Lesson objective
Attach a measured baseline of volume, touch time, elapsed time, first-pass rate and escape point to each step of an as-run plant workflow map, and calculate each figure correctly.
Plain-language explanation
Why this matters now
Module 1 made the case for a baseline: the plant's own numbers, measured the same way before and after, are the only evidence it will have, because this programme's sources contain no manufacturing figure for AI productivity, scrap, downtime or return. This lesson puts the numbers on the map. A baseline attached to steps tells you not only how the workflow performs, but where it performs badly, which is where an AI step might help and where it might hide a problem.
Professional guidance (voluntary framework): NIST AI RMF 1.0 (NIST AI 100-1, January 2023) places MEASURE after MAP for most users, iteratively. It is voluntary and binds no manufacturer. Measuring on the map is a practical way to do both in order.
Professional standard (voluntary consensus standard): ISO says ISO 9001 "can also be used by certification bodies or other third parties to assess an organization's conformity." If your plant is certified, your quality system may already collect some of these measures. This programme has not read the standard's text and quotes no clause, and ISO 9001 contains no AI-specific requirement verified here. The measures below are this programme's recommendation, not a certification requirement.
A question to sit with
Your receiving team says certificate review "takes forever". Is that because each review takes a long time, because certificates wait in a queue, or because a tenth of them bounce back to the supplier? An AI tool fixes at most one of those. Which one does your map show?
What goes wrong
Only total time is measured. A single figure for "how long the workflow takes" mixes the minutes someone spends working with the hours an item waits. An AI tool can shorten the working minutes and leave the waiting untouched, and the before-and-after comparison then shows almost no change and nobody knows why.
Rework is invisible. An item that goes round the loop twice counts once in the volume. Without a first-pass measure, a workflow that now needs more corrections can look just as productive.
Escapes are counted but not placed. A plant knows it had four customer complaints about certificates last quarter. It does not know at which step each error was made or at which step it should have been caught, so it cannot tell whether a new AI step moved the catch point earlier or later.
Shifts are averaged. A day-shift median hides a night-shift queue.
Averages hide the long tail. A mean elapsed time is pulled around by a few items that waited a week. A median with the longest few listed tells the truth more simply.
The measure changes between before and after. A stopwatch before and a tool dashboard after measure different things.
The practical answer
Measure five things on each step of the map.
| Measure | Definition | How to record it |
|---|---|---|
| Volume | Items through the step per month, by shift | From the system of record or a tally |
| Touch time | Minutes a person actually works on one item at this step | Timed sample, same method every time |
| Elapsed time | Time from arrival at the step to leaving it, including waiting | Timestamps from the system of record |
| First-pass rate | Share of items that pass the step without being sent back | Count of items returned at least once |
| Escape point | For each error found, the step where it was made and the step where it was caught | An error log with both fields |
Five measures, per step, by shift. The first three tell you where time goes. The last two tell you where quality goes.
Separate touch time from elapsed time. If a certificate review takes 12 minutes of work but sits 20 hours in a queue, an AI tool that halves the 12 minutes changes the elapsed time very little. The map then shows the queue, not the review, as the problem, and the right fix may be staffing or scheduling rather than AI.
Compute the first-pass rate from returns. Count every item sent back at least once, not every send-back. If 12 of 150 packages were returned at least once, the first-pass rate is (150 − 12) ÷ 150 = 138 ÷ 150 = 92%.
Log escapes by where they were made and where they were caught. For every error found, write the step that created it and the step that caught it. Over a month, the log shows how many errors each review point catches and how many get past it.
In that example an AI tool that drafts discrepancy lists might shorten the 12-minute review. It would do nothing about the 26-hour wait. And the four escapes show that the review step already lets some errors through, so the review design for any AI step must be stronger than the current one, not a copy of it.
Keep the error categories simple and stable. Best practice: use a short, fixed list of error types in the escape log, such as wrong value, wrong revision, missing item and wrong part or customer, so that before-and-after counts compare like with like. Module 13 builds a full error taxonomy.
Record the measurement conditions. For every figure, write the period, the shifts covered, who measured, the method and what was excluded. AI Coalition Network recommendation: keep this with the signed as-run map, so the baseline and the map it describes can never be separated.
Measure what an AI step would change before it exists. If a proposed AI step drafts a document, the relevant measures are the touch time of drafting, the first-pass rate at the review after it, and the escape point of errors it could create. Record all three now, so that the later comparison is possible.
Remember what the numbers are for. The baseline is evidence for a decision a named person makes about whether and how to change the workflow. It is not a target, and it is not a promise to anyone. Do not quote it to a customer as proof that a future AI step will improve anything.
Remember this
- Measure volume, touch time, elapsed time, first-pass rate and escape point on each step, by shift.
- Touch time and elapsed time are different; an AI tool usually changes only the first.
- First-pass rate counts items returned at least once: (total − returned) ÷ total.
- Log where each error was made and where it was caught, so you can see what a review point lets through.
- Record the period, shifts, method, exclusions and measurer with every figure.
Manufacturing example
Hollow Creek Stamping is a fictional stamping plant mapping its first-article inspection package workflow before considering an AI tool to assemble the packages. Quality manager Irena Kovac measures each step over four weeks on both shifts. The plant completes 150 packages a month. Touch time, from a timed sample of 15, is 10 minutes to collect the inspection data, 5 minutes to assemble the package and 3 minutes for the quality engineer to review, 18 minutes in all, so 150 × 18 = 2,700 minutes, or 45 hours a month. Median elapsed time from last inspection to customer submission is 31 hours, and the five longest waits each involved a missing gauge record from second shift. Twelve packages were returned by the reviewer at least once, so the first-pass rate is (150 − 12) ÷ 150 = 138 ÷ 150 = 92%. The error log over the four weeks shows 16 errors: 12 caught by the reviewer, 3 caught by the customer's supplier quality engineer, and 1 caught at the plant's own later audit. So 4 of 16 errors, 25%, escaped the review. Most errors began at the data-collection step on second shift, where the gauge crib records readings on paper. Irena records every figure with its period, shifts, method and measurer, and attaches it to the signed as-run map. Her note to the plant manager says the largest problem on the map is the paper gauge record on second shift, which no AI assembly step would fix, and that any AI step would need a stronger review than the current one, which lets a quarter of errors through.
Responsible-use warning
Do not measure only total time, and do not average across shifts. Separate touch time from elapsed time, or you will credit or blame an AI step for a queue it never touched. Count first-pass rate from items returned at least once. Log where each error was made and caught. And never present a baseline as a target or as a promise to a customer: it is evidence for a decision a named person makes, nothing more.
Human-review requirement
The operation owner takes the baseline, records its period, shifts, method, exclusions and measurer, and keeps it with the signed as-run map. The people who own each step confirm the volume, return and error counts from the system of record. The person the quality system names reviews the escape log, because it shows what the current review points let through. The accountable owner uses the baseline to decide whether and where an AI step belongs, and no figure is shared outside the plant without its limits.
Practical exercise
Use a fictional plant and invented figures only.
- Take an as-run map of one workflow and list its steps.
- For each step, write how you would record volume, touch time, elapsed time, first-pass rate and escape point, by shift.
- Using invented figures, calculate the monthly touch time for the workflow. Show the arithmetic.
- Using invented figures, calculate the first-pass rate at one review step, counting items returned at least once.
- Build an invented escape log of at least ten errors, with the step where each was made and where it was caught. Calculate the share that escaped the main review step.
- Report elapsed time as a median with the five longest items and their reasons.
- Identify whether the workflow's main problem is touch time, waiting, rework or escapes, and say whether an AI step would address it.
- Write the measurement conditions: period, shifts, method, exclusions and measurer.
- Choose a short, fixed list of error types for the escape log, and explain why it must stay the same before and after any change.
- Write the note you would attach to the signed as-run map explaining what the baseline shows, and state plainly that it is not a target or a promise.
Knowledge check
-
What is the difference between touch time and elapsed time, and why does it matter for an AI step?
Answer guidance: Touch time is the minutes a person works on an item; elapsed time includes waiting. An AI tool usually shortens touch time only, so if most elapsed time is queue, the tool changes little and the fix lies elsewhere.
-
Twelve of 150 packages were returned at least once. What is the first-pass rate?
Answer guidance: (150 − 12) ÷ 150 = 138 ÷ 150 = 92%. It counts items returned at least once, not the number of times items were sent back.
-
Why log both the step where an error was made and the step where it was caught?
Answer guidance: It shows how many errors each review point catches and how many escape past it, so the plant can see whether a new AI step moves the catch point earlier or later.
-
Does ISO 9001 require these five measures?
Answer guidance: This programme has not read the standard's text and quotes no clause from it. The five measures are this programme's recommendation. ISO 9001 is voluntary and contains no verified AI-specific requirement.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- International Organization for Standardization (ISO) — ISO 9001:2026 Quality management systems — Requirements (2026-09-16). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.