FIG. 06 The same five pieces, doing a second job

What it takes to make AI work inside a company is what it takes to train a model on how that company works. Read each row across: the thing on the left is already installed, and the thing on the right is what it becomes.

the piece what it does for you today what the same thing becomes

Company Brainprofiled in place

Grounds every answer in how the business actually runs, cited to the source.

The corpus a model of your own is fine-tuned against.

live

Governed tool accessidentity · scope · audit

Safe, audited actions inside real systems, with no raw credentials in an agent's hands.

The real systems a model is trained and tested against.

live

Sandboxes and deployscanned · reversible

Somewhere a non-engineer can ship an app or an agent without a ticket.

The environment where attempts run safely, over and over.

live

Traces and correctionscaptured in the work

What ran, what it touched, and what a person fixed afterwards.

Labelled data and human feedback, produced by your own people doing real work.

live

Private evalsyour outcomes, not benchmarks

Whether a workflow actually worked, measured against the outcome you named.

The benchmark only you own, and the gate a candidate model has to pass.

in build

Fine-tune and distilon your infrastructure

Not yet part of the transformation work.

Small models carrying the routine work, trained and served inside your tenancy.

roadmap

what everyone else does

Authors a copy of the business and learns in the copy, even when the copy runs in your cloud. A replica cannot hold twelve years of exceptions in one company's order-to-cash.

what this does

Instruments the original: your real systems, during real work, judged by the people who own the work. Only possible because the transformation work was already being done.

Capture is the hard part to get, and it only happens inside real work. Nothing new gets installed.