AI for Creating Procedures
Procedures are how organisations remember how to do things: the payroll run, the system restore, the onboarding checklist. Everyone agrees they should exist; almost nobody enjoys writing them, which is why the real procedure so often lives in one experienced head. AI collapses the documentation barrier — a talked-through walkthrough becomes a structured draft in minutes. The verification, though, is physical: someone has to actually follow the steps.
Why this task matters
The cost of missing procedures is invisible until it isn’t: the expert is on leave and nobody can run the process; the new starter learns by interrupting; the audit asks for the documented process and gets a shrug. Procedures don’t get written because the people who know the process are busy doing the process — documentation always loses to operation. Removing the writing effort changes that equation permanently.
The traditional workflow
- The expert performs the process (documentation postponed)
- Eventually: expert writes steps from memory, skipping the ‘obvious’ ones
- A reviewer who doesn’t know the process finds it unfollowable, or nobody reviews it
- Publication; the procedure drifts from reality with every system update
How AI can help
Draft
- A structured numbered procedure from a recorded walkthrough or rough notes
- Variants: the quick-reference card and the detailed version, from the same source
- The ‘if something goes wrong’ section from the expert’s war stories
Summarise
- A rambling explanation into the essential steps in order
- Multiple people’s slightly different methods into one draft for the team to reconcile
Analyse
- Step-logic checks: missing prerequisites, steps referring to things not yet created
- Ambiguity hunting: ‘which steps would a brand-new person be unable to follow?’
- Comparing the written procedure against a fresh walkthrough after a system change
What must stay human
Which method becomes the standard is a decision — when three people do it three ways, AI can document all three but only humans can reconcile them. Safety-critical and compliance steps need qualified sign-off. The test-by-doing is irreplaceably human: a procedure is verified by someone following it exactly, on the real system, ideally someone who didn’t write it. And ownership — who updates this when the system changes — is a governance decision no draft can make.
Traffic light assessment
Drafting procedures from walkthroughs, for routine internal processes. The source is direct observation; verification is following the steps; errors surface immediately and cheaply.
Procedures involving figures, permissions, or customer-affecting steps. Errors propagate to every future execution — test-by-doing is mandatory, plus review by a second operator.
Safety-critical, emergency, or regulated-compliance procedures. These carry qualified-review obligations and consequences beyond documentation. AI may transcribe the expert; sign-off belongs to the accountable specialist.
Example prompt
For the record-a-walkthrough pattern — the highest-value use:
Turn this transcript of me walking through our monthly invoice reconciliation into a step-by-step procedure. Structure: Purpose (two lines), Before you start (prerequisites and access needed), Steps (numbered, one action per step, exact button and menu names as I said them), If something goes wrong (from the problems I mentioned), Who to contact. Constraints: write for someone doing this for the first time, flag any step where my explanation was ambiguous or I said ‘you know, the usual’ instead of guessing what I meant. [paste walkthrough transcript]
The risks
A procedure with a wrong step is worse than no procedure — it fails with authority. AI fills gaps confidently (‘click Save’ where the real system says ‘Submit’), which is why exact names come from the walkthrough and the draft gets tested by doing, on the real system, before publication. Screens change: procedures need owners and review dates or they rot into traps. And walkthroughs of sensitive systems can capture credentials and personal data — scrub before anything enters a tool.
A better workflow
The current way
- Knowledge lives in one head; documentation forever postponed
- Written-from-memory drafts skip the invisible expertise
- Procedures drift silently after every system change
The AI-assisted way
- Expert records a walkthrough while doing the process normally — fifteen minutes, no writing
- AI drafts the structured procedure, quick-reference and gotchas the same day
- A second person tests by doing; corrections go in; an owner and review date go on
What improves
- The documentation barrier drops to nearly zero — the backlog finally moves
- Drafts capture the spoken war stories memory-writing always loses
- Test-by-doing becomes the effort’s focus, which is where quality was always decided
- Key-person risk shrinks one recorded walkthrough at a time
Key takeaways
- Record the walkthrough; let AI do the structuring — the writing barrier disappears
- Exact button and field names come from the source, never from AI’s assumptions
- A procedure is verified by someone following it on the real system
- Safety and compliance procedures keep their qualified sign-off
- Every procedure needs an owner and a review date, or it rots