Eight stages. One register.
The industry sells AI models. The model was never the expensive part.
The hard part of operational AI is everything around the model: assembling the data, understanding the process, knowing where each value came from and who checks it. Once you understand a process and stage the data, doing the thing is the easy part.
So I put the work where projects actually fail, and treat the model as what it is: a component.
Eight stages. Each ends when a named artefact exists — or it hasn't ended.
There's no separate progress status to maintain. The artefact either exists or it doesn't.
Bootstrap
An empty project, ready to work in.
Discover
Your process in your own words, its rules written down, and how it performs today.
Prove
A small test of the riskiest assumption, and what it found.
Specify
Everything to be built, each with what "done" means.
Propose
A fixed price and a written scope.
Build
The working system, on infrastructure you host.
Handover
Runbooks, every decision written down, and training.
Measure
What changed, measured the same way as at the start.
I draft your rules from what is already written down, then check them with the people who apply them.
The checking happens in short working sessions. That usually means two or three of your experienced people, for a few hours each, over the first week or two.
No saving without a baseline.
How your process performs today is measured before anything is built. No before-figure, no saving claimed. Ever.
Scope is written once.
The contract's scope, exclusions, acceptance criteria and project board are all generated from one register — nothing is typed twice, so nothing can drift.
Not in the register, therefore a variation.
Fixed price. Change your mind free.
The register fixes what the system must do — and deliberately not how. The price holds, and you reshape the design every iteration at no cost.
I won't hold your system hostage.
Or leave you one you can't look after.
Some firms keep you dependent and charge rent. Others hand over everything and leave, with code nobody in the building understands. You own what's built outright, and the handover includes what your team needs to run it: runbooks, the decisions written down, and training.
You may run, host, modify, fork, transfer, or engage any third party to maintain or develop what you receive — without permission and without further payment. That's a warranted term, not a promise in a slide.
Before handover, automated checks confirm your system doesn't depend on any tool of mine. If a check finds something, the handover waits until it's fixed.
"What happens when the AI gets it wrong?"
A manual route stays open. That is written into our master services agreement and our terms of business. The agreement says:
"Where a capability depends on a model or an external source, a manual route will remain available so that its failure does not stop the Client working, unless the Statement of Work says otherwise."
Beyond the agreement, this is how it usually works:
WHEN A TRANSFER FAILS
It is retried automatically. If it still fails, a named person in your team is alerted, with the record. A reconciliation report shows what did not arrive, and your team re-runs it or finishes it by hand. I fix defects in the integration itself.
WHEN THE OUTPUT IS WRONG
What the AI produces is a draft, and a named person reviews it before it is submitted or sent. A rejected item is marked in its system of record and finished by hand there, so it is neither lost nor redone by the automation.
What I can say about recent work.
One recent engagement: AI-driven PDF form analysis and completion, for a US permit-processing company.
Software does the routine delivery work, so the build is quicker and costs less. A person makes every decision that matters: no agent can sign a contract, take on a client or close a project.
I don't make extravagant claims.
I work with facts.
If I can't show you the evidence, I won't claim it. That goes for this page too.
Built with the method: Sourcier
A data-quality and automation product for recruitment agencies running Bullhorn — the method's own proving ground, run as a client of itself. Pre-launch, currently invite-only.
Start with the design. Stop after it if you want.
The first engagement is a small, fixed-price piece of work whose output is the design itself: your process in your own language, the riskiest assumption tested, and a priced, generated proposal for the build. You own everything it produces.
Thirty minutes first, no slides. If it's advice you need, we talk about the decision in front of you. If it's a build, by the end you'll have the name of the process worth measuring first.
Free, 30 minutes, no obligation.