ops studio

The process lives in Vera's head

Every ops team has a Vera. She's the one who knows that the Monday stuck-order sweep only works if you pull the export before the 9am sync, that the requeue has to happen one order at a time because bulk requeue silently drops the hazmat flags, that the third vendor invoice of the month is always wrong in the same way. None of it is written down. When Vera takes a vacation the process limps; when Vera resigns the process dies.

AI changed the economics of writing it down. The documented practitioner pattern — record Vera narrating the task once, let a model structure the transcript into a stepped procedure — collapses SOP-writing from an afternoon nobody schedules into a draft you review over coffee. It also introduced a brand-new failure: the model fills every gap in the transcript with plausible invented steps. Menu paths that don't exist. Defaults that are wrong. An SOP that reads beautifully and lies.

This chapter is the ops lead's answer, built as code you run: SOPs as versioned data with a lint that blocks unverified model-filled steps, recurring checklists where the machine computes what's due but a named human signs what got done, exception queues so a failed run can never impersonate a clean one, and a verification pass that traces every number in the weekly report back to a cell in the export before anyone reorders inventory off it.

Why it matters

• Redesigning the process is where the money is. A July 2025 Gartner survey of 1,973 managers found business units that redesign how work gets done with AI are twice as likely to exceed revenue goals — while MIT's 2025 GenAI Divide report found 95% of enterprise genAI pilots delivered no measurable P&L impact (measured on P&L only, so efficiency gains are undercounted — but the direction stands: value concentrated where workflows were actually redesigned). Buying a tool moves nothing. Rebuilding the checklist, the SOP, and the report pipeline around it is the part that pays, and that part is this chapter.
• Some sign-offs are laws, not formalities. Under the OSH Act's General Duty Clause, an AI-drafted SOP that omits a safety step moves nothing off the employer — the procedure's human owner holds the duty. OSHA's recordkeeping rules (29 CFR 1904) put the injury log and the posted annual summary on named people at your company. An automation that auto-completes a compliance item doesn't discharge the duty; it hides that no one did.
• Nobody else teaches this line. The training-market scan behind this path found no scanned offering that teaches operators where the auto-run/named-approval threshold sits for AI-initiated actions, and SOP change-control as a versioned, owned practice is essentially untaught. The market sells tool tours to executives. You run the dock. This is for you.

What AI specifically gets wrong

• It invents the steps you didn't narrate. Feed a model a transcript with a gap and it writes "Open Admin > Batch Tools > Requeue All" whether or not that menu exists. Practitioner accounts of the transcript-to-SOP workflow report exactly this: hallucinated menu paths, caught only because the person who does the task read the draft. A gap should ship as "unknown — ask," never as a confident guess.
• It completes what it cannot legally complete. A checklist bot will happily mark the forklift inspection done. Nobody inspected the forklift. The machine's job is computing what's due; a named human signs what's done — that split survives on purpose, because the duty behind the sign-off doesn't delegate to a cron job.
• It fails silently and calls it success. The export wasn't there, the source was stale, the step errored — and the run summary still says all clear, because nothing routed the failure anywhere. A silent failure is worse than a loud one: it spends your trust in the whole system. Exceptions go in a queue with a reason and an owner, or they go nowhere.
• It writes numbers the export never produced. An AI-drafted ops report is a draft over data: every cycle time, fill rate, and late-order count either traces to the export or it's fiction with good formatting. Exposure analyses specifically warn that ops roles built on report generation are the compressible ones — which is exactly why the verifying, not the drafting, is where your value moves.

What you'll be able to do at the end

• Capture a real process as versioned data — owner, version, date, steps tagged by source — and lint it so no unverified model-filled step and no unreviewed safety step ships in v1 or any version after.
• Release SOP changes the way changes should ship: as a diff, version bumped, reviewed by a named human who isn't the drafter.
• Encode the team's recurring duties as a schedule file, compute this week's due list from data instead of memory, and route every item across the explicit line: auto-run for the reversible drafts, a named approver for everything with a duty or a dollar behind it.
• Run the nightly checklist so failures land in an exception queue with reasons — and read out a run report where "clean" actually means clean.
• Verify every number in an ops report against the export it claims to come from, catch the invented one, and make the SEND / HOLD call with the failing metrics named.

The drills run in plain Python on data shaped like your Tuesday. They're warehouse-flavored because the stakes are easiest to see on a dock — but nothing here depends on the forklift: swap the inspection log for the fire-extinguisher check, the WMS export for the facilities ticket queue or the vendor invoice ledger, and the same lint, threshold table, and receipt check run an office-ops shop unchanged. The logic is the point — it works whether it lives in a script, a spreadsheet, or whatever workflow tool procurement buys next year.