An AI throughput audit is a way to answer one plain question: did AI make the system faster, better, or merely busier?

Use the audit on one workflow at a time. If you audit AI “across the company” in one pass, you will get mush. Pick a real flow with a start, a finish, owners, queues, and a quality bar.

1. Define the work stream

What work are we auditing? Where does it start? What counts as accepted done? Who is the customer of the work? What business outcome does it support?

If the finish line is vague, fix that before measuring AI. A vague finish line guarantees vague ROI.

2. Name the throughput unit

What unit crosses the finish line? A resolved ticket, shipped increment, approved contract, completed onboarding milestone, accepted analysis, closed decision, qualified opportunity, or something else?

Do not count raw AI artifacts unless the artifact itself is the accepted output.

3. Map the pre-AI flow

List the steps, queues, handoffs, reviews, decisions, and rework loops. Capture rough baseline numbers: elapsed time, WIP, queue time, review effort, rework rate, acceptance rate, quality misses, and cost where available.

Precision is useful. Directional truth is enough to start.

4. Identify the original constraint

Where did work actually wait or break before AI? Creation, intake, context gathering, review, decision, QA, integration, approval, customer validation, or expert capacity?

If the AI use case did not touch the constraint, lower your ROI expectations.

5. Describe the AI intervention

What changed? Assist, automate, review, route, summarize, classify, draft, check, synthesize, recommend, or escalate? Who owns the output? Who reviews it? What cases are excluded?

Vague intervention descriptions make measurement impossible.

6. Separate local productivity from system throughput

What got faster for the person using AI? What got faster for the whole workflow? Did elapsed cycle time fall? Did accepted outputs per week rise? Did queue time shrink?

Keep both numbers. The gap between them is where the learning is.

7. Adjust for quality

Did first-pass acceptance improve? Did rework fall? Did reopened cases, defects, corrections, complaints, reversals, or policy misses change? Did reviewer trust improve or degrade?

Faster bad work is not throughput. It is debt with a nicer interface.

8. Measure review and decision load

Did AI reduce expert attention or move more work onto experts? Did decision latency improve? Did managers receive cleaner decision packets or larger piles of analysis?

Review capacity is often the new bottleneck. Treat it as a measured resource.

9. Check WIP and output inflation

Did the workflow accumulate more drafts, tickets, options, experiments, or analyses than it can finish? Did AI create more inventory? Are people spending more time sorting generated work?

More output can lower throughput when the system cannot metabolize it.

10. Name the new constraint

After the intervention, what limits throughput now? Is the new constraint better than the old one? Is it strategic judgment, quality review, decision rights, customer validation, integration, cost, or governance?

This is the heart of the audit. AI ROI is constraint movement plus quality-adjusted throughput.

11. Decide the next action

Choose one:

  • scale the intervention
  • redesign the workflow
  • add or protect review capacity
  • clarify decision rights
  • tighten the quality bar
  • segment the workflow
  • stop the use case
  • run another focused experiment

Do not end with “monitor.” Monitoring without an operating action is usually avoidance.

12. Record the learning

Write down what changed, what did not, what moved, and what the next constraint is. The audit should become operating memory, not a slide that disappears after the AI steering meeting.

The best outcome of the audit is not always “AI worked.” Sometimes the best outcome is discovering that the glamorous use case is irrelevant to the constraint, or that the real opportunity is a boring pre-review check, or that the team needs decision rights more than another model.

That is still progress.

The audit protects leaders from confusing adoption with leverage. People can use AI heavily and still leave the system unchanged. They can save time locally and still produce no more accepted outcomes. They can generate more artifacts and make everyone busier.

The final question should be asked without decoration:

What can the system now do faster, at acceptable quality, that it could not do before?

If the answer is clear, you have leverage.

If the answer is vague, keep auditing.

This is part 10 of 10 in From Productivity to Throughput.