AI Is Not the Enterprise Brain

There is a design pattern spreading through enterprises at the moment that deserves more scrutiny than it is getting. It places a large language model at the centre of a workflow and arranges everything else around it: the data feeds it, the tools serve it, the interface exposes it, and the human reviews its output.

It demonstrates beautifully. It is also, in most cases, the wrong shape.

One region, not the organ

In our model of the organisational brain, models and analytics occupy the Intelligence Cortex. It is a genuinely powerful region and it has transformed what is possible in the last few years. It is still one region among nine.

Consider what the others carry. The Decision Cortex holds decision rights, thresholds and escalation. The Systems Cortex holds architecture, integration and reliability. The Memory Network holds what the organisation has chosen to keep and curate. Governance holds accountability, auditability and the boundaries of automation. Domain holds the industry logic without which a fluent answer can be confidently irrelevant.

Placing a language model at the centre implicitly asks it to do the work of all of them. It is not equipped to, and it will not tell you when it is out of its depth.

What goes wrong, specifically

The failure modes are consistent enough to name.

Unbounded responsibility. The model is asked not only to draft, extract or rank, but also to decide what to do next. The system's behaviour then becomes a function of prompt text rather than reviewable code, which means it cannot be tested in any conventional sense.

Retrieval as an afterthought. Teams invest in the model and neglect the corpus. Retrieval quality is set almost entirely by what the organisation has curated, versioned and retired — a Memory Network problem wearing an Intelligence Cortex costume.

Governance retrofitted. Approval flows, audit trails and human control points are added after a successful pilot, at several times the cost, and they fit badly because the architecture did not anticipate them.

Evaluation deferred. Nobody agreed what "working" means before the build, so nobody can say whether it still works in month six. Silent degradation is the most expensive AI failure precisely because nothing announces it.

Determinism abandoned. Rules that could have been written as ordinary code — eligibility, entitlement, sequencing, permissions — are handed to a probabilistic component because it is faster to prompt than to specify.

The shape we recommend instead

Invert the arrangement. The deterministic system is the architecture; the model is a component inside it with a bounded, well-specified job.

Concretely, that means the workflow engine owns sequencing. Business rules own anything that can be stated as a rule. The permission model owns what data may be touched, enforced outside the model rather than requested of it. Retrieval owns grounding, over a corpus somebody is accountable for. The model owns the tasks that genuinely require language or perception: extraction, summarisation, drafting, classification, ranking. Human approval owns anything consequential or irreversible. Observability and evaluation watch all of it.

This is less exciting to demonstrate. It is considerably easier to operate, to explain to a regulator, and to change when the underlying model is replaced — which it will be, sooner than the business case assumed.

A test worth applying

Before approving an AI design, ask what breaks if the model is swapped for a different one from a different provider next quarter.

If the answer is "we re-tune some prompts and re-run the evaluation set", the architecture is sound. If the answer is "we would have to re-validate the whole workflow because we do not actually know what the model is deciding", the model is doing the architecture's job.

AI belongs in the enterprise brain. It should not be mistaken for it.