A company brain is mostly plumbing

Here is a rough breakdown from our last few deployments: about a fifth of the effort goes into anything you could call AI. The rest is plumbing. Connectors that stay authenticated. Permission maps that match reality instead of the org chart from 2021. Deciding what to do with four copies of the same contract, three of them stale.

This surprises buyers every time, because vendor marketing is model-shaped. Model benchmarks make good charts. 'We spent two weeks reconciling your SharePoint permissions' does not. But the second one decides whether the brain gives correct answers to the right people.

One example from a deployment last year. A client's expired service account had been silently failing for months, so their document index was missing everything newer than March. The old search tool never told anyone. People assumed the documents didn't exist and rewrote them. The fix was not a better model. The fix was monitoring on a credential.

The lesson we keep relearning is that answer quality is set by the index, and index quality is set by boring operational discipline. A mediocre model over a complete, fresh, permission-correct index beats a frontier model over a stale one, and it is not close.

So when you evaluate any company-brain product, ours included, ask about the plumbing first. How does it notice a broken connector? How does it learn that someone changed teams? If the answers are vague, the demo is the best the product will ever look.