What production looks like
The demo is easy. A model reads a document, pulls out the fields, and everyone claps.
Production is different. Production has the supplier who sends a photo of a page. The credit note that arrives before the invoice. The job code that only exists in one person’s head. The month where volume triples.
We build for that version. Most of the work is defining the exception paths, deciding where a human has to sign, and instrumenting the thing well enough that you can see it drifting before your customers do.
How we scope it
A build starts with a workflow definition, written in plain language, that answers four questions:
- What comes in, in every form it actually arrives in
- What has to be true for the output to be correct
- What happens when the system is not confident
- Who is accountable for the output once it leaves
If we cannot answer those, the workflow is not ready and we say so before you spend money on it.
Where the human stays
A reviewer sits where an error is expensive.
We place review points by the cost of an error. A misfiled internal document is cheap and can run unattended. A payment to a new bank account is expensive and irreversible, so it gets a human every time, forever, regardless of how good the model gets.
Teams that put a reviewer on every step end up with a workflow that is slower than the manual process. Teams that put a reviewer nowhere end up with an incident.
Handover is the deliverable
You end up owning this. That means an operator on your side who can read the logs, adjust the rules, retire a prompt and run the evaluation set after a change.
We train that person, write the runbook with them rather than for them, and stay on call while adoption settles.