Turn Repeated Work into Clear Procedures

Use AI to organize observed work into instructions a new person can follow.

An experienced volunteer says, “Check the donation, put it in the right place, and update the log.” A new volunteer asks which donations are accepted, where the log is, and what to do when the count does not match.

The expert has not been careless. Much of the process has become automatic to them. They no longer notice the small decisions that make the work succeed.

AI can help reveal those missing decisions, but it cannot observe a workplace from a vague instruction. A useful procedure begins with evidence about how the task is actually performed and how the organization wants exceptions handled.

Capture the work before polishing it

Ask an experienced person to demonstrate a normal case and one exception. Record the trigger, tools, inputs, decisions, actions, and evidence of completion. Note where they consult another person, check a label, or use an unofficial workaround.

Separate the observed process from the proposed process. If the worker skips a required check, the solution is not to document the shortcut as approved practice. A process owner must resolve the difference between policy and behavior.

For this article, use an entirely fictional exercise: receiving donated notebooks at a community learning center. The observation packet is invented. It says volunteers count notebooks, inspect them, place acceptable ones on Shelf A, and enter a batch record. It also reveals uncertainty about damaged items and duplicate delivery records.

That packet is enough to draft questions. It is not enough to claim the organization has an approved procedure. For the worked example below, we supply explicit fictional rules to resolve the gaps.

Define the scope and completion condition

The fictional procedure covers paper notebooks only. Other donations go to the shift coordinator without being processed under these instructions. The procedure starts when a notebook delivery arrives and ends when every received notebook is assigned an outcome and the batch totals reconcile.

Required materials are the donation log, a unique batch label, a counting surface, Shelf A for accepted notebooks, and a labeled Hold bin for items awaiting coordinator review. The shift coordinator owns exceptions and approves final disposition of held items.

For this exercise, an acceptable notebook has blank usable pages, an intact binding, and no visible damage requiring review. A notebook that fails one of these checks goes to Hold. Volunteers do not remove pages, repair items, or decide their final disposal under this procedure.

A clear completion condition is: received count = accepted count + held count, with all three values recorded and physical placement matching the log. “Donation processed” is too vague to test.

Write steps someone can observe

Here is a complete miniature procedure for the fictional task.

Procedure: Receive notebook donations
Version: 1.0, exercise draft
Owner: Shift coordinator
Review trigger: Any failed trial, change to acceptance rules, or change to log/storage locations
Status: Ready for a training trial; not validated in a real organization

1. Confirm scope. Check that the delivery contains paper notebooks. Set other items aside in the designated incoming area and ask the shift coordinator to handle them separately.
2. Create the batch record. Obtain an unused batch ID from the donation log. Record the date, your initials, and the batch ID on both the log and batch label. Do not invent donor details or collect them unless the organization’s approved process requires them.
3. Check for an existing record. If the delivery already has a batch label or matching entry, stop and ask the coordinator whether it has already been logged. Do not create another count while the duplicate is unresolved.
4. Count the notebooks. Place the batch on the counting surface and count each notebook once. Record this as the received count before sorting.
5. Inspect each notebook. Check usable blank pages, binding, and visible condition using the acceptance rule. Put acceptable notebooks in one group and items needing review in another. If unsure, use the review group.
6. Count the two groups. Record accepted and held quantities. Check that their sum equals the received count. If it does not, keep the batch together and recount before proceeding.
7. Place and label the groups. Move accepted notebooks to Shelf A with the batch ID. Put held notebooks in the Hold bin with the same batch ID and a short reason, such as damaged binding.
8. Complete the log. Enter the storage locations and mark the batch “received and sorted.” Record held items as awaiting coordinator review, not accepted inventory.
9. Hand off exceptions. Tell the shift coordinator about held items or unresolved issues using the approved local handoff method. Record the handoff when it actually occurs.
10. Check completion. Confirm the counts reconcile, both groups carry the batch ID, and the log matches their locations. Leave any unresolved batch open with the reason documented.

The procedure identifies actions, locations, and decisions. In a real workplace, replace fictional labels with verified names that a newcomer can find. “Use the usual spreadsheet” is not a location.

Work through a normal case and an exception

Suppose a synthetic batch contains 15 notebooks: 12 meet the acceptance rule and three have damaged bindings. The record should show received 15, accepted 12, held 3. The arithmetic reconciles because 12 + 3 = 15.

The accepted group goes to Shelf A; the held group goes to the Hold bin. The batch can be recorded as received and sorted after placement and logging, while the three held notebooks remain awaiting a disposition decision. Do not confuse completion of receiving with completion of every later review.

Now imagine the initial count says 15 but the groups total 14. The instruction is to keep the batch together and recount. AI should not “correct” the received count to 14 merely to make the table balance. Reconciliation is an investigation, not an invitation to alter evidence.

If the log is unavailable, the real process needs a defined contingency. For the exercise, pause the batch and ask the coordinator; do not guess the next ID or create an untracked personal record. In another organization, an approved temporary log might be appropriate. Document the actual decision.

Ask AI to find the hidden assumptions

Review this draft procedure as if you are a new volunteer. Identify undefined terms, missing prerequisites, ambiguous locations, unsupported policy decisions, and actions without an observable completion condition. For each issue, propose a clarification question. Keep observed facts separate from suggested process improvements. Do not mark the procedure tested or approved.

After the questions are resolved, ask for a shorter training checklist derived from the approved procedure. The checklist should reference the same rules rather than inventing a second process.

A useful checklist for the example is: correct scope, unique batch, received count, inspection, accepted/held counts, reconciliation, placement, log, and exception handoff. It is a memory aid for trained users; a newcomer may still need the full instructions.

Test without silently coaching

Give the draft to someone unfamiliar with the task and ask them to process a practice batch. Explain that you are evaluating the instructions, not their intelligence or suitability. Use safe training materials and allow them to stop whenever uncertain.

Observe where they hesitate, ask a question, skip a step, or interpret a term differently. If you must intervene, record the intervention. Otherwise, you may later claim the procedure worked independently when your coaching supplied the missing instruction.

A trial record can say: “At step 3, the reader could not determine which log field identifies an existing batch. Revision: name the batch-ID column and show a fictional example.” That is a concrete improvement. “Needs more detail” is not yet an edit.

Maintain the procedure as work changes

Assign an owner, version, and review trigger. Record what changed and why. Archive superseded copies so users can find the current version without comparing several files themselves.

A procedure may drift when the room layout changes, a system field is renamed, or a policy exception becomes common. Invite users to flag these changes. AI can compare versions and draft a change summary, but a process owner must verify that the new instructions reflect the intended work.

For equipment operation or specialized tasks, use the relevant approved instructions and qualified review. A general language model should not invent technical limits or substitute a generic procedure for the equipment’s actual requirements.

For students: document a process another team can repeat

A class can use notebook receiving, equipment checkout, or a fictional office workflow. One group drafts from observations; another follows the instructions without seeing the original demonstration. Compare the result with the defined completion condition.

Follow course rules for AI assistance and record revisions. If no trial occurred, submit the procedure as untested with a trial plan. The exercise rewards clear communication and reproducibility rather than claims of professional deployment.

Practice: draft, trial, and revise a procedure

Use the fictional notebook packet or an authorized low-risk process. Write scope, prerequisites, numbered steps, exception handling, and completion criteria. Ask AI for ambiguity questions. Conduct a training trial if possible, then revise from observed problems.

Completion check: A reader can follow the procedure, identify the exception owner, and recognize when the task is complete. Your version record distinguishes proposed instructions from approved rules and actual trial observations from planned checks.

For a stretch exercise, create a one-page training checklist and show which full procedure step supports each item.