Every RAG demo works. That is exactly the problem. You point a model at a folder of PDFs, ask it three questions you already know the answers to, and it responds beautifully. Everyone in the room nods. Then it goes live, a real employee asks a real question, and the system confidently cites a policy that was retired two years ago. The demo was a feature. What the business actually needed was a product.
Retrieval-augmented generation — connecting a language model to your own documents so it answers from your content instead of guessing — is the single most requested thing I'm asked to build. It is also the one most often underestimated. Teams treat it as a checkbox: "add RAG." In practice, the model is the easy part. Everything around it is the work.
A chatbot answers. A company brain is accountable.
The difference is in how each one is judged. A chatbot is judged on whether the reply sounds right. A company brain — a system employees and customers rely on to make decisions — is judged on whether the answer is correct, current, permitted, and traceable. Those four words contain almost none of the model and almost all of the engineering.
Generation is a solved-enough problem; you can get a fluent answer from any modern model. The hard questions are the ones the demo never asks: Which version of this document is authoritative? Is this person even allowed to see this? How do we know the answer was right? And what happens when it isn't?
The five things a real company brain needs
When I scope one of these systems, the model choice is one line near the end of the document. These five things are the rest of it.
Document governance
One authoritative version of each document, with ownership, expiry, and a way to retire the stale ones. Most "wrong" AI answers are right answers to old documents.
Retrieval quality
Chunking, embeddings, and ranking tuned to your content. If the wrong passage is retrieved, no model can save the answer — it will fluently summarise the wrong thing.
Permissions
Retrieval must respect who is asking. A salesperson and the CFO ask the same question and must get answers scoped to what each is allowed to see.
Evaluation
A real test set and a score. "It felt good in the demo" is not a metric. You need to know, on a fixed set of real questions, how often it is correct — and notice when that drops.
Human escalation
A clear, designed path for "I'm not sure" — hand off to a person with the full conversation, rather than inventing an answer to fill the silence.
A demo runs on documents you curated this morning and questions you chose. Production runs on a decade of duplicated, contradictory, half-retired files, on questions you never imagined, asked by people with different permissions — some of them adversarial. The demo measures the best case. The business lives in the worst case.
In a company brain, the model is the easy 20%. Retrieval, governance, and escalation are the 80% that decide whether anyone trusts it.
Scope it like a product, not a plugin
If you're about to commission a "company brain," you can de-risk the whole thing before a line of code by answering five questions:
- What must it answer? Write the 30–50 real questions it has to get right. That list is your test set and your scope.
- Who is allowed to ask? Define the permission boundaries up front, not after launch.
- Where does truth live? Name the authoritative source for each topic and who keeps it current.
- How will you know it's right? Decide the score and the threshold before you build, so "good enough" isn't a feeling.
- When must it defer? Design the moment it stops answering and hands off to a human.
Answer those, and the model becomes a swappable detail. Skip them, and you've built a confident, fluent, beautifully-worded liability.
Thinking about a company brain?
I design and build RAG systems that are correct, permissioned, and measurable — not just demos that impress in a meeting. See RAG / Company Brain, or book a consulting call.