The value is usually in one expensive, boring process — not a chatbot.
We start where a business already loses hours: purchase orders re-keyed, invoices checked line by line, sessions nobody has time to write up. A model goes behind that one step and a person stays on the result. Two ways in, depending on whether you're deciding where to start or already know what you want shipped.
Purchase orders arrived as PDFs and were re-keyed by hand, then price-checked against a book that changes quarterly. A model reads them now, every line is matched against the manufacturer's files, and anything that doesn't reconcile waits on a person before it leaves the building.
Two ways in
Same practice, different starting point.
For owners and operators
Adoption advisory
You know a process is expensive. You don't know whether a model can carry it, or which one to start with. We follow the documents, sit with the people doing the work, and come back with a ranked shortlist, an honest cost, and the steps we'd tell you to leave alone.
For teams who already have the spec
AI engineering
Extraction from a document type. Answers grounded in your own corpus. Transcription that has to stay on the device. We build and ship it — with the fallbacks, the diagnostics, and the review step that keeps a wrong answer from becoming a wrong order.
How an engagement runs
Small, in order, and cheap to stop. Nothing here requires a platform decision before you've seen the thing work on your own documents.
01
Audit
One week
We follow one workflow end to end and write down what each step costs in hours. The output is a document: what a model can carry, what it can't, what it would cost, and what we'd leave alone. Plenty of steps come back “don't.”
02
Pilot
Four weeks
One process, in production, with your staff and your ugliest real inputs — the vendor whose columns are swapped, the invoice that was photocopied twice. A demo on clean documents proves nothing. If it doesn't hold up, you've spent weeks finding out instead of quarters.
03
Rollout
Ongoing
The second process costs less than the first, because the review screen, the correction loop, and the diagnostics already exist. Staff corrections become the test suite — every fix ships with the real document that broke it.
Ground rules
A human confirms anything that leaves the building
Citations are attached by retrieval, not written by the model
Never invent a value to fill a field
A staff correction is ground truth — it becomes a test
It keeps working when the model is down, or it says so plainly
Measured in hours off a real week, not demo quality
TickA terminal-flavored Mac app for projects, nested todos, and deadlines. Local-first SQLite, quick capture, pinned deadline strip, and an MCP server so Claude and friends can read and write the same checklist.
