How To Make The Right Buy VS. Build Decisions & Calculate ROI For GenAI Platforms

This week, Marc Benioff and everyone I talked to from Salesforce repeated the same mantra: “Don’t DIY your AI.” It’s a compelling catchphrase. Companies can step back from all AI’s complexity, cost, and hype. All the promised benefits will be gift-wrapped and delivered with a bow. Of course, it’s not that simple.

Salesforce’s platform supports the operating model and helps manage interactions with current and potential customers. The products a business sells must still be developed internally. Data teams are critical to the business because customer expectations are changing rapidly. AI will soon be expected, and if Apple intelligence is all it’s hyped up to be, Apple will be the company that ramps up consumer expectations.

In this article, I will explain the agentic AI platform because you’ll be a user and builder. That means companies must be able to select and define the ROI of the AI platforms they use. They must also innovate and deliver AI platforms to meet their customers’ changing AI expectations. No matter where you sit, agentic systems will impact your job and business in multiple directions. Getting a handle on what’s coming is a matter of survival.

A Gentle Introduction To Agentic Systems In Production

An increasing number of companies will bring agentic products to consumers within the next 12 months. Customers will want all the technology they interact with to be AI-enabled. Change never takes hold overnight, so many businesses will not see the signs until it’s too late.

That is a Salesforce customer pain point: ‘I need a way to see changing preferences coming and get recommendations for what to do.’  That would require multiple apps and complex data analysis in a digital paradigm. It would require asking the right questions and knowing what data contains the answers in the copilot paradigm.

Salesforce is using an agentic paradigm. We all know this is coming, but do these platforms work, or is the technology too immature to support them? Salesforce set up the Agent Launchpad at Dreamforce because you don’t get how dramatic the agentic shift is or trust it until you see it.

A customer sits down and talks through their problem with a product manager. The conversation is recorded, and an LLM builds an agent to solve it. I lurked in the zone and watched a dozen conversations where customers used natural language to:

• Build the initial version of their agent around their desired outcome.

• Test it with multiple prompts they came up with and generate dozens of prompts automatically.

• Validate the agent’s behavior with JSON descriptions of the resources used to generate the workflow and deliver the outcome.

• Update the workflows, add steps to fill in the blanks for autogenerated workflows, and add resources for the agent to use to deliver the outcome.

• Define limits and guardrails for agent behavior; what actions were off limits.

It looked like it worked in the demos. I talked with a robotics supplier from Australia who used the Agent Builder. He went through the demo, delivered the output to his Salesforce Engineer, and implemented it in a half day. He said it worked and can probably be implemented faster in the future.

Saks saw Benioff preview an agent for their customer service use cases. Their data team said, “Challenge accepted,” and tried to implement what was demoed. They started on Thursday and the agent was running in production on Monday. Benioff updated his keynote the night before giving it because the demo wasn’t theoretical anymore.

After seeing demos fail to deliver for over a decade, I'm a natural AI skeptic. There’s evidence that Agentforce works and will work in product environments. My biggest concerns are with scale and infrastructure diversity. Benioff announced plans to launch a billion agents in a year. Doing a back-of-napkin calculation, the resource usage for serving conversations at that scale is mindboggling. Time will tell if its cloud infrastructure can keep pace with that demand.

A New Pricing Model & ROI Calculation

Marc Benioff's outcomes-based business model emphasizes delivering tangible outcomes through AI agents rather than simply enabling tasks as traditional digital apps do. Task-centric products focus on improving efficiency and productivity. Outcome-centric products require a different approach to monetization—one that aligns with the value of the outcome achieved rather than the task completed.

Agentforce offers AI-powered customer service starting at $2 per conversation, which should be evaluated against the ROI of each conversation.

The Saks example illustrates how outcome-based pricing works. An AI agent successfully transformed a potential product return into an exchange, effectively salvaging a high-margin luxury sale. For retailers dealing in premium goods, the value of a saved sale far exceeds the $2 cost of an AI conversation. We can explain ROI for different conversation categories based on the conversation outcome.

Subscription services like Disney+ can have their paths to profitability derailed by high customer churn rates. If a customer is on the verge of canceling their $13.99 monthly subscription, an agent that resolves the issue retains the customer and generates recurring ROI. $2 spent on a support call saves a $13.99 monthly subscription. That’s a 7X return on investment in just the first month.

The Pitfalls of Productivity & Efficiency ROI Metrics

Calculating AI product ROI based on employee productivity or headcount replacement is shortsighted. For example, AI agents can eliminate customer service wait times without increasing operational costs. However, framing the value of an AI conversation purely in terms of employee time saved ($2 of AI time saves 10 minutes of onshore customer service time) runs the risk of undervaluing the AI's impact.

A narrow focus makes the ROI fuzzier and doesn’t explain the new paradigm. There’s no consideration of the broader business outcomes, such as increased customer retention, higher satisfaction, and long-term revenue growth.

AI agents’ ROI should be based on the value of their outcomes, not their capacity to perform tasks or replace human labor. This shift in thinking enables businesses to capture the actual value of agentic AI, especially in high-margin industries where small customer satisfaction or retention gains lead to significant financial returns.

Making The Right Buy VS. Build Decisions

Businesses must build AI business models and platforms targeting the highest value outcome their data can support. Data is always the competitive advantage, and models are increasingly commoditized. It makes more sense to buy than build for AI that supports internal operations and workflows. The rationale all comes down to data.

Salesforce possesses operational data on how thousands of businesses use its platform, enabling it to train highly reliable AI models to support their workflows. These models generalize effectively across a wide range of workflows and business processes because they are trained with datasets that cover diverse use cases. This scale of data creates a significant advantage for companies like Salesforce, allowing them to offer more accurate and adaptable AI solutions.

In contrast, most companies have access only to their own operational data. It’s valuable but limited in scope. Internal data is critical for understanding a business’s unique processes and workflows. It pales compared to the datasets available to AI vendors with access to multiple organizations’ operations.

The best approach is to customize bought solutions with internal data. This hybrid approach leverages the best of both worlds: businesses can preserve their competitive advantage by applying their domain-specific data and expertise while benefiting from pre-built AI models' scale and efficiency.

Companies can dramatically reduce costs and deployment times by customizing instead of building from scratch. The models are already proven and reliable, so businesses can focus on fine-tuning them to suit their unique operational needs rather than developing AI systems internally. Buying also frees up data team resources to build customer-facing products that create new revenue streams for its customers.

What’s Next?

There are several implications in the move to agentic systems. Data remains the competitive advantage, so companies should decide to buy when vendors have access to more data. The implication is that the data team will no longer manage most internal-facing initiatives. We’re about to see the rise of more AI product teams that drive revenue and business growth. That’s a good thing.

If models are commodities, where does that leave data scientists? Engineering capabilities are more important than ever as we shift focus to delivering models for customer-facing use cases. In the next article in this series, I’ll explore the changing role of BI Engineers, Data Analysts, and Data Scientists. I sat down with the Chief Product Officer at Tableau to get a look at the next wave of platforms. Our roles are changing, and we must adapt with them to remain relevant.

I want to thank Salesforce for partnering with me on this article and making its product leaders so accessible during Dreamforce.