Five convictions — and why they matter.

Here is the uncomfortable truth about enterprise AI: most of it has amnesia.

An organization deploys a chatbot. It answers questions. But tomorrow, it has forgotten everything it learned from today’s conversations. An automation pipeline processes a thousand documents. Next week, it cannot recall a single pattern from that work.

Everyone is fixated on the model. But the model is the one thing every competitor can rent, from the same handful of providers, at the same price. It is a commodity you have no monopoly on. What no one can copy is the accumulated context underneath it — the decisions, the corrections, the institutional knowledge that only your organization has lived through.

The instinct most enterprises follow is to build a knowledge base from their subject-matter experts — capture the specialists’ wisdom, connect it to a model, and call it done. It is a reasonable start and a dangerously small target. Ask yourself: if you fired every employee except the SMEs, would the business continue to function? Of course not. The knowledge that actually runs a company is distributed across every function — how a deal really gets approved, how a project actually moves, where the workarounds live when the official process breaks. That operational knowledge is tacit, cross-functional, and almost never held by the people labeled “experts.” An SME-only approach captures the documentable layer and misses the layer where compounding advantage actually lives.

A memory layer that captures organizational knowledge across every function — not just the expert tier — is the foundation everything else builds on. Connected to custom applications, tools, data sources, and orchestration layers, it is what transforms a standard model into something that operates with the full context of how your organization actually works. Memory before orchestration — always. Build the road, not just the engine.

Here is the part that should give you pause: almost no one builds this. They buy it — the same strategy, from the same firms, as everyone they are trying to beat.

For most of my career, the honest answer to “can we build that?” was “not yet.” The technology was the ceiling. That ceiling has largely fallen away.

The constraint now is rarely the tooling. It is imagination and will — the willingness to actually attempt the thing rather than commission a report about whether the thing might someday be feasible.

The models, the APIs, the tools, the infrastructure, RAG, MCP — they are building blocks. Everyone has access to the same ones, from the same handful of suppliers, at roughly the same price. The pieces confer no advantage on their own.

The edge is entirely in the design — how you select, sequence, and assemble those components into something that solves a problem no off-the-shelf product addresses. The parts are commodity. The architecture is the craft. That is the whole game, and it is why the same stack in two organizations produces stalled pilots in one and a compounding system in the other.

The enterprise AI space is crowded with people who have read every paper and shipped nothing. They can draw a beautiful architecture on a whiteboard, but they have never sat in front of a production system at 2 AM working out why the pipeline is dropping a third of its context.

There is a kind of knowledge that only comes from building — from deploying inside real organizations with real constraints, real politics, and real consequences. It lives in the scar tissue of failed deployments and the muscle memory of the ones that worked. I would rather prototype the answer than theorize about it. Nothing is off limits until you have actually tried.

And the first thing you build is never the prize. What matters is what it leaves behind — the memory, the patterns, the muscle — for the next build to stand on.

There is a principle underneath all of this, and it is really a philosophy of management more than a philosophy of AI: authority should follow knowledge, and accountability should follow authority.

That is not a call for organizational anarchy. The CEO still decides what the company is trying to accomplish. Finance establishes the economic constraints. Legal establishes the legal ones. Security sets the security requirements. Domain experts contribute their particular expertise. The structure stays intact — every function still owns what only it can own.

What changes is who solves what. You do not need fifteen people collectively solving every component of every problem. That is how a $30,000 problem becomes a $500,000 one — whole-organization opinion routed through every constituent part until the boldness bleeds out, and what ships is the version everyone could agree on.

The alternative is simple to state and harder to practice. Decompose the problem. Put the right expertise at the right point. Integrate the result. Hold someone accountable for the whole. Authority sits with the person whose knowledge is closest to the part; accountability sits with whoever owns the integration.

For me, that principle becomes a simple operating bargain: give me the objective and the constraints, challenge my reasoning aggressively, give me authority proportional to the responsibility, and then hold me accountable for the result. I want the informed criticism — the goal is not agreement, it is the best answer. The line that matters is between informed challenge and uninformed veto.

This is exactly how the best AI systems get built — broken into parts, each part handed to the function with real authority over it, then assembled by someone accountable for the outcome. But the pattern is not unique to AI. It is how any hard thing gets built well. The technology only makes the cost of ignoring it plain to see.