Why boring AI consultancies win.
Most AI consultancies sell narrative. Almost none ship software. Why the boring path — fixed bids, on-call, telemetry — is the actually scarce thing in 2026.
F.001 · Read note
Field Notes · Vol. III · 04.2026
Notes from the field.
Mostly opinionated.
01 — Latest
Notes ordered the way we’d want to read them.
What we learned shipping to 19 rural clinics.
The model was the easy part. The hard part: the receptionist's screen, the WiFi that dies at 3pm, and the doctor who's been billing on paper since 1997.
F.002 · Read note
Eval harnesses that catch real bugs.
Public benchmarks lie. We share the harness we built for criminal defense workflows and what it caught that MMLU never would.
F.003 · Read note
A defense attorney’s first week with CriminalFlow.
Years of paper case files, indexed in a weekend. What the paralegal flagged on day three. The feature we cut after seeing what they actually used.
F.004 · Read note
AI is the floor, not the ceiling.
If the model can't get you to MVP, you don't have a product. If the model is the only thing standing between you and competition, you don't have a moat. Both are common; both are fixable.
F.005 · Read note
Why our case studies say ‘log in and break it.’
Slideware is cheap. Logins are not. The clients I want are the ones who'll click around for ten minutes before the kickoff call.
F.006 · Read note
The vertical SaaS playbook, in ten-plus live apps.
What survived from one industry to the next. What didn't. The four primitives I copy verbatim and the three I rebuild every time.
F.007 · Read note
One operator, too many pagers.
How on-call works when you're the whole team. The runbook that runs the runbook. Where AI agents take the page and where they should not.
F.008 · Read note
Quarterly digest · No spam
The field-notes digest,in your inbox.
Four times a year — the essays, the field notes, the post-mortems, and one thing we shipped that we’re still arguing about internally. Read in seven minutes, archived forever.