AI Has Booted Five Billion-Dollar Features. What’s Next?

The first time I heard the words “Product-Market Fit” (PMF), I was being asked by Microsoft’s Head of Commercial Marketing if my little product had it. Takeshi Numoto is now Microsoft’s CMO, and he didn’t want to strongly sell Microsoft Graph Data Connect if it didn’t meet a strong demand. I had no idea what he meant, and the meeting didn’t go anywhere from there.

I’ve figured Product-Market Fit out since then, of course. If you’re as clueless as I was, PMF is a kind of “I’ll know it when I see it” phrase. Essentially it just means that you are selling a whole lot of the product, and it’s fairly easy to sell.

There are a few AI features that are selling “a whole lot.” Is that odd? AI has been powering on for three years. So far, only five modules have passed their PMF checks. Only five product types represent the entire multi-trillion-dollar market!

Or you could say that it has only been three years, and there are already five durable products. Let’s check the boot logs: explore what AI features have PMF, what they have in common, and what may be the next billion-dollar industry.

Five features with PMF

There are five breakout features in this wave of AI that every company is chasing: programming assistance, chat companion, internet search, meeting notes, and writing research briefings.

SYSTEM OK: Programming

Programming assistance is the first durably useful product from Large Language Models. GitHub Copilot actually predates ChatGPT by 17 months, although it was in a simpler form then. GitHub Copilot started with GPT-3, fine-tuned on programming languages. It would help complete a line or two of code after you had begun typing it, with a feature we now call “code completions.” Code completions are still part of AI-assisted programming, giving experienced programmers both a way to type more quickly and to correctly use the many functions of little-used libraries.

Code completions progressed to a chat feature after ChatGPT was released. This combined the easily verifiable nature of programming with the value of search–this time applied to the codebase instead of the internet. There were competitors by this time as well, helping users understand how the code worked and helping to diagnose bugs.

Ultimately, chat in the programming domain evolved to “vibe coding” as coined by Andrej Karpathy. When vibe coding, you entirely ignore the code that the AI is writing and only use the Chat feature. As OpenAI Codex and Claude Code progressed, the AI systems were increasingly able to view the error logs and run commands themselves, essentially vibe coding themselves.

It’s clear that vibe coding itself will give way to “specifying.” You can see this in products like Loveable, which creates entire frontend applications from your request. In larger codebases, the GitHub Copilot coding agent can just take tasks off the list and complete them. Other companies will follow suit.

These coding agents won’t be able to do everything; software engineering is not going away. Software engineering is a fairly niche market however. Generative AI would only take off when ChatGPT added a personal interface for everyone else.

PROCESSOR ACTIVE: Chat companion

If you’re reading this blog, I know that you use ChatGPT or a competitor. This was the hackathon-project-turned-breakout-hit of 2022. This first version didn’t have web search capability or integrations into any other system. It was a lot less capable than ChatGPT today!

ChatGPT and its competitors have grown to be many things, but one thing they have remained is a chat companion. I’m using this phrase to describe a broad swath of ability approximately equivalent to a conversation with a good friend. That friend might be able to help you practice a foreign language, describe a concept, or just chat about your day.

OpenAI recently learned a hard lesson about how many people rely on ChatGPT as a chat companion when they replaced the GPT-4o model with GPT-5. Suddenly, the entire internet was furious that the personality of GPT-4o had been removed, as they had been treating it like a friend. Many people had taken several steps beyond “friend” towards “girlfriend” and even “husband.” You don’t have to browse Reddit very long before you learn how seriously people consider their relationship with an AI.

Companies like Replika and Character.ai focus exclusively on this use case, but it’s clearly front-and-center in the minds of folks running ChatGPT, Microsoft Copilot, and Anthropic Claude as well. The original ChatGPT suffered from troublesome hallucinations until this chat interface was connected to a search engine.

DISK INTEGRITY SCAN: Search

Bing added Chat only three months after ChatGPT was released, in February 2023. And then it was the only major search engine with native AI capability for 11 months! ChatGPT gained SearchGPT in October 2024, and Google added Gemini to search overviews in January 2024. Google’s delay here is considered a major blunder, but they are working hard to solve for AI search now.

xAI attempts unique value by being able to search X (Twitter) in addition to the web, while Perplexity is solely focused on web search. There’s a reason this is so popular: it’s very good! Getting an answer quickly is very helpful, and so is being able to ask follow-up questions. It helps that web search has gotten terrible over the last several years. SEO-optimized garbage dominates results to the point where I wonder how much of the open internet is even left anymore.

It’s hard to know just how much AI-powered search has disrupted traditional search, but we do have some information on click-through rates. Organizations like the New York Times report an 8% reduction in traffic from Google, for instance.

Increasingly, AI search is moving on from searching the internet to searching the intranet: the internal documents and communications of businesses. This remains the most frequent use of Microsoft 365 Copilot’s chat feature. I’m not sure if it is the most used feature of all Microsoft 365 Copilot, because that may be AI-generated meeting notes.

INITIALIZING INPUT DEVICE: Meeting notes

The first version of AI-written meeting notes in a major product was Intelligent Recap in Microsoft Teams in May 2023, although some smaller and third-party systems existed first. One of my coworkers led the feature, and I remember him asking me how they should charge for it. The major choices were “Teams Premium” or “the AI SKU” as Microsoft 365 Copilot hadn’t been released yet.

This feature is hugely valuable if you have a lot of meetings; it really saves a lot of time. No one I know would be willing to go back. The incredible thing about meeting notes is that it is remarkably difficult to generate great notes, even for humans that were in the meeting. There is so much nuance and ambiguity in an average meeting! I am still often impressed by how well the notes come out.

After these four, there was a bit of a lull in new scenarios. Skeptics were claiming that AI would never make inroads into real work beyond programming. They were very wrong!

OUTPUT COMPILED: Briefings

Although it is similar to search, I consider researching and generating briefings from the results a new feature with Product-Market Fit. Unlike search where the goal is to find an answer for yourself, the best briefings create new and sharable knowledge out of unstructured information.

Every major AI product has their version of this feature:

• ChatGPT: Deep Research
• Perplexity: Research mode
• Claude: Research function
• Gemini: Deep Research
• xAI: DeepSearch
• Copilot: Deep Research
• Microsoft 365 Copilot: Researcher

Each of them limits how often you can use the feature. Those limits are there because it is expensive to run, but limits also give an opportunity to upsell into a more expensive plan. The demand is there!

What they have in common

These are excellent features, but they are not worth trillions of dollars of valuation by themselves. Are we in history’s largest-ever bubble, or will we find several more features with Product-Market Fit? Let’s examine what these five features have in common:

• Obviously, Large Language Models work with language. This doesn’t have to be a spoken language; there’s a reason we call Python and Java “languages.” This showed up first in programming features, and then features for natural language.
• Increasingly, the most valuable forms of these features are personalized. ChatGPT was the first time in history where a machine could communicate with us in the way we naturally talk.
• Each of these features relies on a large knowledge base. You could look at these features as applying an LLM to a new database: the public internet, a codebase, or verbose chat transcripts.
• They are flexible enough to handle ambiguous tasks. For these scenarios, there are many possible solutions even for fully specified tasks, which is where AI thrives.

The personalization aspect is the most interesting to consider. Chat companions will tell you exactly what you want to hear, while personalized search and research products should consider what you already know. If AI gave the same answers to everybody, it wouldn’t be very useful.

What language tasks are still out there that require ambiguity and benefit from personalization? Quite a few!

READY FOR INPUT: What’s next?

With five examples and common factors, identifying the next billion-dollar market is simply a matter of taking a large database of (somewhat ambiguously defined) tasks, and using some kind of language processing on it. I happen to have a tool for that!

I asked Copilot with GPT-5, Gemini 2.5 Pro, and Grok 4 to each give me several ideas. Most ideas were bad, but a few have already been bouncing around my head lately:

• Business analysis. Especially when you consider the ability of language models to generate SQL, I believe there’s an opportunity to cut out the pre-calculated dashboard step of analyzing data. I personally have never found a pre-calculated dashboard to fully satisfy my curiosity about a topic. Inevitably, it’s missing the bit of data that would allow me to answer the next question. We’ll be able to move directly from clean, organized data to important insights about a business.
• Personalized ads. Advertising is already one of the most sophisticated data industries, personalizing anything they can. I suspect the field will move from deciding what ad to show users towards personalizing the ad itself to your preferences. The TikTok algorithm is already a weapon of mass destruction on attention spans, and it will come for consumerism next.
• Asset production. This product one I wish I had myself to manage translating content to social media platforms. But this goes beyond top-tier poasting, AI should soon be able to turn your talking points and graphics into broad marketing campaigns.
• Scientific data processing. Some scientific fields make heavy use of graduate student brainpower to perform a lot of the work. Scientists will still be setting direction, and they will soon be much more productive with AI. There may not be a lot of money involved, but there is still an excellent opportunity for PMF.

Two of the AIs I asked suggested personalized education, but I don’t buy it. You can already audit MIT courses for free; few people do because you need to supply your own motivation. There’s also too much incentive for keeping universities exactly the way they ar Other AI suggestions that I don’t believe for various reasons are entertainment creation, end-user legal services, supply chain optimization, policy simulation, and B2B purchasing.

The industry is still seeking Product-Market Fit beyond these initial five features. They are changing the world, and more is still to come. It’s not too late for you to get started: pick one of these scenarios that is well-suited for AI and for the market. Subscribe to a top-tier AI and build the Product that the Market is waiting for.