The race for specialized intelligence

Yesterday morning, Town announced a $108 million Series B and a billion-dollar valuation based on this idea: every employee should have an AI assistant plugged into their email, docs, calendar and systems that can learn their people, projects, and voice. The news comes a week after Ramp announced they’re launching Glass, their internal AI workspace, to other companies under the tagline: AI made every employee faster, Glass makes the company smarter.

In the past year, other companies have also started circling around this problem. In February, OpenAI launched Frontier, its platform for AI coworkers "grounded in their company's context." Glean, which began as enterprise search, started describing its job as connecting models to a company's systems, people and knowledge. Last week, Uvi raised a seed round to build a "reasoning layer" for agents on the premise that models will come and go but a company's context won't. And the most recent YC batches are full of two- and three-person teams whose entire pitch is some version of "we learn how your company actually works."

There are many different companies coming from many different directions, but they’re all running at the same problem: in 2026, every company wants to bring agents into its workflows but can’t get past one thing… the agents have no idea how their companies work.

While agentic automation has raised the stakes of knowing how your company works, the problem has always been around. When I was at McKinsey, one of the first things that became clear to me was how little visibility companies have into their own workflows. I was one of an army of consultants that businesses paid millions for, all so we could sit next to their employees and write down how work gets done. We did, diligently, but by the time we left the client site, the work had already changed. Even in businesses people are quick to dismiss as slow-moving, ways of working don’t hold still. 

‍

—

‍

This is the problem I’ve thought about for the last decade: the illegibility of work. When we started Scribe in 2019, long before companies were trying to balance their AI ambitions with their token budgets, we set out to address it by giving employees easy tools to capture what they do without having to stop and write it down. It was the thing I’d spent my twenties doing, but fast, shareable, and inexpensive. 

Then the frontier models arrived, captivated business leaders, and rapidly improved. And with every improvement, the stakes of the problem I’d dedicated my life to grew. 

Ramp, Town, Glean, and the startups behind them are all seeing the same thing. What was easy to dismiss as a technology constraint (and once legitimately was) no longer is, and the deeper, more pernicious challenge of transforming work is now in every company’s direct line of sight. Businesses are burning through tokens but getting no leverage, and massive investments in AI transformation are resulting in uneven gains at best. We’ve stepped into uncharted territory, where intelligence, something that businesses used to work tirelessly to capture by hiring experts and the best employees available, is a common resource. The models you use know almost everything imaginable, but your competitors know too. To claim the economic advantage these technologies promise, businesses are rushing toward being first movers, but they are still ultimately moving toward the place all their peers will end up too. General intelligence is everywhere, and it won’t set anyone apart. 

The thing that will differentiate companies is what the models don’t know: how each idiosyncratic business works. The catch is that for it to be an advantage, companies have to know it themselves.

A frontier model can pass the bar exam, calm a disgruntled customer, and absorb every playbook in existence, but it is lost when the right call is to ignore the playbook, something your best employees do instinctively. AI can read your documentation (if it's current) and do a plausible version of basically any job you can describe with rules and success conditions, but it cannot do it in the way only your company does. Not until it knows what you actually do. 

Town, Ramp, Glean, Uvi: they're all trying to bottle the same thing—specialized intelligence, the unique know-how that lives in the daily choices, workarounds, judgment calls, and exceptions that define how a company operates and that can’t be found anywhere on the internet. In an age where information asymmetries are becoming near-impossible to claim, what's left is the knowledge that can only be learned from the inside: how your particular organization turns work into outcomes.

The data already in your systems is part of that intelligence. But looking at the data alone is like trying to prompt an agent with sentences that only contain nouns. Specialized intelligence is the verbs: how a good rep qualifies a deal or finance closes the month. It lives in the ways your people work. 

Town describes it the same way I’ve always thought about it: this is the context "trapped in individual heads and inboxes," and it should be accessible to the company that built it.

We are watching the birth of an industry whose promise is to make the way work happens legible, to people and agents alike. The question that will decide the race, and the one I've spent a decade on, is where that knowledge lives and how you get it out.

‍

—

‍

So far, every approach to solving this problem can be sorted into one of four categories: ask, infer, attempt, capture.

Ask is the oldest method in the book. It’s also the one most companies are already using AI to do. You turn SOPs into prompts, wikis into skill files, training decks into custom agents, and send them to work with all the answers to all the questions you thought to ask. Most transformation leaders I talk to are standing up teams to do exactly this. It's the smart place to start, but it's also where the limits of companies’ current knowledge show up first. People are poor witnesses to their own work. Ask a subject-matter expert how they do their job, and they will tell you what they think is worth telling, which is, at best, the most common version of their workflow. That most common version is rarer than anyone assumes. In our data, built over seven years of helping companies capture how they work, the most common way of completing a given task shows up only 7 percent of the time. Asking employees to describe their work so AI can read the description is a lossy process full of stale information. The companies that make it work have uncanny discipline and a culture of writing. But even for them, the picture they get is incomplete. 

Infer will likely be the most common approach and the place where I expect more companies to enter the specialized intelligence category. It’s a smart approach because the data exists and you just mine the exhaust: emails, chats, docs, calendars, tickets, CRMs, and every exchange an assistant has. Then you let an agent learn from it. This is Town's model, and Glean's and Rovo’s, and most of what Copilot does. It's the fastest way to surface how a company works, and if I were starting an assistant-only company today, it's where I'd probably start too. The challenge is that exhaust is still only a sliver of how work happens. Exhaust records the inputs and the outputs, but the actions in between mostly leave no trace. Institutional knowledge lives in the steps employees skip as much as the steps they take, and most of those never generate a record. Inference works on nouns. It struggles with the verbs. It also over-weights whoever writes the most, inherits contradictions in your records, and can't tell how something was done. It's a hard problem, and I'm curious to see who solves it.

Attempt is the newest approach, and the one AI advances have only just made possible: let the agent try the job and learn from correction. This is the computer-use bet, and the one the labs are most excited about because, in theory, it needs nothing from you and the most from them. In practice it requires an unlimited token budget and high tolerance for mistakes inside production systems, and most companies have neither. 

Capture is the simplest in premise and, I think, the most promising. Get the work itself, at the moment it happens, from the person doing it. It's what consultants with notebooks have done for decades, except now the notebook is free, can sit with multiple teams at once, and the AI that reads it produces structured knowledge instead of a 90-page appendix. Capture is the only one of the four approaches that gets the actual process rather than a description of it, a residue of it, or a guess. It's also where the newest entrants seem to be heading: the last few YC batches have produced at least three companies that hope to help agents understand work that has never been written down.

Unsurprisingly, capture is the one I'm bullish on. It's the bet we made at Scribe in 2019, years before there were agents to hand the knowledge to. And the reason it made sense then is the reason it makes sense now. 

It also has to be done well. Capture that monitors people instead of workflows will not be accepted, and a company that rouses fear in its own employees will learn nothing from them. That story played out at Meta this spring, when the company installed software on employee laptops that logged keystrokes, mouse movements, and screenshots to train its agents. More than 1,500 employees signed a petition, and within weeks Meta was walking parts of the program back. The products that win here will be built with the employee in mind, because the person doing the work is the one who has to want it understood.

These four aren't mutually exclusive, but they also are not equal. For companies without infinite budget and energy, the first three are proxies that fail in the same way—producing confident agents that only perform the documented version of a job that no human actually does. Only the fourth corrects for it because it's the only approach that learns where the work is actually happening.

‍

—

‍

As models stop mattering as a differentiator, the race will go to whoever builds the deepest, truest picture of how a company actually works, and whoever that company trusts to hold it.

That second part matters as much as the first. Specialized intelligence is the most valuable thing a company owns that isn't on its balance sheet, and for the first time it can be captured, kept, and handed to the tools and people that need it. 

I started working on this problem when it was a consultant problem, but companies should see the emergence of new entrants to this category for what it is: a clear signal that claiming your specialized intelligence is about to become the most important infrastructure problem in enterprise software. And, if you run a company, the practical implication is simple. Don’t spend this year agonizing over which model to pick; that decision will get cheaper and easier every quarter. Spend it finding out how your work actually gets done.

‍