How I work.
Four verbs, in order. Each one ends with something you can check.
I work in short, scoped engagements. I learn how your week actually runs, we agree what’s worth changing, I ship the build with you, and I come back a month later to check it stuck. That’s the whole shape. What follows is the longer version — what actually happens inside each verb, and what you’re holding at the end of it.
Learn
I start by sitting with your team and watching how the week actually runs. Not the org-chart version — the real one: the inbox that takes the first hour, the report assembled by hand every Friday, the question that gets asked in Slack every Tuesday. I’m hunting for the hours that leak, and for the two or three jobs that are repeating, text-shaped, and big enough that handling them moves a number. This part usually takes a few days. At the end of it you get a plain written read: what’s eating the week, and what’s worth fixing first.
The result: the two or three jobs worth fixing, found and written down.
At Meta: I evaluated thirty-plus AI platforms and kept three. Most of this stage is working out what isn’t worth your week.
More on this: Your AI is a mirror · AI vs automation
Agree
Then we decide together what’s worth changing. One job, not five. We agree the scope, and we agree the test it has to pass — what working means, written in a sentence, before anything gets built. We agree the number too: typical builds land in the four-to-low-five-figure range, and you know it before anything starts. This is also where I say no. I don’t build decks that go on a shelf, and I don’t build pilots designed to demo well. If it isn’t going to be part of someone’s real week within a month of shipping, better we find that out here, where finding out is cheap.
The result: a scope, a test it has to pass, and a number, all agreed before any code.
At Meta: the procurement rebuild started here, with checkpoints agreed up front to catch opex and capex issues before they hit.
More on this: Recurring AI jobs and the bolt-on trap
Ship
The build happens in your accounts, on your stack. If your team lives in HubSpot, it runs in HubSpot, not in a separate dashboard nobody opens. I treat the model as a smart junior, not an oracle: it does the first pass, a human reviews, and the review layer is sized to what a mistake would actually cost. A wrong line in an internal brief gets a light check; anything that touches a customer gets real eyes. Most builds take a few weeks end to end. When it’s running, I hand it over so your team can run it without me. You own all of it — the code, the prompts, the accounts it lives in.
The result: a build your team runs without me, in accounts you own.
At Meta: this stage took procurement approvals from ten days to two, and produced the custom AI tool rebuilding a 300-person org’s budget platform off Airtable. Product owner and builder, same person.
More on this: Your AI is a junior, not an oracle · The two loops every AI process needs · the guides
Check
A month after shipping, I come back to see if it stuck. Stuck means one thing: it’s part of someone’s real week, run because it saves them the hour, not because they were told to. If it didn’t land, you don’t pay. Stuck is measured against the test we wrote down together in agree, before any code, not a feeling either of us argues about. And the deal assumes the build got a fair run: if it never got switched on, or the goalposts moved mid-build, that’s a conversation, not a refund. I also look for drift — models change, edge cases show up that didn’t exist in week one, and a build that worked on Monday can quietly degrade by week six if nobody is watching. What you get from the visit is a short read with numbers in it: what it’s handling, what it costs to run, what got caught in review. If you want me back on a cadence after that, we agree it up front.
The result: proof it stuck, or you don’t pay.
At Meta: the check was measured adoption. Usage tripled every two-week measurement, zero to org-wide, and once I’d trained the first internal implementers it kept growing without me.
More on this: Your AI didn’t break, it drifted · The system that built this site
What this has produced
This shape has a record. At Meta it produced the daily brief that 75+ ops people run before standup, and gave those operators five hours back, every week. It cut a thirty-plus tool sprawl across three teams down to three. It rebuilt procurement approvals from ten days to two. And adoption was measured, not hoped for: usage tripling every two weeks until the whole org ran it. Your week has its own version of those numbers, and finding them is what the first verb is for.
I build the way I advise. The system I do this work with runs my own operation and built this site; the detail is in the proof of work on my about page. And if you want the thinking behind the category, I’ve written up what I mean by applied AI.