What is the knowledge stack?

The knowledge stack is the three layers an AI system runs on beneath the model: Knowledge (what the organization knows and how it is structured), Dissemination (how that knowledge reaches people and systems), and Decision (how the organization actually decides and whether the reasoning is recorded). The model sits on top and amplifies whatever those three layers hand it. The model is the one layer you can replace with a contract; the three below it are the ones only you can build.

How is the knowledge layer different from buying an AI model?

A model is a capability you license. The knowledge layer is the material that capability works on: your structured record of what the business knows and how it decides. Buying a model changes what you can do with your knowledge, and the knowledge stays as it was. If it lives in people's heads and scattered decks, a better model reads the same gaps a weaker one did. Rebuilding it is slower, internal work that no vendor can sell you.

Why do AI pilots produce confident answers nobody can use?

Because the model was handed an organization that could not give it a clean account of what it knows or how it decides. The reasoning behind past decisions was never recorded, knowledge moves by hallway rather than by any system a model can read, and the ontology is out of date or was never drawn. The model works on the material it receives, so a thin input yields a fluent output with no grounding. A fluent answer built on a gap can look like competence, which makes the gap harder to see.

When should an executive audit the knowledge stack?

Before approving the next significant AI purchase, not after. Run the audit when the tooling roadmap is growing faster than the record underneath it, when a pilot produced an answer nobody trusted, or when the reasoning behind key decisions lives with specific people rather than in a system. The wrong moment is after a vendor is already in flight, because by then the audit can end up justifying a decision already made instead of informing it.

What is the first step to fixing a weak knowledge stack?

Write the reasoning behind one recent decision in a single page, using only the existing record and without calling the people involved. If you cannot, you have located the weakest layer in one sitting. That one page is also the template for the fix: it shows what a legible decision record looks like, which exposes what the dissemination and knowledge layers were missing. Writing the page before shortlisting models tells you whether the problem is a model problem at all.

The Three Layers Under the Model That Decide Your AI Return

The knowledge stack is the three layers an AI system runs on beneath the model: what your organization knows, how that knowledge moves, and how the organization decides. The model sits on top of all three. My view is that most organizations still handle those three layers in an archaic way while leadership attention goes to the top one, and that is how an AI budget can keep climbing while the return flattens out.

Leadership attention, as I see it, goes to the technology: which model to license next, which platform to roll out across the business. The layer the tooling stands on gets far less of the conversation. Check your own AI roadmap and count how many lines are about knowledge or decisions.

Two questions for your own leadership team

Here are two questions worth putting to your own leadership team before the next model decision. First: where is the reasoning behind your last three significant decisions written down? Who weighed what, which numbers they trusted, what they ruled out, and why they ruled it out.

If the honest answer is "in people's heads" or "in a deck somewhere on the drive," you have learned something about the bottom of your stack. Picture a pricing call whose logic sat with one person who has since left the company. The decision still stands. The reasons behind it left with him, and so did any chance of a model learning from them.

Second: if you hired a sharp new analyst on Monday, how long before that person could reconstruct how one part of the business actually works from what is written down? If the answer is weeks or months, a model plugged into the same record hits the same gap, and it has no colleague down the corridor to ask.

A team in that position is about to point a capable model at an organization that cannot hand it a clean account of what it knows or how it decides. The ontology, by which I mean the map of what the business knows and how the pieces relate, may be years out of date or may never have been drawn. Knowledge travels by hallway and habit, through no system a model could read. The decisions the AI is meant to support were never recorded as decisions, so there is nothing for the model to learn the pattern from and nothing for it to check its own output against.

There, a pilot can produce a confident answer that no one can use, because the model works only from what it was handed. People with context can make an undocumented process run. A model has no context you did not give it.

Repeat that across several pilots and the cost grows. Each new model lands on the same undocumented base, and each pilot meets the same gap from a slightly different angle.

A note I wrote in my journal in October put my view plainly: leadership keeps focusing on technology, when knowledge and decision making are the moat. Your competitor can buy the same technology next week from the same vendor. I would expect rebuilding the knowledge you have structured, and the way you decide, to take them a good deal longer.

Where my August essay on the moat argued for domain depth, this one is about the plumbing that makes depth usable: three layers you can inspect and fix.

The knowledge stack

I call this the knowledge stack. Read it from the bottom up, because each layer depends on the one beneath it.

            [ Model ]          replaceable with a contract
   ---------------------------------------------
   Layer 3   Decision          how the org actually decides
   Layer 2   Dissemination     how knowledge reaches people and systems
   Layer 1   Knowledge         what the org knows, and how it is structured

Layer 1, Knowledge. This is what your organization actually knows, and whether that knowledge is captured and structured well enough for anyone, or anything, outside the original author's head to use it. The archaic default is knowledge locked in individual experts, scattered decks, and chat threads nobody can search. Imagine a sales team whose real win rules, which discounts clear and which do not, live only in the head of a regional lead who has run the desk for nine years. The rules work beautifully as long as he is in the room. The day you ask a model to help price a deal, there is nothing for it to read, because the knowledge was never written down. The test is simple. Pick one function and ask whether a motivated outsider could rebuild its logic from your written record alone. If the honest answer involves "you would have to ask Priya," Layer 1 is thin, and no model will repair that.

Layer 2, Dissemination. This is how knowledge gets from where it lives to the people and systems that need it at the moment they need it. The archaic default is that it moves by meeting, by forwarding, and by the luck of asking the right person. Imagine a policy that changes in one department while the others keep running the old version, because nothing carried the update to them. The test is the new-analyst question from earlier. If the time to reconstruct how the business works is measured in weeks or months, dissemination is broken, and a model plugged into that pipe inherits the same blockage.

Layer 3, Decision. This is how the organization actually makes a call, and whether those calls leave a legible trail. In my view, the archaic default is that decisions happen in a room, get announced as outcomes, and lose the reasoning on the way out the door. Picture a company where, a few months on, nobody can say why the company walked away from a market or doubled down on a product line. The test is the first question above. If you cannot show the reasoning behind your last three decisions, you have no record for a model to support, audit, or improve. You have outcomes with no logic attached, which is the one thing a model cannot reverse-engineer for you.

The model on top amplifies whatever the three layers hand it. With strong layers, a capable model compounds them across the whole business. With weak layers, the same model turns thin inputs into confident wrong answers at scale.

Why leadership keeps shopping on the top floor

If the bottom three layers decide the outcome, the obvious question is why leadership teams keep spending on the top one. They default to the top floor because it is the only floor anyone built to be bought.

The model layer has vendors who call you, demos that look finished, a price you can put in a budget, and a logo you can show a board. It converts into a slide in an afternoon. The three layers underneath have none of that. Few vendors sell you a clean ontology of your own business, and few have a demo for "we wrote down how we actually decide." The work is slow, internal, and invisible until the quarter it quietly saves you, which means it never competes well against a tool that promises a result next month.

There is also a measurement trap. Tool spend is easy to report. A team can report that it bought the platform and trained hundreds of people on it, and that reads as progress. Knowledge work resists that kind of counting. "We made our pricing logic legible" has no license count and no adoption dashboard, so it loses the room to the thing that does, even when it matters more.

So I read a tooling-heavy AI roadmap with no line for knowledge or decisions as a warning sign. Keep buying models. Just hold the approval on the next one until someone can say, in plain terms, what it will stand on.

Three things to do this week

Run the one-page test on one real decision. Take your last significant call and try to write the reasoning behind it in one page, from the record alone, without calling the people involved. If you cannot, you have found the weakest layer in your stack, and you found it in one sitting.
Draw one page of your ontology. Pick a single function and sketch what it knows and how those pieces connect. Keep it to one page and expect gaps. The gaps are the finding: they show how much of the real picture lives only in people.
Split the next AI budget line before you approve it. For the next AI purchase on the roadmap, write two numbers next to it: what you are spending on the tool, and what you are spending on the knowledge and decision layers it will depend on. If the second number is zero, hold the purchase until it is not.

What to read next

Related essays from the archive:

About

Anees Merchant writes one essay every Tuesday about enterprise AI, agentic systems, and the human side of the work. He is the author of Merchants of AI, a TEDx speaker, and a doctoral researcher at the Swiss School of Business and Management, where he studies how AI agents can help organizations map and redesign the way their work gets done.

The newsletter version, with extra commentary, goes out separately on Mondays. Subscribe at /newsletter.

See you Tuesday.