Lesson objective

Explain how volume, shift handoffs, line and customer variation and shared inputs change the risk of an AI-assisted plant task, and convert a measured correction rate into an expected monthly count.

Plain-language explanation

Why this matters now

Level 2 taught one careful person to run one AI-assisted plant task well: triage it before any tool sees it, prompt it from the released document set at a stated revision, check every value against the print, the router or the specification, and keep the task AI use record while the work happens. Level 3 is about what happens when that task stops being one person's habit and becomes something the plant runs on every shift, every line and every customer programme.

A shift-handover summary drafted by one experienced production supervisor, who reads it against the downtime log and the scrap tags before she posts it, is a task. The same summary produced at the end of every shift on four lines, posted by whichever lead is closing out, is a system. The tool can be identical. The risk is not.

Be clear about why a plant scales. This programme's source register contains no manufacturing study that shows a productivity, scrap, downtime or return figure for AI, and no figure for how many manufacturers use it. The nearest controlled evidence does not come from a plant at all. Research (controlled field experiment, consultants): Fabrizio Dell'Acqua and colleagues (Organization Science, vol. 37, issue 2, March 2026) ran a preregistered experiment with 758 Boston Consulting Group consultants, randomly assigned to work with no AI, with GPT-4, or with GPT-4 plus a prompt engineering overview. On a complex task chosen to sit outside the AI capability frontier, subjects using AI "were 19% less likely to produce correct solutions compared with those without AI." These were consultants using a 2023 model, so the finding reaches plant work only by analogy. The lesson it carries matters at scale: the same tool helps on some tasks and hurts on others, and nothing on the screen tells the user which kind of task is in front of them.

The reason to run AI-assisted plant work as a disciplined operation is therefore the consequence of an unverified output reaching an operator, an inspector, a customer or a regulator, multiplied by the number of runs.

A question to sit with

A quality engineer who checks every AI-drafted nonconformance narrative against the inspection record catches a wrong dimension or a wrong lot number about twice a month on her one machining cell. Next quarter the plant runs the same method on four lines across three shifts, with nine people drafting. How many wrong narratives a month should the plant expect, who will catch them at two in the morning, and how would anyone know if they stopped being caught?

What goes wrong

Volume turns a rate into a count. A method that needs correcting in a small share of outputs feels safe when it runs a few times a week. Run on every handover, every nonconformance narrative and every supplier message across a plant, the same share becomes a steady flow of wrong part numbers, wrong lot numbers and wrong quantities. Nothing about the method changed except how often it runs.

Shift handoffs strip away context. The engineer who built the method knew that one customer's prints give dimensions in millimetres with inch reference values, and that the second-shift gauge crib logs readings under the operator's badge rather than the inspector's. The third-shift lead who inherits the method knows neither. Each handoff, from the builder to the team, from first shift to second, from one line to the next, passes on the output and loses some of the knowledge needed to judge it.

Lines, customers and programmes multiply the variation. An automotive stamping line, a medical-device cell and a defence subcontract have different requirements, print conventions and data rules. A method tuned on one produces fluent output on another that fits badly.

Shared inputs make errors travel together. Nine people each working their own way make different mistakes. Nine people using one shared prompt, one narrative template or one library of part descriptions make the same mistake on every line at once. Consistent output then looks like quality when it is really a common cause.

Silent errors raise no flag. Most operations count exceptions: a file that would not open, a blank field. A confidently wrong output passes every such counter, which is why an exception count is not an error rate.

The review designed for day shift does not survive the night. On first shift the engineer who wrote the method is down the hall. On third shift the lead has no one to ask, a line waiting to start, and a summary that reads well.

The practical answer

Start by separating the risks that scale multiplies from the risks it creates. They need different controls.

Risk Present in one careful task? What scale does Control that addresses it
An invented value, revision or part number Yes Multiplies it Checking against the released document, as in Level 2
Customer or controlled data pasted into the wrong tool Yes Multiplies it Tool approval and access control (Module 7)
Shifts or lines running different versions of the method Barely Creates it A written procedure with one owner (Module 4)
Checking quietly stops Barely Creates it Sampling and quality review (Module 13)
One defect in a shared template reaches every line No Creates it Version control and staged rollout
Nobody owns the operation or may pause it No Creates it A named owner and stop line (Lesson 3)

Rows three to six are what this level means by operations. A better prompt fixes none of them.

Do the frequency arithmetic before rollout. Take the correction rate your plant has actually measured, multiply it by the planned volume, and write the answer as a sentence.

If that sentence is unacceptable, the question is not whether the tool is good. It is which control changes the number: a narrower task, a heavier review step, a smaller rollout, or a different point in the workflow.

Ask four questions of any task you plan to scale. These are an AI Coalition Network recommendation, not a legal or professional requirement.

Question Why it matters If the answer is worrying
How often will it run? Frequency converts a rate into a count State the expected monthly count of wrong outputs
Who will run it, on which shifts, lines and customer programmes? More people and more programmes mean more variation Write the procedure down before widening use
How far is the reviewer from the source? Review needs the print, router or record, not a summary of it Keep a sampled path back to the released document
What do all the runs share? Shared inputs cause errors that arrive together Put every shared prompt, template and library under version control

Treat anything shared as a controlled asset. A prompt used by nine people is not a personal note. Best practice: give every shared prompt, template, library and tool setting a version number, an owner, a change record and a way back to the last good version, the same way the plant already controls a work instruction.

Widen in stages and watch between them. Start on one line and one shift, then add a shift, then a line. A defect found on one line and shift costs that rework; found on four lines and three shifts, it costs twelve times as much, plus the work of tracing which version each shift used.

Keep the tool on the read side of the plant. AI Coalition Network recommendation: at this tier, nothing an AI tool does writes to a control system. It may read from the ERP, MES, QMS, maintenance system or historian; it never reaches a PLC, a robot controller, an HMI or a safety function. A model output is an input to a decision a named person makes. A flag is not a disposition, a drafted narrative is not a nonconformance decision, and a summary is not a release.

Keep the accountability lines where they were. However many shifts run a method, the person the quality system names still dispositions a part and signs a certificate of conformance, the employer still owns every duty under the Occupational Safety and Health Act, and the officer who signs an export or government-contract representation still signs it. Lesson 3 names these owners.

Map before you manage. Professional guidance (voluntary framework): NIST's AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, January 2023) says that, after instituting its GOVERN outcomes, "most users of the AI RMF would start with the MAP function" and continue to MEASURE or MANAGE. It is voluntary, binds no manufacturer, and NIST's web page says 1.0 is being revised; nothing here describes a revised version. The point for operations is plain: understand how the work actually runs before deciding how to control it. Module 2 builds that map.

Remember this

  • Scale multiplies some risks and creates others: version drift between shifts and lines, fading checks, errors that arrive together and operations nobody owns.
  • Convert your measured correction rate into an expected monthly count and write it down before rollout.
  • An exception count is not an error rate; a confidently wrong output raises no flag.
  • Every shared prompt, template and library needs a version, an owner and a way back, and the tool stays on the read side of the plant.
  • No source in this programme shows an AI gain in manufacturing; the case for discipline is the consequence of an unverified output.

Manufacturing example

Hollow Creek Stamping is a fictional tier-two automotive supplier running four press lines on three shifts. Production supervisor Dana Okafor builds a method on Line 2: an AI assistant drafts the end-of-shift handover summary from the downtime log, the scrap tags and the lead's notes, and Dana reads every line against the log before posting it. Over two months she corrects the prompt twice, once to stop it rounding scrap counts and once to make it list open quality holds first. Plant manager Victor Salas likes the results and rolls the method out to all four lines and all three shifts. Four months later, two things have happened that nobody decided. First, the prompt has forked: first and second shift use the version Dana emailed in week one, but third shift on Line 4 uses a copy a lead shortened, which drops the instruction to list open quality holds. Nobody recorded which shift used which version. Second, the checking has thinned. The procedure still says to agree each summary with the downtime log, but at the end of a long shift leads check the scrap figure and skip the rest. In month five, first shift on Line 4 runs a coil that third shift had placed on hold for a burr problem, because the hold never appeared in the handover. The parts are caught at the plant's own final inspection, but Victor asks two questions: which summaries used the shortened version, and in which weeks. Nobody can answer from the records, and the team rebuilds the list from the posting times over two days. Dana's method was sound. The operation around it had no version control, no sampling of the check and no named owner.

Responsible-use warning

Do not treat a successful pilot as proof that a method is safe to run on every shift and line. A pilot run on day shift by the person who built it, on a line and a customer they know, with checking they do themselves, tests the method under the one condition that disappears at scale. Version drift between shifts, fading checks at the end of a long shift, a shared defect reaching every line and an operation nobody owns are all hidden in a pilot, because one careful operator suppresses them.

Human-review requirement

The person who owns each output still owns it at scale: the shift supervisor for the handover, the quality engineer or the person the quality system names for a nonconformance record and its disposition, and the officer who signs any export or government-contract representation. At the operations level the operation owner also judges whether the review actually performed on every shift still matches the review the procedure describes, and whether the volume is within what named reviewers can genuinely examine. If not, their job is to narrow the scope, add reviewers or pause, never to sign a record that overstates what was checked.

Practical exercise

Use a fictional plant or anonymised material with no real customer, supplier or employee data.

  1. Choose one AI-assisted task your plant performs or plans to perform, such as shift-handover summaries, nonconformance narratives or supplier status messages. Describe the steps as they are actually done, in no more than one page.
  2. Record the current volume in outputs per month, and the volume if every shift and line adopted the method.
  3. From your last 20 reviewed outputs, count how many needed a correction. Calculate the rate, multiply it by the projected volume, and write the result as one sentence.
  4. List every shared input: prompts, templates, libraries and tool settings. Beside each, write who owns it and how a change would be recorded today.
  5. Identify the steps that exist only to catch errors. For each, write what evidence, other than asking people, would show that the step had stopped happening on night shift.
  6. Name the shared input whose defect would reach the most lines at once, and estimate how many outputs a month that would be.
  7. Using the risk table in this lesson, mark each risk in your method as multiplied or created by scale.
  8. Confirm the tool reads only from business systems and writes to no control system, and mark any point where an output could reach a disposition, a release or a signed representation without the named person seeing it first.

Knowledge check

  1. Which risks does scale multiply, and which does it create?

    Answer guidance: It multiplies risks already present in one careful task, such as an invented value or controlled data pasted into the wrong tool. It creates risks that barely exist in one task: shifts running different versions, checking quietly stopping, one shared defect reaching every line, and an operation nobody owns.

  2. A pilot found 3 corrections in 150 handover summaries, and the plant plans 1,350 a month. What should the rollout note say?

    Answer guidance: The rate is 3 ÷ 150 = 2%, and 1,350 × 0.02 = 27, so the note should say the plant expects about 27 wrong summaries a month and that the review design must find them before an incoming shift acts on one.

  3. Why is an exception count not an error rate?

    Answer guidance: Exceptions are items the tool itself could not handle, such as a file that would not open or a blank field. A confidently wrong output raises no exception, so only checking outputs against the released source document reveals it.

  4. What did the Dell'Acqua experiment find outside the AI capability frontier, and how far does it reach a plant?

    Answer guidance: Consultants using AI on a task outside the frontier were 19% less likely to produce correct solutions than those without AI. It studied 758 consultants with a 2023 model, so it applies to plant work only by analogy: the same tool can help on one task and hurt on another.

Sources and further reading