The checklist that lied
A warehouse team wires their recurring checklist into an automation. It's a good automation — it knows the cadences, it fires on schedule, it never forgets. Six months later an auditor asks who performed the forklift daily inspections in March. The checklist says they happened. Green ticks, every day.
Nobody inspected the forklift. The bot marked the item done because the item was due, and "due" and "done" had been wired to the same wire. The checklist didn't just fail to catch the miss — it manufactured evidence that there was no miss. That's strictly worse than having no checklist at all, and it's the failure this whole lesson exists to make impossible.
Some sign-offs are laws
For an ops lead, this isn't a philosophy question. The floor is statutory:
- Under OSHA's recordkeeping rule (29 CFR 1904), covered employers must maintain the Form 300 log of work-related injuries and illnesses, post the 300A annual summary where employees can see it, and — depending on size and industry — submit the data to OSHA electronically, with annual submissions due March 2. Employees have the right to review the log. These are duties held by the employer and its named people.
- Under the General Duty Clause, the workplace must be kept free of recognized hazards — a duty that doesn't move onto whatever tool drafted or scheduled the procedure.
A script that auto-completes "post the 300A summary" hasn't discharged a duty. It has hidden the fact that nobody did. When the gap surfaces — audit, incident, lawsuit — the record your own automation faked is the first exhibit.
The line, drawn explicitly
So every recurring item gets classified, on purpose, before automation touches it:
- Auto-run — reversible, no duty, no dollar: regenerate the report draft, refresh the due list, compile the reorder suggestions. If it runs wrong, you rerun it. The machine may do these end to end.
- Named approval — anything with a legal duty, a signature, a safety act, or spend behind it: inspections, compliance postings, invoice approvals, anything that binds the company. The machine computes that it's due and routes it to a named human. Only that human marks it done.
Here's the part that should annoy you: the training market barely teaches this. The scan behind this path found no scanned offering that teaches operators where the auto-run/named-approval threshold sits for AI-initiated actions — what a system may execute versus what needs a human, and who sets the threshold. It's taught implicitly, if at all, as Lean tradition. You're about to encode it as a field in a data structure, which is where a threshold actually holds.
Next: the checklist as schedule data, with the line built in.