An AI work design audit should answer one question: did AI improve the way work happens, or did it get grafted onto the old system?

Do not start with tool usage. Usage is too easy to fake and too weak to trust. Start with one real workflow. Pick something that happens repeatedly and matters: support resolution, renewal prep, month-end variance commentary, contract review, implementation planning, incident response, lead qualification, product discovery synthesis, QA, or operating-review packet creation.

Then walk the workflow from input to outcome.

1. Name the work outcome

What is this workflow supposed to produce? A resolved customer issue, approved contract, accurate forecast, shipped change, completed onboarding, risk decision, clean handoff, or executive decision.

If the outcome is vague, AI will only make vague work faster.

2. Identify the current unit of work

Write down the input, owner, steps, systems, handoffs, review points, exceptions, and definition of done. Include the unofficial parts: Slack checks, spreadsheet lookups, manager memory, copied context, and side-channel approvals.

If people say "it depends," capture what it depends on. That is usually where the design work lives.

3. Look for AI grafting

Ask what changed after AI was introduced. Did the process map change, or did one existing step get an AI helper? Did any handoff disappear? Did review change? Did role expectations change? Did quality standards become clearer? Did exception paths improve?

If the answer is mostly no, you probably grafted AI onto the old workflow.

4. Assign the mode for each step

For each meaningful step, choose assist, review, exception, or automate.

Do not assign one mode to the whole workflow. A single customer support flow may use all four. A finance process may automate low-risk reconciliation, assist variance explanation, review commentary, and escalate unusual movements.

Mode decisions should reflect risk, reversibility, input quality, judgment required, and customer impact.

5. Make hidden work visible

Where are people assembling context, checking facts, cleaning outputs, updating records, fixing tone, handling edge cases, or teaching the system through repeated corrections?

Hidden work is not failure. It is design material. Decide what should be removed, supported, automated, or formally owned.

6. Measure review capacity

Who reviews AI-assisted work? How much time does it take? What are they checking? What slips through? Where is the queue growing? Which reviews could move to sampling, automated checks, or exception-based review? Which reviews need to become stricter?

If review capacity is not measured, it will be consumed quietly.

7. Redesign the role

For each role in the workflow, clarify what AI now handles, what the person owns, which judgments matter more, what gets escalated, and what quality means.

If the role definition has not changed but the work has, people will invent their own version of the job.

8. Write the accountability contract

Name the outcome owner. Define the quality bar. Define review responsibility. Define exception ownership. Define evidence requirements. Define who changes the system when it fails.

This can be lightweight. It cannot be absent.

9. Check the metrics

Old metrics may push the wrong behavior. If AI speeds drafting, volume metrics may reward noise. If AI improves classification, accuracy and rework may matter more than speed. If AI shifts work to review, review load and exception rate should appear in the dashboard.

Measure the work unit, not the tool.

Useful metrics include cycle time, rework, reopen rate, escalation rate, review time, exception volume, quality defects, customer impact, manager touch time, and learning-loop completion.

10. Create the improvement rhythm

A redesigned workflow still decays. Sources go stale. Prompts drift. Policies change. People route around bad steps. Review rules become outdated. New exceptions appear.

Set a rhythm for improving the work system. Weekly may be right for high-volume workflows. Monthly may be enough for slower ones. The review should ask: what failed, what repeated, what can be removed, what should be escalated sooner, what can now be trusted more?

The audit should end with a decision.

  • Keep the workflow in assist mode while learning.
  • Add review checks before scaling.
  • Move a stable slice to exception-based handling.
  • Automate a low-risk step.
  • Stop the pilot because hidden work exceeds value.
  • Redesign the role before expanding adoption.
  • Fix source data before blaming the model.

That is the point of the audit. Not to admire AI activity. To decide how the work should change.

A good AI workflow feels calmer over time. Fewer unclear handoffs. Better exceptions. Cleaner ownership. Less duplicate checking. Stronger roles. More useful review. Faster learning.

A bad one feels like more output with more managerial anxiety.

That difference is the signal. Follow it.

This is part 10 of 10 in Work Design for the AI Era.