Generative AI Agents Are Powering A Wave Of Disruption And Most People Won't See It Coming

Bill Gates isn’t known for his hot takes, but he dropped a fireball this week. Generative AI could put Google and Amazon out of business. It’s a bold statement, and it’s difficult to see the market playing out that way. Most people agree that Generative AI tools will be disruptive, but that statement refers to disruption on another level.

It’s not the first time I have heard someone say that Generative AI would take down Google. In 2019, I attended a talk where the speaker made a similar case and said Google would no longer be the top name in search within 5 years. They went even further and said it was likely that Google would go under because of that disruption.

Generative AI was more mature than you probably remember in 2019. The original GPT and BERT had been out for about a year, and the underlying architecture was maturing rapidly. There was enough clarity for some to see the implications for search business models. Oddly enough, even though Google had BERT in 2018, it didn’t move to implement Generative AI tools until another company moved first.

The end of Google was a fringe hypothesis. I remember mentioning it on a video call with Demetrios Brinkmann in the Spring of 2020. We were preparing for my upcoming talk on the path to production and monetization for the MLOps Community. His response was stunned silence. The case was compelling enough to keep it in mind for the next 3 years.

Generative agents are the beginning of the cycle Bill Gates sees coming. This post will explain why Bill Gates is mostly right with this take and how the cycle will play out over the next 12-24 months. I’ve broken the post into multiple parts because it’s a long, somewhat dense read.

This article is focused on workflows, so you will take away enough information to develop more than just a product strategy. If you see the pattern of workflow impacts, you’ll be equipped to identify opportunities and design products in this category.

Gates is referring to a specific class of Generative AI products. For the next five years, agents or intelligent assistants will be among the most disruptive product categories. While the product isn’t new, models like GPT give them new capabilities and broader utility.

Amazon has already signaled it intends to go into generative search. Scanning through recent job postings reveals they’re hiring with a new vision for their internal search products. This vision includes integrating generative models and changing how people interact with Amazon Prime. Amazon’s vision extends to intelligent assistants for marketers. The marketing assistant product Amazon envisions will help people build video and image assets for ads.

At this stage, it’s unclear how Google will respond, but it’s hard to imagine Google’s leadership doesn’t see the implications of intelligent agents on its search business. The problem for Google is its current business model. Intelligent agents don’t easily lend themselves to an advertising business model. Let me explain why.

What Is An Intelligent Assistant, And What’s The Need They Fill?

An intelligent assistant is a product that leverages advanced models to deliver expert knowledge and answer questions. This product category will dramatically reduce the cost of accessing expert advisors. Many search use cases involve people looking for expert knowledge and advice, which is why this product category will be so disruptive.

Legacy search portals are search engines or the search bar at the top of every application. While the text box allows users to enter whatever they want to, we all know that natural language search is beyond its capabilities. As a result, most search use cases are poorly met. We’ve all gotten used to shortening and abbreviating what we need in search bar lingo.

Engaging with an expert is simpler and more natural. That’s one of the reasons why human assistants are so widely preferred over search capabilities. If I have a choice between talking with an expert software engineer or searching Stack Overflow to solve a challenge that I’m having while coding, I’m always choosing the expert.

Search is great for simple questions, but the more complex the use case, the more complex the search query gets. Complexity also leads to follow-up questions. In the current search paradigm, that means multiple searches and frequent dead ends.

Search solutions also create an interruption in most workflows. In the coding example, I must leave the IDE and go to Bing, Google, or Stack Overflow to find answers. We are desensitized to workflow disruptions and context switching, but every time we hop from one app to another, the productivity hit is significant. It’s the same when we must search while writing emails or any other content.

In the Amazon app, the product search functionality is better integrated into the shopping flow. If I’m not finding what I need, the search bar is always at the top of the screen. I can initiate another product search without navigating somewhere else or switching apps.

However, the search bar’s capabilities don’t extend beyond the product catalog. Shoppers often need more than product information to make a buying decision. That requires users to transition to a different app like Google or Bing. There’s also a matter of trust. Which link do I trust to give me accurate and unbiased information? Search results for the best men’s t-shirt or beard trimmer are filled with affiliate marketing disguised as reviews.

It can feel like our search needs are well met. If you evaluate the workflows, you’ll discover that few truly are.

How Do Intelligent Agents Change The Search Workflow For The Better?

Replacing search with intelligent agents or assistants will enable the Super App paradigm. Super Apps have functionality so broad and generalized that users don’t need to leave them as we do now with most apps. Think of every workflow requiring multiple applications and/or web pages.

While I was in Orlando last week, I bounced between apps on the way to the airport. I started in Uber to book my ride, and the app asked me for the terminal I was departing from. I had to go to the United app to find that out and return to Uber.

An intelligent assistant allows you to consolidate this type of workflow. For this use case, it would be a single app that handled all things transportation. I ask the intelligent assistant to get me a ride to the airport. It handles everything after that. In my example, I was being lazy. I didn’t check multiple ride-share apps for the best price or fastest pickup. An intelligent agent would.

The application ecosystem that’s being built inside ChatGPT is a good example. OpenAI’s goal is to enable complex workflows within a single app. When ChatGPT’s ecosystem fills out and becomes more functional, it will handle more complex search use cases. The demos have shown a user prompting ChatGPT for a recipe, selecting their favorite from the results, asking to see the ingredients, and using Instacart to get them delivered.

The night before I left Orlando, my American Airlines flight plans were disrupted by a delay. I would miss my connection and needed to book a new flight out. It was a painful process that involved multiple searches, trips to different airlines’ websites, and an attempt to get someone from American Airlines’ customer service team on the line.

As intelligent agents become more capable, that complex workflow will get easier. I will give the agent instructions like we create prompts today.

“Find me a flight option from Orlando to Reno, leaving tomorrow morning and getting me into Reno before 5 pm. I prefer to switch my flight on American to another one on the same airline. United and Delta are my next two preferences if that’s not possible. If nothing else is available, I will take a Southwest flight. I prefer flights with aisle and window seats, but I’ll take a middle if it’s all that is available. If the flight is less than $500, book it. Otherwise, ask me before booking. If you don’t find any options on the first attempt, run new searches with the same parameters every 15 minutes. If I’m asleep, you may wake me to ask about booking a flight option that costs more than $500.”

It’s a lengthy, detailed prompt. Still, spending 2 minutes writing or saying it is far less painful than what I had to go through to get a new flight. This functionality is a few months away from being a reality, and I have seen a few early proof of concepts that handle the workflow.

My prompt can be mostly handled by the ChatGPT-Bing integration as they work today. ChatGPT converts the prompt into a search query, interprets the results, and iterates through the filter conditions to execute the entire request. The only things missing are integrations with the airlines for seat selection and booking.

In the next 12 months, intelligent agents will be developed to handle more complex use cases. I won’t need to use the American Airlines, United, or Uber apps directly. I will access them indirectly through the ChatGPT (or some other intelligent agent) ecosystem. The simplified workflow cuts Google and its ads out altogether. Creating an intelligent agent will challenge Google to transform its search business model to fit the new workflow.

This represents one use case class for intelligent agents: managing content discovery and consumption at scale. Google is so useful because there is no way for us to distill the Internet to find what we need. It is a master aggregator. For many people, that small search bar is their only portal to the Internet and single means of interacting with its immense complexity. Intelligent agents will become the new portals and make content consumption at scale even easier.

Enabling New Search Classes

Several more use cases deal with content discovery and consumption at scale that are entirely unserved. Amazon’s product search is a good example and a core reason they’re focusing on generative search. They want to be the Super App for anyone who needs to shop and discover products. Amazon wants to be equally ubiquitous as a tool for advertisers. If they can win at product search and ad generation, that goal is within reach.

Amazon Prime isn’t a single storefront. Under the covers, it’s actually a marketplace with thousands of sellers. Sellers join Amazon’s marketplace to get access to distribution across Prime’s 200 million+ subscriber base. An intelligent agent improves support for current searches and enables a new search class.

Instead of looking for a shirt, jeans, and shoes separately, Generative AI supports a search for outfits. The results present options as a set instead of individual pieces. Smaller models can run on the user’s device and implement personalization without transferring personal data to a company like Amazon. The user’s image gallery can provide examples of the outfits they wear today.

For Amazon, the monetization path extends its current ad business. The model makes 30 outfit recommendations, and companies on Amazon’s marketplace bid for ranking or to have their outfit featured prominently. The intelligent agent provides personalized recommendations and hyper-targeted advertising.

What’s Next?

I have covered the near-term implications of intelligent agents. These use cases can be managed today by existing technology. Products like these will be released in the coming months.

In the next post, I will extend further out and cover what’s coming in the next 12 months. Intelligent assistants will evolve to become advisors. Assistants leverage Generative AI and model chaining to automate more complex workflows requiring some simple decision-making. Advisors use the same fundamentals to prescribe actions. Some prescriptive actions will be built into these products’ automation, while others will be delivered to the user, who will decide what to do next.

Intelligent advisors require retraining or rebuilding models with domain-specific (for expertise) and individual (for hyper-personalization) data sets. I will explain this paradigm in the next article, provide use cases, and detail more workflows.