Retrieval is not discovery

“AI that learns your business” is on every vendor’s homepage now — ours included. But “learns” is doing a lot of hidden work. It can mean three very different things, and they are not equally hard to build or equally valuable to own. If you are deciding what to buy, the most useful question you can ask is which of the three a system actually does. Most stop at the easiest one and let you assume the rest.

Three rungs, not one

Borrow a ladder from an unlikely place. Recent category-theory work on scientific discovery — Fiona Wang and MIT’s Markus Buehler, 2026 — draws a sharp line between three things an agent can do with knowledge: retrieve, search, and discover. It was written about protein mechanics and materials science, but the rungs map cleanly onto the far messier world of running an organization. The distinction is not academic hair-splitting; it is the difference between a tool you rent and an asset you build.

Retrieval: finding what is already written down

Retrieval is looking something up. The fact exists, someone recorded it, and the system fetches it on demand. This is what most “AI memory” and “chat with your documents” features actually are: a search index over things you already wrote, returning the closest match. It is genuinely useful, and it is table stakes. But it can only ever hand back what is already in the store, in the shape it was stored. The vocabulary is fixed. Ask it about something nobody wrote down, and it has nothing to give you.

Search: new combinations of the same parts

Search is recombination. Instead of fetching one fact, the system chains steps, plans, calls tools, and assembles an answer that no single document contained. This is where most “agentic” demos live — multi-step workflows stitching together known pieces — and it is where a lot of real value sits. But notice the ceiling. Search finds new paths through the categories the system already has. It gets better at using its vocabulary; it does not change the vocabulary. Every answer is still assembled from parts that were defined in advance.

Discovery: changing the categories themselves

Discovery is the rung almost no one ships. It is what happens when a system notices that its current categories no longer fit the world — and revises them. In a business, that is not an abstract event. It is the contract clause that matches none of your templates. The pipeline failure with no name in the runbook. The customer segment that behaves unlike every bucket you defined. Discovery is adding a new type, not another row: the system changes what it treats as a thing worth tracking, because reality forced the question. That is a different operation from retrieving a fact or recombining known ones, and it is the one that turns a record-keeper into something that actually learns.

Retrieval returns what you knew. Search recombines it. Discovery changes what counts as worth knowing.

Why most systems stop at rung one

Because retrieval is easy to build and easy to demo. Drop your documents in, ask a question, get a passage back — it looks like understanding. But the value in a real business hides in the exceptions, and exceptions are exactly where a fixed vocabulary fails. A system that only retrieves will answer confidently and miss the one case that did not fit the mold — which is usually the case that mattered. The gap between rung one and rung three is not effort or model size. It is architecture: whether the system can revise its own model of you, or only query a frozen one.

What to ask a vendor

When someone tells you their AI “learns,” ask which rung. Is it a better search box over your files? An agent that chains steps? Or does it genuinely update its model of how your business works when the old model stops fitting — and can you see and audit each change? The first two are useful tools. Only the third is an asset, because only the third is worth more next quarter than it is today. The honest vendors will tell you where they actually sit. The rest will let the word “learns” do the work.

Where we stand

We build toward the third rung, and we build it grounded. Not autonomous invention — disciplined, inspectable learning. A knowledge base you own, that accretes structure as it does real work in your environment, surfaces the exceptions instead of smoothing them over, and records every revision so you can audit why it believes what it believes. The frontier labs are chasing discovery for science. We are after the quieter, equally valuable version: a system that keeps revising its understanding of how your organization actually runs — and can show its work.