Why AI Agents Can't Run Your Company (Yet): Research Notes on the Missing Layer

Here's a strange fact about 2026: frontier AI models can reason across hundred-page documents, write production code, and hold their own in graduate-level analysis — and yet almost no company has successfully handed real internal work to an AI agent.

Not because the agents aren't smart enough. Because they're geniuses with amnesia. They know how refunds work *in general*. They have no idea how refunds work at *your* company — who approves them, what the exceptions are, or why the rule changed last March.

I spent the last few weeks researching this gap: reading the ecosystem's framing of it, auditing the existing tool landscape against it, and pressure-testing the obvious solution with a panel of frontier models. These are my notes.

In a recent Request for Startups, YC partner Tom Blomfield (co-founder of Monzo) named the bottleneck directly: the blocker to AI automation is no longer the models — it's domain knowledge. Every company's critical know-how is scattered across people's heads, old email threads, Slack conversations, support tickets, and databases. The company functions because humans vaguely remember where that knowledge lives and how to apply it.

AI agents can't operate like that. Blomfield's proposed primitive is a "company brain" — a system that pulls knowledge out of fragmented sources, structures it, keeps it current, and turns it into what he calls an *executable skills file* that AI systems can act on. Crucially, he's explicit that this is not company-wide search or a chatbot over documents. It's a living map of how a company actually works: how refunds get handled, how pricing exceptions are decided, how engineers respond to incidents.

Once you look for it, the entire problem space splits along one line.

A knowledge base a human searches answers the question "how do we handle late refunds?" — to a *person*, who then decides and acts. A process representation an agent executes gives the *agent* the conditional logic itself: *if the order is over 30 days old, the lead approves, unless it's an enterprise account with an open incident* — so the agent can take the action.

Retrieval ends at an answer. Execution ends at a completed task.

Every tool on the market today is excellent at the first and absent at the second. That's not a criticism of those tools — retrieval is genuinely valuable and genuinely hard. But it explains why buying more knowledge software has never made a company's AI agents any more capable of doing the work.

The lazy answer is "nobody thought of it." The real answer is that extraction is brutally hard in ways that only show up when you look closely at what company data actually contains.

The logic isn't in the text. When a manager writes "yes, refund them, they're enterprise tier," the *real* reason — the customer threatened to churn on a call an hour earlier — never made it into Slack. Chat logs contain deltas and exceptions built on unstated human context. A model can confidently synthesize a clean rule from that mess, but a plausible rule is not the canonical rule.

Policy vs. precedent vs. exception. Company data can't tell you whether a past decision was official policy, a one-time judgment call, or a mistake that got repeated. There are no labels for "this was correct" or "this was overridden." A citation to a wrong decision is still wrong.

Contradiction is the norm. Policies drift. Different managers answer the same question differently. Half the history reflects a rule that changed six months ago. Detecting that a new message genuinely *contradicts* an old rule — versus being an exception, a joke, or a one-off — is a subtle inference problem.

The category error. Perhaps the deepest issue: extraction treats a *normative* question ("what should we do?") as a *descriptive* one ("what did we usually do?"). Often there is no canonical policy to discover. Writing one down is a political act — someone's version of the truth wins — and no amount of model capability makes that a purely technical task.

Three things genuinely changed.

Models can finally do the reading. Extracting candidate patterns across thousands of threads, tickets, and PRs — with citations — requires long-context reasoning that simply didn't exist a few years ago. The *drafting* half of the problem is newly tractable, even if the *adjudication* half remains human.

A standard for agent–tool interaction exists. Anthropic's Model Context Protocol (MCP), introduced in late 2024 and since adopted broadly, means an "executable knowledge layer" can be exposed to any agent through one interface rather than N bespoke integrations. The plumbing stopped being the hard part.

The sources are reachable. Slack, GitHub, ticketing systems, and meeting tools all expose APIs and webhooks. Ingestion — step 1 — is now weeks of work, not years. Which is exactly why it's a commodity, and why the value has moved up the pipeline to the steps nobody has cracked.

The "company brain" is a real gap, badly named. Nobody buys a company brain — they buy a fix for a specific painful workflow. But underneath the VC framing sits a genuine unsolved problem: the layer between raw company data and reliable AI automation doesn't exist yet, and the teams racing toward it are mostly rebuilding step 1 while steps 2 and 3 sit untouched.

My honest conclusion after this research: the winner here won't be whoever ships the grandest platform. It'll be whoever picks the narrowest wedge — one workflow, one buyer, one measurable outcome — survives the cold-start period where the brain isn't yet trusted, and earns the right to expand.

I'm continuing to dig into this space, and there may be more to share soon. If you're thinking about the same problem — or think I've got part of this wrong — I'd genuinely like to hear it.