Why the AI Factory is a dead end
Load documents, build a RAG index, plug in an LLM. Every enterprise AI programme starts the same way — and most of them end the same way too. The problem is not the model.
Over the past two years a whole category has appeared: the AI Factory. Every pitch deck looks the same.
Load your documents. Build a RAG index. Plug in an LLM. Get an AI assistant.
It looks sufficient. It demos beautifully. And then, in a striking number of enterprises, the project quietly stops somewhere between the pilot and production.
Not because the model was bad. Not because there were too few GPUs. Not because the retrieval could have been tuned better.
The problem sits a layer deeper.
What an LLM actually knows about your company
A language model generates plausible continuations of text. That is an enormous capability, and it is the wrong capability to build an enterprise decision system on top of, on its own.
A model does not know which of your documents carry legal force and which are drafts someone left in the shared drive. It does not know that an order from March supersedes the instruction from January. It does not know that a rule applies only to a specific class of goods, a specific customer segment, a specific jurisdiction. It cannot tell you why it decided something in a form an auditor would accept. It cannot check its own conclusion against anything except more text.
And, most importantly, it has no notion of accountability — the property that makes a corporate decision a corporate decision. Somebody signs it. Somebody can be asked to explain it. Somebody can reverse it.
None of that is in the weights. None of it is in the vector index either.
Why the demo works and the rollout does not
The demo works because a demo asks questions. "What does our returns policy say?" "Summarise this contract." "Which suppliers are affected?"
Retrieval-augmented generation is genuinely good at that. It is a very strong search interface with a natural-language front end.
The rollout fails because production asks for decisions and actions. Classify this shipment. Approve this invoice. Issue this response to a regulator. Update this record. And a decision has requirements a search result does not:
- It must be traceable to the specific norm or record that justifies it.
- It must be consistent — the same case decided the same way tomorrow.
- It must be bounded — the system must not be able to take an action nobody authorised.
- It must be reviewable — a human must be able to intervene at the points that matter, and only at those points.
An AI Factory hands you none of these. It hands you a model and an index, and leaves the four hard properties as an exercise for the integrator. So integrators build them ad hoc, per project, in prompt templates and Python glue — and that is precisely the layer that does not survive contact with real volume, real audits, and real staff turnover.
The two failure modes
In practice, AI Factory projects fail in one of two directions.
They shrink into enterprise search. The honest outcome. The organisation ends up with a better intranet search box. That has value, but it is not the transformation the budget was approved for.
They become an expensive chat over documents. The unhappy outcome. The system is deployed to people who need answers they can act on, produces answers that are usually right and occasionally confidently wrong, and quietly loses trust. Once a domain expert has caught the system fabricating a reference twice, they stop using it — and no amount of accuracy improvement wins them back, because the issue was never the average accuracy. It was that nothing in the architecture distinguished a well-grounded answer from an invented one.
Both outcomes come from the same root: the architecture ends at generation. There is nothing after the model.
The missing layer
Think about what actually happens when an experienced employee makes a decision in a regulated domain.
They form a hypothesis quickly — that is the intuitive, associative part, and it is the part LLMs genuinely replicate well. Then they do something a language model does not do at all: they check it. Against the classifier. Against the current version of the regulation. Against the precedent. Against what they are personally allowed to sign off. If it does not hold, they revise it. If it holds but sits above their authority, they escalate it.
The hypothesis is neural. The check is symbolic. Competent professional work is the loop between them.
Enterprise AI has, almost universally, built only the first half. It has industrialised hypothesis generation and left verification, authorisation and explanation to be improvised.
That is why we think the industry took a wrong turn. What is needed is not another AI Factory — a pipeline that ends when the tokens stop. What is needed is a different kind of architecture: an execution environment for intelligent systems, where the model proposes, the domain model validates, policy decides what is permitted, tools act, and a human owns the decisions that must be owned.
We call that a harness system.
Next in this series: why RAG will never become the brain of a company, and what a semantic model of a domain adds that retrieval cannot — Why RAG will never be a corporate brain.