How Enterprise Knowledge Powers AI Workers

We keep designing enterprise knowledge as a briefing to be read before the work starts. An autonomous worker doing real work does not read a briefing — it holds a conversation, and half of what it asks for could not have been anticipated when the task began.

An Autonomous AI Worker is forty minutes into resolving a supplier billing dispute. It began the way anyone would want it to begin: it pulled the master services agreement, the current rate card, the disputed invoice, and the three months of purchase orders that preceded it. On that basis it could already explain most of the discrepancy, and if the task had been a question rather than a piece of work, it would have been finished. But somewhere in the payment history it finds a credit memo issued the previous quarter that references a pricing exception approved by a regional operations director. Nothing in the contract mentions an exception. Nothing in the rate card contemplates one. The worker now needs to know who in this organization is permitted to grant a pricing exception, whether that authority is delegable, and whether an exception granted for one shipment silently establishes a precedent for the next — three questions that did not exist forty minutes ago, that could not have been asked at the outset, and that only became askable because of what the earlier retrievals turned up.

This is the ordinary texture of real work, and it is almost the exact opposite of the model that most retrieval systems were built around. The prevailing mental image of enterprise knowledge is a briefing packet: you state your question, the system assembles the relevant material, and you set out with what you need. It is a comfortable image because it maps onto how a person prepares for a meeting. It is also, for an agent doing multi-step work, wrong in a way that quietly determines whether the work finishes. What an autonomous worker needs from Enterprise Knowledge is not a packet delivered at the start. It is an interlocutor available throughout, one whose answers reshape the next question, because the shape of the task is discovered while performing it rather than declared before it.

The question that did not exist ten minutes ago

Consider what actually happens inside a task of any real length. The worker begins with a stated objective and a rough theory of what it will need, and that theory is correct for perhaps the first two retrievals. Then it encounters something — an unexplained credit, a clause that references a schedule not attached, a customer record whose region field disagrees with the address on the contract, an approval logged by someone whose role it does not recognize — and the encounter generates a demand for knowledge that is entirely conditional on the encounter. The dependency runs the wrong way for a lookup. You cannot precompute the query, because the query is a function of the findings, and the findings are a function of the earlier queries. Each retrieval changes the state of the task, and the changed state is what determines what should be asked next.

This is why the sequence of retrievals in a real AI Mission looks nothing like a set of parallel searches issued from a form. It looks like an investigation, where the first few questions are broad and orienting, the middle ones are narrow and specific in ways that would have been meaningless earlier, and the last ones are verification — checking that the thing the worker is about to do is consistent with a policy it only learned existed halfway through. A Reasoning Core running that arc is accumulating Observations, and each Observation is simultaneously a partial answer and a new question. The knowledge layer is not being consulted once with high stakes. It is being consulted continuously with escalating specificity, and the retrieval that decides whether the work is correct is usually the fifth or the ninth, not the first.

The practical consequence is that the value of a knowledge layer to an autonomous worker is concentrated in exactly the places that lookup-shaped evaluation never measures. Benchmarks tend to score the system on a set of well-formed standalone questions, which is a reasonable proxy for a person typing into a box and a poor proxy for a worker asking its seventh follow-up in a context the first six established. A system can perform respectably on the standalone set and fail the seventh question badly, because the seventh question is underspecified on its own and only makes sense against the accumulated state of the task. Judged as a lookup engine it looks fine. Judged as a participant in a working dialogue it is unusable, and the unusability only surfaces when something real is being attempted.

Retrieval was designed for a reader who scans, reformulates, and knows when to stop

The deeper reason lookup-shaped retrieval misbehaves under an agent is that it was tuned, quite reasonably, for a very particular reader. A person at a search box is a remarkably forgiving consumer of results. They scan ten of them in three seconds, discard the obviously stale ones by the look of the header, notice that result four is from a deprecated wiki and ignore it, recognize that their phrasing was wrong and rewrite the query without ever registering that they did so, and open two documents to read around the passage that matched. Every one of those behaviors is a repair mechanism, and the retrieval system has been implicitly designed to lean on them. It can afford to return a mixed bag, because the reader sorts the bag.

An Autonomous AI Worker repairs almost none of that by default. Handed the top three passages, it tends to treat them as a faithful description of what the enterprise knows, which means an omission reads to it as an absence and a stale document reads as current unless something in the result says otherwise. It does not see the header, the folder it came from, or the fact that a colleague quietly superseded the page last spring. It cannot read around the snippet unless the surrounding context is offered. And when a passage is retrieved that is merely adjacent to the question — the general expense policy when the worker asked about a specific regional exception — a person would register the mismatch immediately, while a worker with a plausible-looking passage in hand has every incentive to proceed. The failure is not that the retrieval was bad by search standards. It is that search standards assume a reader who compensates, and this reader does not.

The dialogue framing also raises a demand that a single lookup never has to satisfy, which is consistency across the arc of a task. A search box answers one query and ends; nothing about the second search is obliged to cohere with the first. A worker holds all of it at once, so a knowledge layer that returns one passage at step two and a differently-scoped passage at step nine has not given the worker two results, it has given it a contradiction to carry. The same holds for permissions, which are easy to think of as an admission check at the door: the worker's authority to see a document must be evaluated on every retrieval, not just the first, because the ninth question reaches into corners of the corpus that the opening question never approached and the entitlement that cleared the opening question says nothing about them.

Designing knowledge for a reader that keeps coming back

If knowledge is consumed as a dialogue, then the interesting design questions move from ranking to conversation. What does the layer return when the worker's question is precise but the corpus holds only something adjacent, and can it say so rather than substituting the adjacent thing silently? Can the worker narrow, expand, or ask what else lives near a passage it just received? Does a result carry enough about its own nature — that this is an approved policy rather than a proposal, that it applies to one region, that it was superseded — for the Reasoning Core to weigh it rather than merely read it? These are not refinements of relevance scoring. They are the difference between a source that participates in the work and one that hands over text and leaves the worker to guess at its standing.

They also explain a pattern visible across the reporting on autonomous operations, including the body of work published on the autonomous enterprise as an operating model: organizations that connect a capable agent platform to an existing search index frequently find the agent competent on simple, single-hop tasks and unreliable on exactly the multi-step work they bought it for. Gartner's prediction that over forty percent of agentic AI projects will be canceled by the end of 2027 points at costs, unclear business value, and inadequate risk controls rather than at retrieval specifically, but the mechanism is easy to recognize once you have watched a worker stall on its seventh question. Multi-step work is where the value was supposed to be, and multi-step work is precisely where a knowledge layer built for a search box stops keeping up.

The reframing worth carrying is small and changes almost everything downstream of it. Stop picturing Enterprise Knowledge as a library the worker visits before setting out, and start picturing it as a colleague the worker keeps interrupting — someone who is asked a vague question first, a sharper one after the first answer lands, and a very specific one an hour later that only makes sense to someone who has been following along. Every good property of a knowledge layer follows more naturally from that image than from the library one: it should remember the thread, disclose the standing of what it hands over, admit the edges of what it holds, and re-check on every exchange whether this worker is entitled to what it is about to be told. A library optimized for the first question will look excellent in a demo and disappoint in production. A colleague who is still useful on the ninth question is what autonomous work actually runs on.