For years, the limits of enterprise software were a convenient form of camouflage. Systems were narrow enough that the contradictions between them stayed politely out of sight. Each team optimised its own corner, each tool spoke its own dialect, and the gaps were quietly absorbed by humans who learned — through years of unwritten apprenticeship — how to translate between them.
AI ends the camouflage. Not because it is hostile, but because it is literal.
What the model actually sees
When a capable model is pointed at the enterprise, it does not see departments. It sees claims. And it notices, immediately, that the claims do not agree. The CRM says the customer churned in March; billing says they renewed in April. The product catalogue lists a SKU the warehouse no longer stocks. The policy document forbids what the approval workflow routinely permits. The KPI on the executive dashboard is computed from a query no one can locate.
None of this is new. What is new is that the contradictions are now legible to a system that will act on them. The mess was always there. It was simply tolerable because no one was forced to read it end-to-end.
The smarter the model, the less it tolerates ambiguity. And the less it tolerates ambiguity, the more the organisation's ambient chaos becomes a visible operating risk.
Why this feels like the AI is broken
It is not. The model is doing exactly what it was asked to do: reason across the available context. The discomfort comes from discovering that the context itself is incoherent. Most "AI failures" in the enterprise are, on inspection, definitional failures (two systems using the same word for different things), provenance failures (a number whose origin no one can name), process failures (the documented workflow and the real one diverged years ago) or authority failures (no one can say who is allowed to decide).
A human navigates this terrain by instinct. A model navigates it by asking the question the organisation has been avoiding: which version is true?
The uncomfortable gift
This is, in the end, a gift — though it rarely feels like one in the first quarter of deployment. For the first time, the enterprise has a counterparty that will not paper over its own incoherence. The work that follows is not AI work. It is operating work: define the entities once, centrally, with an owner; make provenance non-negotiable; reconcile the documented process with the real one and pick which one survives; name the decision rights, out loud, in writing, per domain.
The organisations that will compound advantage in the agentic era are not the ones with the best models. They are the ones willing to let the model show them what they have been refusing to see — and to act on it before the next model release shows them more. The smarter AI becomes, the less room there is to pretend. That is the threat. It is also, finally, the opportunity.