Boyle and Aguinis ran 5 studies across 198 samples and 633,263 people and found that individual performance does not follow the bell curve we build compensation bands, calibration sessions, and headcount plans around. It follows a power law.

A small group produces output far beyond the median, and the median sits well below the mean. Every HR system in the enterprise assumes normality. Every consultant who has spent a quarter inside a business unit knows it is Paretian. The 6-2-2 Rule is my field version of a finding that has been sitting in the literature for much longer.

These are the teams that resist irrationally. As soon as technology brings transparency, the game is up for most of the team. The 2 high performers will get the credit that often gets misallocated to the 6 who have time for playing politics. Eventually, the 6 know they’ll be let go, so they will fight progress until the end.

We can already watch that fight play out in how the 6 use AI. BetterUp Labs and Stanford’s Social Media Lab surveyed 1,150 US desk workers and found 41% had received the now-infamous ‘workslop’ in the previous month. Each incident cost the receiver just under 2 hours. It cost the sender standing. 42% of recipients trusted that colleague less afterward.

Workslop is what political capital looks like once it is given an LLM to amplify itself with. The 6 are not refusing to adopt ALL AI. They are using it to manufacture the appearance of output faster than the 2 can produce the real thing, and the cost lands downstream where the measurement can’t see it yet.

A Changing Perception Of Productivity At The C-Level

In the past, there was 0 appetite for addressing 6-2-2 teams in the C-suite. A certain level of inefficiency has been tolerated since the recovery from the Great Financial Crisis. The idea behind forced ranking (GE’s 20-70-10 vitality curve, Microsoft’s stack ranking until it dropped the practice in 2013, Amazon’s 6% unregretted attrition target) was to keep running the most efficient business possible.

Everything was more competitive, and growth was harder to come by in the early 2000s. C-level leaders did the work to build organizations that were more like 1-1-9 than 6-2-2. But looser monetary policy and multiple high-margin tech growth cycles padded enterprise wallets so inefficiency stopped being a focus area.

Talk to any strategy or efficiency consultant, and they will have a version of The 6-2-2 Rule. Some say half of all teams’ budgets are wasted. Others say it’s closer to a 90:10 ratio, with 1 high performer carrying most teams’ value delivery.

In any case, none of us shared it outside of think-tank-style presentations. It is unpopular, and there’s no will to address the issue. Why touch that third rail?

But AI has changed that. As businesses pour more money into AI, they must offset that spending somehow. Efficiency is back on the menu, and in the last 6 months, I have seen C-level leaders be a lot more receptive to addressing the 6-2-2 organizations.

The budget math is doing most of that persuading. Companies plan to spend around 1.7% of revenue on AI in 2026, roughly double the 0.8% they spent in 2025. 91% say they will increase AI investment again next year, even with only 1% reporting significant returns. That gap has to close somewhere.

AI: The Excuse For Finally Addressing the 6

AI and the transformation required for enterprises to get the most out of it are cover to do the technical, cultural, and strategic debt that have gone unaddressed for so long. CEOs can do almost anything and follow it up with, “Because AI…” and investors reward them for it. The need to reallocate budget to AI without impacting margins or free cash flow, coupled with the blanket pardon of AI, is creating the perfect conditions to fix 6-2-2 organizations.

The pardon is already being used broadly. Resume.org surveyed 1,000 hiring managers and 59% admitted they emphasize AI when explaining layoffs and hiring freezes because it plays better with stakeholders than citing financial constraints.

Gartner surveyed 350 executives and found companies cut roles over automation whether or not the technology was returning anything. Sam Altman named it ‘AI washing’. Thomas Davenport’s survey of 1,006 global executives found that cuts are happening in anticipation of AI’s impact, not because of its measured performance.

But AI vendors like Microsoft, Anthropic, and Salesforce must be wary of the backlash. If AI becomes the harbinger of layoffs, the job of FDEs becomes impossible. Everyone will resist AI because they see how many people get laid off once an organization adopts it. The real reason doesn’t matter. If the spin is “AI layoffs,” then that becomes the perception.

The reversals are already playing out, and every business unit facing an FDE engagement can recite them to you.

Klarna said its assistant was doing the work of 700 support agents, froze hiring for close to a year, then started rehiring humans after service quality fell.

Commonwealth Bank reversed a decision to replace 45 service roles when the bots left the remaining team so overwhelmed that team leaders started taking calls.

Salesforce took support from 9,000 people to 5,000, then rehired into sales and professional services and started talking publicly about grounding, governance, and keeping humans in the loop.

Notice where the cost of each of those stories landed. The CFO who ordered the cuts moved on. The vendor inherited the perception problem, and the next deployment got harder to sell.

The thing about measuring ROI is you actually have to deliver it. That has always been a holistic, enterprise-wide problem, not an isolated technical challenge.

Changing The Perception To Get Buy-In

I learned that the only way to get organizations that resist technology to change their behaviors is to frame intentional technology use as an opportunity for improvement vs. a reason to lay everyone off. This is a big part of meeting the business where it is.

When technology adoption is viewed through the lens of improvement vs layoffs and disruption, executive leaders see a different path back from 6-2-2. Yes, there’s some upside that can be gained by laying the 6 people off. Or you could use technology to turn those 6 people into high performers. The increase in value created by the 1-1-9 team is always higher than the incremental cost savings of running a smaller team.

There is now hard evidence that this is the better trade. Brynjolfsson, Li, and Raymond studied the staggered rollout of a generative AI assistant across 5,179 customer support agents and found productivity rose 14% on average. The average hides the finding that matters here. Novice and low-skilled workers improved 34%. The most experienced, highest-skilled people saw almost no gain at all.

The system worked by capturing what the best people already did and distributing it to everyone else. That is the 6-2-2 to 1-1-9 conversion. AI doesn’t manufacture high performers out of nothing. It industrializes the tacit knowledge your 2 already have and hands it to the 6. Capability and transparency arrive in the same deployment.

And that leads to the other side of the productivity crisis. You only need higher productivity if the business is growing. That’s why several surveys have shown that the businesses that get the most out of AI are also growing the fastest. You need both: AI as a productivity driver and as a growth accelerator.

The gap between those two groups is wider than most leadership teams assume. BCG’s future-built companies post 1.7X the revenue growth of laggards, 3.6X the three-year total shareholder return, and 1.6X the EBIT margin. They report roughly 5X the revenue gains and 3X the cost reductions from the AI work itself.

McKinsey found 80% of organizations set efficiency as an AI objective, while the high performers add growth and innovation on top of it. Efficiency on its own gets you a smaller version of the business you already had. Growth is what turns the productivity you unlocked into something worth having, and it is the only version of this story where the 6 have somewhere to go.

Without that framing, improvement vs downsizing, the 6 will continue to resist until the bitter end, derailing even the most capable FDE teams. What’s worse is their executive leadership will back them up. No one wants to see their headcount reduced by 60%.

Growth must go hand in hand with productivity improvements, or the whole thing has a very low ceiling. AI strategy and AI product management must align with growth first, then efficiency to support scaling operations to handle that new growth.

If you’re ready to get certified as an AI Strategist or AI Product Manager, it’s time to check out my courses.

In this article, I’m partnering with Tricentis to explain how to translate an emerging need into the language and framing that opens up wallets and gets C-level buy-in.

Understanding The Need

The most important sentence in Tricentis’ 2026 Quality Transformation Report is its last one: Software quality is becoming a boardroom concern, akin to cybersecurity before it.

But does your C-suite and board understand that? It should be obvious. AI tooling makes development a continuous process. AI tools create new quality risks and scale existing risks. Businesses need a new approach to quality that enables them to adapt and reap the benefits of AI coding tools without falling victim to their risks.

More changes are checked in more often. More code is AI-generated, and purely human validation processes threaten to offset the productivity gains. The solution is continuous automated testing, issue detection, and escalation.

The obstacle to what should be an obvious conclusion is that the people who would have to carry quality up to the board can’t yet describe it in language that the board can act on. There’s a number in the data that proves it.

The 23 Point Gap That Should Worry The Board

93% of C-level respondents are confident their testing strategy covers the most critical risks to the business. Only 70% of practitioners agree. That 23-point spread is a population of executives who are confident in a strategy, while the people closest to it are not. They’re signing off on something their own experts are uneasy about.

42% of C-level leaders believe their developers and executives are fully aligned on what good software looks like. Among QA and DevOps leaders, just 22% agree. Most of the confident executives have a gap they can’t see. They are reassured by an alignment that doesn’t exist.

It’s easy to get visibility and approval for AI tools. They are front of mind, and enterprise vendors have done an excellent job of connecting the dots between adoption and value creation. But AI development tools fundamentally change the development lifecycle. Those upstream changes require new tools to support a new kind of quality assurance lifecycle.

Those are a lot harder to get the budget for, and the root cause is the same as the one that causes the 23-point gap.

The Real Job Is Connecting The Dots

Technical leaders understand how the technology creates value. That fluency is what makes us all bad at explaining it. We can see the whole causal chain from better test selection to fewer escaped defects and lower change-failure rate. So we assume the value is self-evident, but outside of the tech bubble, it isn’t.

Everyone who controls a budget needs you to connect that chain to something they already care about, and “the regression suite is more resilient” is not on the list. The big reveal is that not all technology sells itself. Either we translate it, or it dies in committee.

There’s a structure to the translation, because there’s a structure to the room. In any real funding conversation, there are two kinds of deciders, usually at the same table.

The opportunity decider leads with upside: growth, speed, or market position. That’s the CEO, CRO, and often the head of product.

The risk decider leads with downside: loss, exposure, compliance, or the thing that ends up on the incident bridge or in front of a regulator. That’s your CFO, your CISO, your audit and board-risk people.

A capability gets funded only when you can speak to both. Lead with opportunity to a risk decider, and you sound reckless. Lead with risk to an opportunity decider, and you sound like a barrier. Every translation is 3 parts:

1. Here’s what the technology does.

2. Here’s the business consequence or impact.

3. Here’s how you explain that consequence to each decider.

Let’s run Tricentis’ platform through it, because it’s a good example of a quality platform investment built to survive this conversation, and because seeing the moves on a real product is how you learn to make them on your own.

Start With The Artifact

Every discipline that climbs into the boardroom arrives carrying a recognizable object. Cybersecurity didn’t get elevated on a message of more security tools or explanations of what the tools do. It got elevated around a control plane or a single pane of glass, the audit trail, and continuous monitoring. These are the things a CISO could put on a slide and say ‘This is how we know’. Quality needs the same kind of objects, and that’s the center of Tricentis’ design.

What it does. The platform is organized around an AI Workspace that Tricentis calls a control plane and system of record for agentic quality engineering. It coordinates the AI agents doing the actual quality work, holds shared context across roughly 200 enterprise systems plus web and custom apps, connects to the tools teams already run (Jira, GitHub, Tosca, qTest ), and embeds governance, approvals, and an auditable decision trail directly into how releases happen, with humans kept in the loop for the judgment calls.

The business consequence. It is, for the first time, a system of record for release decisions. You can prove after the fact why something shipped, and stop something before it does. Quality stops being a checkpoint bolted on at the end and becomes a governed, continuous layer.

How you say it.

To the opportunity decider: “This is how we release at AI speed without growing the QA team in lockstep. This platform enables us to scale output and productivity, not headcount.”

To the risk decider: “This is our auditable control plane for autonomous releases. When the board or a regulator asks how we govern AI-generated code, this is the documented, provable answer.”

Knowing What To Ship

What it does. Agentic Quality Intelligence continuously reads change and risk signals across the lifecycle, decides what actually warrants testing, judges release readiness, and escalates to a human only when judgment is genuinely required. It’s the engine that blocks untested changes and runs only the tests the change demands. It gives us risk-based selection instead of brute force.

The business consequence. You stop testing everything and start testing what matters, which is the rare lever that makes you faster and safer in the same motion. It’s the operational form of what Tricentis’ own CEO, Kevin Thompson, argues in the report: You don’t need to test everything. You need to understand the changes, their impact, and where the risk truly sits.

How you say it.

To the opportunity decider: “We shorten release cycles because we only run the tests a given change actually warrants. That removes the blanket regression tax on every deploy.”

To the risk decider: “We always know our release-readiness posture, and we can show the risk basis behind every go/no-go decision.”

Keeping Coverage Ahead Of AI-Written Code

What it does. Two agents work the production line. Agentic Test Creation lives inside qTest and turns plain-language requirements into reusable test cases, so generating coverage no longer depends on scarce specialist expertise. Agentic Test Automation runs and maintains those tests through Tosca’s automation engines across SAP, web, and custom applications, intelligently reusing modules instead of rebuilding them.

The business consequence. This is the answer to the report’s central challenge that code volume is scaling faster than teams can validate it. Coverage now scales with AI-generated code without a matching rise in cost or headcount. The bottleneck that’s been capping AI’s productivity dividend is removed.

How you say it.

To the opportunity decider: “Our ability to validate finally scales with our ability to generate. We unlock the productivity AI development promised, but quality risks were throttling.”

To the risk decider: “AI-written code stops outrunning our ability to verify it. The gap between generated and trusted closes instead of scaling.”

Surviving Production

What it does. Agentic Performance Testing puts autonomous agents across the analysis, design, and execution of performance and load testing, from individual APIs to full end-to-end systems, surfacing performance risk early. Tricentis reports that the agents accelerate time-to-insight by up to 90% to 95% over manual expert work.

The business consequence. You find the outage-class defect in the pipeline rather than in production. Discovery happens before it becomes a customer-trust event, which the report ties directly to lost revenue and lost partners.

How you say it.

To the opportunity decider: “We can commit to performance under scale to our biggest customers and actually prove it before we promise it.”

To the risk decider: “We catch the thing that takes the system down before customers do, and we catch it in the pipeline, not on a 2 a.m. incident call with the brand on the line.”

Why This Stops Being Overhead And Becomes Governance

String those four capabilities together through the control plane, and something changes about the category that the spend lives in, and the category is everything to a check writer.

Pitch this stack as more tooling (a testing platform, test automation coverage, and QA management), and it lands as a cost. There, it competes with every other tool, gets benchmarked on price per seat. Overhead is what a board looks to cut.

Pitch the same stack as the governance layer for AI-scale release decisions and the system of record that lets the company move at machine speed without going blind on its risk exposure. It lands beside the security stack, the audit function, and the controls a board is obligated to fund. Governance is what a board is accountable for.

This is the journey cybersecurity has already made. A decade ago, it was an IT line item you funded grudgingly until breaches got quantified, regulators arrived, customer trust got a price, and security walked up the stairs into the boardroom with its own committee and its own budget logic. Once the framing changed, that all changed with it.

The financial data is now doing the same thing to quality. 1 in 5 companies is losing up to $5 million a year to poor software quality. 45% are losing between $500,000 and a million. 40% of large enterprises sit in the $1–5 million band. Those are the shareable metrics we can use to start the conversation with C-level leaders and the board.

The moment poor quality has a dollar figure and a customer-trust line item, it stops being an engineering metric and becomes a category of enterprise risk. That’s important because enterprise risk has a higher-level owner than engineering metrics.

And to name the thing technical leaders avoid at all costs: none of this is spin. You are not inflating the value of the platform. You are translating its real value into the frame that the executive actually uses to allocate capital. Persuasion is helping someone see what’s true through a lens they can act on. Manipulation is getting them to act against their interests. The risk is genuinely board-level; saying so out loud is just accurate.

But with information, you can explain complex patterns. With one taxonomy, I’ll explain how I have diagnosed client technology failures so they understand the root cause and we fix it right the first time. I will use it to diagnose 6 real-world success and failure cases from the last 48-hours. But that’s just the first-order impact of this information.

What I have done as a consultant for 14 years is now turning into new roles businesses are staffing themselves. The only thing keeping consultants afloat is that there aren’t enough people who can do this. That’s the second-order impact, and this taxonomy will allow me to connect the dots between emerging roles and what’s driving them.

A good taxonomy beats thousands of data points, especially when you realize they can be represented mathematically…but let’s leave that for another article. I’m already revealing enough IP in this one.

The Bottleneck Taxonomy

The Bottleneck Taxonomy identifies where enterprise value becomes trapped between organizational capability and business outcomes. Its central premise is that most enterprise bottlenecks are not resource shortages or technology limitations. They are failures in the shared models organizations use to understand the people and systems governing their outcomes.

This is why I designed my curriculum around teaching systems, models, and frameworks. Without the model and framework that make it actionable, it is difficult to be effective in a business setting. You may understand the model in your column, but does the rest of the business? Can you work with the people who need your model and align with their (model) needs or desired outcomes?

A shared model is an explicit, transferable representation of how a group thinks, decides, or operates; how a technology works, is built, and is monetized; or why a goal was decided upon. It is sufficiently accurate that two people using it independently would describe the group’s decision logic (or another element) in substantially the same way. The taxonomy examines the shared models of five groups:

• G1 Technologists: How technology is built and monetized.

• G2 C-level leaders: How executives decide what to fund.

• G3 Business units: How operating units create and capture value.

• G4 Customers: How customers determine what they need and will pay for.

• G5 The marketplace: How competitors, partners, and market forces behave.

For each group, a shared model can fail in one of three ways:

• F1 Don’t understand: No accurate model exists. Assumptions substitute for evidence.

• F2 Can’t work with: An accurate model exists, but it remains trapped with the person or team that created it. It does not reach the decision and action owner in a usable form.

• F3 Can’t align: The model exists and has reached the decision and action owner, but strategy, implementation, or execution still diverges from it.

These failures must be diagnosed sequentially. First determine whether an accurate model exists. If it does, determine whether the relevant decision owner can use it without its author being present. Only then determine whether strategy, implementation. and execution align with the model.

Diagnosis stops at the first failed test because downstream failures are not yet actionable. An organization cannot translate a model that does not exist or align execution to a model that never reached the decision maker.

The intersection of the five groups and three failure types produces 14 valid bottleneck categories. There is no G5/F2 category because the marketplace is not a counterparty with which the enterprise directly works. It is a system the enterprise must understand and align itself to.

The taxonomy classifies a condition at a particular moment rather than assigning a permanent organizational trait. Bottlenecks recur as customers change, executives turn over, technologies evolve, and markets move. A model that was once accurate can decay into F1, become trapped in F2, or drift away from execution in F3.

That’s why transformation must be continuous. Once the causal bottleneck is identified, it becomes the input to the MAIT cycle:

1. Model: Build the shared model of the group, goal, or technology that isn’t understood.

2. Apply: Put the model to work with the group that owns the decisions and actions, and measure the outcome.

3. Improve: Iterate the model until the outcome that the business needs shows up. This is where the cycle goes from removing bottlenecks to delivering higher value more efficiently.

4. Transform: Scale the proven model across the enterprise so the advantage isn’t local. This is where fixing a bottleneck becomes a capability that creates competitive advantages.

Model, Apply, and Improve create an information flywheel implemented with WITA (Work, Information, Transparency, Augmentation) cycles that produces Decision Dominance: people across the enterprise make decisions from the same current models faster than the competition can.

Continuously running the full cycle produces Transformation Dominance: the organization removes bottlenecks faster than changes in its customers, market, or operations can recreate them. It pulls ahead of competitors faster and further with each cycle.

Examples Of The Bottlenecks From The Last Few Days

News Event: Stardust’s product contradicts its privacy promise.

Causal Cell: (G4, F3) Customers × Can’t align

Diagnostic Result: T1 passes; T2 passes; T3 fails

Stardust explicitly positions the product around privacy, so the customer model exists and reached product leadership. Yet Mozilla found the app sending sensitive reproductive-health data to an analytics provider. A fully informed actor building for the stated customer requirement would not ship this implementation. This is an unusually clean F3 example.

News Event: KPMG Australia faces major cuts after its audit-leak scandal.

Causal Cell: (G4, F3) Customers × Can’t align

Diagnostic Result: T1 passes; T2 likely passes; T3 fails

A professional-services firm unquestionably understands that clients buy confidentiality and trust; those requirements are institutionalized through professional standards. Yet KPMG allegedly mishandled confidential information and the subsequent whistleblower process. The model was the, but the operating behavior diverged from it. The reported consequences include potentially hundreds of job cuts, partner-pay reductions, and senior departures.

News Event: TV Time shut down after failing to monetize a large user community.

Causal Cell: (G4, F1) Customers × Don’t understand

Diagnostic Result: T1 fails

TV Time had millions of users but no accurate model of which customers would pay, what they would pay for, or how the service could capture enough value to remain viable. The company’s own explanation says the free model was unsustainable and demand for a paid version was insufficient. Large-scale usage was mistaken for monetizable demand. Sound familiar? Some of the most common failure modes in tech, especially for startups, are G4, F1.

News Event: OnePlus withdrew new products from North America and Europe.

Causal Cell: (G4, F1) Customers × Don’t understand

Diagnostic Result: T1 fails

OnePlus originally had a precise customer model: enthusiasts wanting flagship specifications at a midrange price. Its range expanded, flagship prices rose, U.S. share dropped below 1%, and demand weakened across key markets. The original model appears to have decayed without being replaced by an accurate current one. This is also a strong illustration of a once-correct customer model reopening or regressing into F1 as the company and market changed, but the model failed to evolve.

News Event: Anthropic and Blackstone formalized a $1.5 billion enterprise-AI implementation company.

Causal Cell: (G1, F2) Technologists × Can’t work with

Diagnostic Result: T1 passes; T2 fails

This is evidence of a bottleneck market forming, rather than an enterprise failing. Capable models and technical knowledge already exist, but enterprises cannot reliably turn them into operating systems and workflows without embedded engineers carrying that knowledge into the business. Ode’s forward-deployed model almost literally satisfies the F2 description: the technical model travels only if its author accompanies it.

News Event: Vint Cerf joined an effort to establish identities for internet-based AI agents.

Causal Cell: (G1, F1) Technologists × Don’t understand

Diagnostic Result: T1 fails

There is not yet a shared model of what an agent’s identity signifies, where its authority originates, what commitments registration creates, or who is accountable for its behavior. Multiple standards are emerging, and two technologists could currently describe the necessary trust model differently. This is an ecosystem-level instance of G1, F1 rather than a strictly internal enterprise case.

News Event: Apple’s newly approved launch with Alibaba and Baidu.

Causal Cell: Resolved (G5, F3)

Diagnostic Result: Success

Apple knew China required a localized technical and regulatory strategy, but Apple Intelligence remained unavailable while implementation failed to conform to that marketplace. The partnerships operationalized the model sufficiently to obtain approval.

Deep Dive: IBM’s Big Miss On Revenue & Growth

Causal Cell: (G4, F1) IBM lacked an accurate, current model of how customers would allocate budgets, sequence purchases, and delay commitments under rapidly changing infrastructure and cybersecurity conditions.

The Narrative: Arvind Krishna’s statement that IBM “did not adapt and move quickly enough” makes the miss sound like G4, F3: execution failed to align with known customer needs. But he’s doing that to marginalize the true extent of the strategic failure.

The taxonomy’s ‘sequentiality’ rule prevents that conclusion. IBM also admitted that it underestimated the magnitude of the customer-spending shift. That means the customer model itself was inaccurate. Execution cannot be evaluated against a model IBM did not yet possess. The failed deals are evidence of F1 propagating downstream.

IBM also appears to have had a marketplace-modeling failure G5, F1. It underestimated the combined speed and effect of:

• AI-driven demand for servers and memory

• Infrastructure supply constraints

• Anticipated component-price increases

• Rapidly evolving cybersecurity threats

• The resulting competition for enterprise technology budgets

This is G5, F1 because IBM did not accurately model how these market forces would change customer behavior. However, G4, F1 is the more immediate diagnosis. The revenue disappeared because customers made different purchasing decisions than IBM expected.

The diagnosis would change if IBM’s account teams already knew customers were moving budgets but that knowledge did not reach forecasting, product, or executive decision makers. That would be G4, F2: Customer understanding existed at the edge but failed to reach corporate planning in a usable form.

IBM’s July 22 earnings call may clarify this. The question to ask is, ‘Did IBM’s sales teams fail to see the spending shift, or did they see it and fail to change the corporate forecast?’ If they did not see it, the diagnosis remains F1. If they saw it but leadership could not act on it without individual account teams carrying the information, it becomes F2.

The MAIT Implication

IBM’s Model step should produce a dynamic customer-capital-allocation model vs a general demand forecast. It must represent:

• Which budget pools IBM products compete for

• What events cause customers to reallocate those budgets

• Which purchases substitute for or delay IBM transactions

• How supply constraints and expected price changes alter timing

• How cybersecurity events override planned buying sequences

• The probability that major transactions slip under each condition

The fundamental failure was not that IBM lacked customer relationships or sales opportunities. It lacked a sufficiently current model of the decisions customers would make when their constraints suddenly changed. This is why the Information Flywheels are so critical. Models must be continuously updated, or your stock drops 25%.

Your New Superpower Is The Future Of Work

You can now diagnose and communicate any bottleneck up or down to get buy-in for what needs to happen to resolve it. You also understand the approach to resolving each category of challenge. This is the Rosetta Stone for business failure modes and the top-level blueprint for putting a managed process in place to address each one.

This is also a view into the future of work and emerging roles. It’s no longer enough to just be a technologist. To get past the senior-level, you must be a technologist who can remove at least one of the three bottlenecks across multiple groups.

This is more than me pitching my courses (green) and V Squared’s services (purple). It’s cool to be ahead of where jobs are going, but I want you to see it too.

The pattern underneath all of it is that new roles are emerging, but not because the technical work got harder. AI is collapsing the cost of the technical work, and when the “doing” gets cheap, the value moves to the boundary-crossing. Every new job title is the market looking for a solution to a specific bottleneck cell. The technical skills are table stakes; the bottleneck you remove is the new job.

Emerging Roles On The Boundaries

The Forward Deployed Engineer removes the customer–technologist bottleneck at all three failure levels. The FDE embeds directly inside the customer’s organization, and the first thing they build isn’t code. Palantir calls it a customer-specific ontology that grounds the system in the customer’s own nouns and verbs.

That’s the G4, F1 modeling cell. They construct an accurate model of how that customer actually decides and operates. Since they’re embedded rather than handing a spec over a wall, the model never gets stuck in translation. G4, F2 is removed by physical presence.

Since they ship and stay accountable for production code rather than delivering a report and exiting, the build aligns to the model, which is G1/G4, F3. Then they route what they learned back to Palantir’s platform, which is MAIT’s Improve→Transform loop.

This is bottleneck-removal that goes beyond engineering. OpenAI created the FDE role to solve a specific bottleneck: customers stuck bridging trial to production. The analysis linking it to enterprise AI failure is proof that the models aren’t the problem; the deployment is. The FDE exists because the 95%-of-pilots-fail number is a modeling-and-translation failure, not a technology failure.

Andreessen’s ‘builder’ role removes the translation bottleneck inside the product org. The “three-way standoff” he describes among engineer, designer, and product manager, each now believing they can do the other two, is really a description of three groups that have always generated translation failures at their shared borders. That’s the design intent that doesn’t survive the handoff to engineering or the customer insight the PM can’t get built as specified.

Collapsing them into one person deletes the handoffs where meaning leaks and bottlenecks form. That’s why LinkedIn’s move to a single “full-stack builder” role was described as eliminating communication bottlenecks. The builder is an F2-removal role aimed at the internal build team, and the reason it’s viable now is that agents can competently fill the gaps across all three functions.

That’s also why Anthropic’s own Claude Code creator predicts the software engineer title fades toward builder. Once you see the pattern, more of these roles become obvious. You just need to map the failure modes and let the taxonomy do the predicting.

The Analytics Translator is another case. McKinsey coined it as a bridge between the technical expertise of data scientists and engineers and the operational expertise of frontline managers. That’s a G1↔G3 translation role with a projected demand of two to four million in the US alone.

It exists because roughly 85% of data science projects fail, most often from poor business alignment. Practitioners know it’s a skill set rather than a job title, which is what you’d predict for a role that is nothing but bottleneck-removal with no technical deliverable of its own. That’s why I don’t see Analytics Translator surviving as a stand-alone job. It will get folded into a different category of ‘builder’.

The AI Product Manager is described in the same language. It’s a “translator” role coordinating engineers, designers, C-level leaders, and business units. That’s an F2 role sitting on the G1, G4 line between what the model can do and what ships.

Design Engineers and Product Engineers are the builder’s narrower cousins, removing the design-to-engineering translation tax. Solutions and Deployment Engineers are the FDE’s lighter-weight relatives. They are customer-embedded, but a step down in bottleneck coverage because they don’t own production code, so they clear G4, F2, but leave G4, F3 partly open.

That gradient is a taxonomy prediction. The more cells a role closes, the more valuable and the harder to hire, which is why FDE compensation is as high as it is.

When AI commoditizes the labor, a job title stops emphasizing the labor or artifacts it produces and starts describing the boundary it dissolves. It explains the roles that have already emerged because each one maps to a hot, persistent cell.

It predicts where the next ones appear. Wherever a bottleneck cell is both expensive and stubborn, a role will form to remove it. The G2 executive cells are the obvious next frontier. The “AI translator to the C-suite” doesn’t have a clean title yet, but the taxonomy says it’s coming.

And it’s the one I built my most recent course (Executive Presence and C-Level Influence) to address.

The Move Towards Outcomes

Just as businesses are increasingly expected to deliver outcomes, employees are too. The larger shift in the market is hiring people who deliver more complete and complex outcomes instead of more complex artifacts. For decades, engineers were promoted based on the complexity of what they could build, but times have changed.

Engineers are now advancing based on the value of what they can build. That’s why removing bottlenecks has become a critical capability set. That’s the top-level theme. We get paid for outcomes, so we have become outcomes engineers. FDEs and ‘builders’ are just another name for an engineer who is expected to deliver the business or client’s outcome from 0 to 1.

Satya Nadella published a long post on X this weekend that got 3.7 million views. He coined the phrase the ‘Reverse Information Paradox’. His thesis is you pay for AI twice, once with money and once with the proprietary knowledge you have to reveal to make the model useful.

He borrowed Alex Karp’s demand that customers should own the means of production, and Karp made the case on CNBC. He told enterprises that token-metered AI was a deal where “something has gone completely wrong.” You pay for the tokens and the labs keep your IP.

Marc Benioff has been saying a gentler version of that for a year. He says that your data is not Salesforce’s product. Salesforce sells an agentic platform, but Benioff is positioning its Data Cloud as the enterprise system of record for everything that touches the customer without transferring that data into Salesforce’s hands.

All three are dancing around the same perception shift. Models are commoditizing, and the moat is moving from the model to the proprietary knowledge that makes models and agents valuable. Nadella calls that knowledge tacit, stored in private evals and corrections. Benioff calls it context and grounding. Karp calls it the ontology and the alpha. It’s the same asset and competitive advantage with three different brands that suddenly find they have a common interest.

However, they are only describing the easy half of the problem. It’s enough to get you to buy the platform, so that’s where these vendors stop. Unfortunately, this partial narrative is dangerous to businesses that need to get value out of their information and AI investments. It hides the additional work required to actually get what those vendors are promising.

The One Assumption All Three Of Them Share

The same assumption hides underneath each narrative: that the knowledge already exists. It is sitting in your datasets and your employees’ heads right now, valuable and intact, waiting for you to wall it off and monetize it.

‘Protect your context’ only makes sense if you have context worth protecting.

‘Own your learning loop’ assumes the loop is already producing something.

Nadella’s term for it is revealing. He calls the valuable byproduct exhaust. It’s the traces, corrections, and evals that leak out of the engine as you work. Exhaust is what you get when you are not capturing information intentionally.

Here is what I see working with a range of clients, and others I talk to have a similar experience. Most start with datasets that contain almost none of the information that would make an agent valuable. The transactions are all there, but the context that explains them is not.

You have a record that a deal closed at a discount, but nothing about who pushed for it, what the customer threatened, which competitor was in the room, or whether the rep would do it the same way again. You have the output of a thousand decisions without the reasoning behind any of them. The knowledge the three CEOs want you to protect is rarely captured because nobody engineered the systems to do it.

That is the part the vendors have every incentive to understate. “You already have the context, we’ll help you unlock it” is a much easier sale than “your pipelines throw away the most valuable thing they touch, and we need to rebuild them.” The first is a platform purchase. The second is a strategy and architecture problem. Guess which one shows up in the keynote.

Context Is Manufactured, Not Found

Most of my engagements start with the unglamorous work of re-engineering data pipelines so they gather data contextually in the first place. Context requires capturing the decision, correction, reason, and the outcome alongside the transaction, rather than the transaction alone. Then turning that captured context into a knowledge graph and structural causal model an agent can reason over, instead of a warehouse it can only query. The raw material of the moat is a byproduct, and byproducts only accumulate when you build the machine to collect them deliberately. Left alone, they vent into the air like exhaust.

This reframes what the high-value information capability actually is. Protecting information is a control problem, and it is largely solved with zero-retention tiers, tenant boundaries, and on-prem weights. That is the platform the three of them are selling.

The capability that separates businesses over the next decade is acquiring information, especially new information. That means getting good at:

Engineering access to the processes and workflows that generate high-value information

• Designing the decision workflow so the reasoning is transparent

• Capturing corrections the moment a human overrides the model

• Running experiments that tell you something you did not already know

• Wiring the outcome back to the action that caused it

Every new information set you can manufacture this way makes the business more valuable. When the information set is genuinely novel, something no competitor holds and none can easily reconstruct, it stops being an asset and becomes an advantage. That is where the moat gets built. Not by protecting the context you have, but by engineering the flow of context you do not have yet and protecting that engineering work from being easily copied by competitors.

Information Flywheels Are The Strategy

This is the piece Nadella hints at with his fifth C, Compound, and then leaves as an incomplete thought. Compounding is not a property you declare. It is a machine you build and an iterative action: The Information Flywheel.

The mechanics are simple to state and hard to implement, so I can’t blame them for stepping over them. You engineer a workflow so that using it generates information you did not previously have. That information feeds back into the system (the knowledge graph, structural causal model, evals, and models) and makes the next use more valuable, which drives more use, which generates more information.

Turn the wheel and the asset compounds, and the gap between you and a competitor who started later widens as the product or agent’s usage scales. The three CEOs describe the asset sitting at the end of The Information Flywheel. It is the difference between owning a reservoir that’s stagnating and owning the mechanism that continuously replenishes it.

Which is why the real strategic question is not the defensive one every vendor is currently posing: How do we keep our data from leaking to the labs? Enterprises must go on offense to generate more revenue: What flywheels can we build that our competitors structurally cannot duplicate easily?

Protecting a static asset has become table stakes, and that’s what most vendors are afraid to confront. Manufacturing compounding information architecture is the endgame. You can patch the leak, and still have nothing worth protecting. You can win the second and make the first fight irrelevant, because a competitor who copies your data as of today is copying a snapshot that has already moved.

Information Flywheels are critical drivers of decision and transformation dominance.

Where This Lives: The Learn Layer Of Orchestration

In the agentic orchestration architecture I’ve been using and writing about, this capability maps to one of four layers: Represent, Decide, Act, and Learn.

Everything Nadella, Benioff, and Karp describe lives in Represent. Grounding context, protecting the ontology, owning the semantic layer, and keeping the evals inside the boundary are all the representation problem. That is the right place to start. It is also the place most enterprise AI work ends, which is why this narrative is so dangerous.

Information Flywheels live in Learn. It’s the layer that turns action-outcome feedback into information the graph did not contain before. Represent protects the knowledge you have. Learn manufactures the knowledge you don’t. One is a shallow moat you inherited from BI. The other is a moat you compound to enable agents and AI. Learn is, by a wide margin, the hardest of the four to build correctly, for a reason most teams discover only after they have shipped something that looks like it is learning, but isn’t.

That is where I pick up in my next article. For now, look at your own stack honestly. How much of it is built to protect what you already know, and how much of it is built to manufacture what you don’t? Have you completed an information assessment that details what context exists in your datasets? Have you done a gap analysis between what you have and what agents will need to support a workflow and deliver an outcome?

If your business is like most that are early in their information and AI maturity journey, there isn’t enough context for monetization to be feasible. It is better to understand that upfront than to build high castle walls around data that isn’t worth protecting.