Organisational memory beats another model
Swapping models is easy. Teaching the organisation to remember how it works is the hard part.
Every company already has AI-shaped knowledge: policies, precedent, who can approve what, which customers are edge cases. Most of it lives in people and chat threads.
When that memory is not structured, every agent and every new hire starts from zero. You get generic answers, repeated mistakes, and meetings that re-litigate settled decisions. Then someone suggests a new model, as if the amnesia were a compute problem.
Models are interchangeable. Memory is not.
A model is a part. It will be replaced. The layer that should outlast it is how the organisation remembers: what is allowed, what was decided, what this customer is like, which document is the real one.
If your next investment is only another model subscription, ask what will still be true when that model is obsolete next year. If the answer is “nothing,” you are renting cleverness, not building a system.
What to structure first
Start with the decisions the desk already makes, not with a knowledge-base project that tries to boil the company. Approved answers. Exception paths. The fields that must be true before a write-back. The document that counts.
That is why invoices and documents and research work fail when they are treated as “just add a chatbot.” The job is matching, memory, and a person on the exception — not a paragraph generator.
People still decide
Memory is not a replacement for judgement. Humans still decide. The system remembers. That split is the whole point. An agent that cannot find the last decision will invent a new one, confidently.
Context as infrastructure
Our delivery work treats knowledge, decisions, and context as infrastructure. Instinct is the longer version of that idea: organisational context, not another stranger in the tools. Models sit on top. They should be swappable without the company forgetting how it works.
Keep reading.
Start with the work. Not the AI.
The useful first question is not which model to buy. It is which process still depends on memory, retyping, and whoever happens to be at the desk.
Read the note
Demos aren't desks
An agent that dazzles in a slide deck still has to survive your exceptions, approvals, and Tuesday afternoon volume.
Read the note
The aftercare question
Consulting, build, or build-and-run — the wrong depth wastes more money than the wrong model.
Read the note