Thinking

Field notes.

Notes from building AI that has to work: evals, sign-off surfaces, systems that survive contact with real data.

17 ESSAYS · 83 MIN TOTALEST. FEB 2026

  1. · 28 July 2026

    Document intake in an accounting practice: what actually eats the hours

    Client records arrive in every format a client feels like sending, and somebody re-keys them. A system can read the pile and file it. A qualified person still signs anything a regulator will read.

    6 MIN READ

  2. · 11 July 2026

    Private AI, in-house: the shift to internal models and agentic infrastructure

    A growing set of regulated institutions are moving AI behind their own walls: open models, agentic infrastructure, and a human on the consequential call. The constraint was never the model.

    6 MIN READ

  3. · 10 July 2026

    Why your e-commerce ad pipeline stalls at ten variations

    Ad performance runs on variation, and hand-run creative stalls around ten a week. The bottleneck is production, not ideas.

    5 MIN READ

  4. · 9 July 2026

    Fraud screening at 16 million decisions a month: what holds

    Sixteen million decisions a month for more than five years. What keeps a screening system reliable is never the model.

    5 MIN READ

  5. · 8 July 2026

    From dashboards to decisions: what AI analytics is for

    A dashboard shows the past and waits for a human to act. A decision system reads the data, proposes the call, and keeps a person on the ones that matter.

    5 MIN READ

  6. · 7 July 2026

    The AI app you ship in eight weeks: what production-ready means

    A demo is a proof of possibility. Production-ready is a proof of reliability, and it means a specific set of things or it means nothing.

    5 MIN READ

  7. · 6 July 2026

    Document intelligence for clinic groups: intake without the retyping

    Multi-location clinics type the same patient details three times. Document intelligence reads the referral and drafts the entry; a person approves in one click, and nothing patient-facing runs alone.

    5 MIN READ

  8. · 5 July 2026

    What AI can actually automate in a 3PL back office (and what it can't)

    A 3PL back office runs on rekeying. AI can take the typing, the anomaly-spotting, and the status-chasing. It should not touch the rate exceptions or the disputes.

    5 MIN READ

  9. · 20 June 2026

    Strategy that ships

    AI strategy is worthless as a deck. It is worth a lot as a working proof.

    4 MIN READ

  10. · 10 June 2026

    Human-in-the-loop is a feature, not a disclaimer

    The teams that win with AI design the sign-off in, not bolt it on.

    4 MIN READ

  11. · 1 June 2026

    Why most AI never reaches production

    The gap between a demo and a system that runs is where most AI projects die. Here's what closes it.

    5 MIN READ

  12. · 27 May 2026

    Volume is a system property

    Five times the ad output did not come from better prompts. It came from a pipeline with thirteen states.

    4 MIN READ

  13. · 6 May 2026

    In regulated care, design the escalation first

    Twenty medical agents in production: the design that made them safe to run.

    4 MIN READ

  14. · 14 April 2026

    Retrieval at 30 million documents

    Search quality is an evaluation problem long before it is an infrastructure problem.

    5 MIN READ

  15. · 24 March 2026

    Bank-grade is a discipline, not a badge

    Sixteen million fraud decisions a month will teach you what validation actually means.

    5 MIN READ

  16. · 5 March 2026

    What 1,322 automations taught us about 30 industries

    The shapes repeat. The last mile never does.

    5 MIN READ

  17. · 18 February 2026

    Integrate, don't replace

    The fastest way to kill an AI project is to make it wait for a migration. Layer on top of the stack that already runs.

    5 MIN READ
WRITTEN BY THE BENCH · NO GHOSTWRITERS

MORE WHEN THE WORK TEACHES US SOMETHING WORTH WRITING DOWN.

IF ANY OF THIS MAPS TO A PROBLEM YOU ARE CARRYING · SEND THE BRIEF. A PRINCIPAL READS EVERY ONE.