Lesson objective

Write behaviour statements for a plant AI session that name the trigger, the role, the action, the source it is checked against and the person who owns the result.

Plain-language explanation

Why this matters now

Module 1 described the people in the room and why telling them a rule changes little. This module turns that into a design method. It starts at the end: what a learner must do on the next shift, at a particular bench, desk or terminal, when an AI-drafted document or value arrives in front of them.

Most plant AI sessions are designed forwards. Someone collects what they know about the tools, puts it in order, and trims it to fit the slot. The result is a session about AI. Designing backwards produces a session about a job: the quality technician's inspection record, the planner's customer reply, the maintenance lead's close-out.

Nothing in this module rests on learning-science research, because this programme holds none. Backwards design is presented here as this programme's recommended practice, and every choice below, from how a behaviour statement is written to how long a segment runs, is labelled as practice. You can defend it to a plant manager on its logic, not on a citation.

A question to sit with

Think of the last training you sat through at work. If someone had followed you onto the floor the next day, what would they have seen you do differently? If the answer is "nothing I could point to", what should the person who designed it have written down first?

What goes wrong

1. The objective is a state of mind. "Participants will understand the risks of AI" cannot be seen, checked or practised. Nobody can tell on the next shift whether it happened.

2. The objective is about the tool, not the job. "Participants will be able to use the summarise feature" teaches a button. The behaviour that matters is what the person does with the summary.

3. The trigger is missing. A behaviour without a moment attached never fires. "Check AI output" does not say when; "when an AI-drafted value is about to go onto an inspection record" does.

4. The source is missing. "Verify the dimension" does not say against what. The drawing at the revision on the traveller is a source; "your experience" is not.

5. The owner is missing. A behaviour statement that never names who owns the result lets the room assume the tool, the instructor or "quality" owns it.

The practical answer

Best practice: write the behaviour before anything else. Before a slide, a demonstration or a handout exists, write three to five behaviour statements. Use the five-part shape.

Weak objective Behaviour statement
Understand AI risks When an AI-drafted value is about to go onto an inspection record, the inspector checks it against the drawing at the revision on the traveller and initials the verification column; the person the quality system names owns the record.
Use AI to write emails When an AI-drafted customer reply names a part, revision or date, the customer service lead checks each against the order and the schedule before sending; the customer service lead owns what the customer is told.
Be aware of data rules When about to paste a print, process sheet or customer file into any tool, the user stops and confirms the tool is approved for that information and the data has been screened, asking the named export or data contact if unsure; that contact owns the determination.
Know that AI makes mistakes When an AI summary of a maintenance log is used to plan work, the maintenance lead opens the log and reads the entries the summary depends on; the maintenance lead owns the plan and any return to service.

Each right-hand statement can be observed, practised on fictional material and checked on the next shift. Each names a source and an owner.

Why the owner is part of the objective. Professional guidance (voluntary): NIST's AI Risk Management Framework, AI RMF 1.0 (NIST AI 100-1, January 2023), says in Appendix C that human roles in decision making and oversight "need to be clearly defined and differentiated", and that human-AI configurations can span "from fully autonomous to fully manual." The Framework is voluntary, binds no plant, is not something a company can be certified to, and NIST states on its web page that AI RMF 1.0 "is being revised"; cite it as 1.0 and date it. It still gives you a sound, citable reason to write the owner into every objective: a behaviour that does not say whose decision it is leaves that role undefined.

Best practice: choose behaviours where a check changes an outcome. Not every AI use deserves a session. Pick behaviours where an unchecked AI output would reach a customer, a controlled document, a shipment or a machine. A rewritten internal meeting note rarely qualifies. A value on a certificate of conformance always does.

Best practice: say where the tool's help ends. For each behaviour, write one line naming what the tool may do and what it may not. The tool may draft the nonconformance description; it may not disposition the part. The tool may summarise the maintenance log; it may not decide a machine is ready to run. That line becomes the most repeated sentence in the session.

Why the verification step belongs inside the behaviour

Research (Goddard, Roudsari and Wyatt, Journal of the American Medical Informatics Association, 2012): a systematic review of clinical decision support in healthcare, before large language models, defines automation bias as "the tendency to over-rely on automation", and lists among its mitigators "training and emphasizing user accountability". Naming the accountable person in the objective is one way to make accountability concrete. The review studied healthcare, not manufacturing; applying it here is this programme's analogy.

Professional guidance (voluntary): NIST AI 600-1 (July 2024) defines confabulation as "the production of confidently stated but erroneous or false content". That is why every behaviour statement names a source record to check against: the tool's own explanation is not a source.

Best practice: test each statement against the next shift. Read each behaviour statement to a supervisor from the floor and ask: could you see this happen on your shift, and would you know if it did not? If the supervisor cannot picture it, rewrite it.

Best practice: keep the list short. Three to five behaviours for a session is this programme's recommendation, because each one needs its own practice and its own check. A list of fifteen behaviours is a lecture in disguise.

Remember this

  • Design backwards: write the behaviour a learner must show on the next shift before any slide exists.
  • Every behaviour statement names the trigger, the role, the action, the source and the owner.
  • NIST AI RMF 1.0 is voluntary and under revision; cite it, dated, for why human roles must be clearly defined.
  • Name what the tool may do and where its help ends; AI never dispositions, releases or clears anything.
  • This method is this programme's practice, not a learning-science finding.

Manufacturing example

Hollis Valve Components, a fictional maker of bronze valve bodies, asks its quality engineer, Anneke Mbeki, to design a thirty-minute AI session for inspectors and customer service. Her first outline lists four objectives: "understand generative AI", "know the risks", "use the approved tool" and "be aware of data rules". She reads them to the fictional second-shift supervisor, Carl Iverson, and asks what he would see on his shift if they worked. He says: "Nothing. I wouldn't know." Anneke rewrites them backwards. The first becomes: when an AI-drafted value is about to go onto a first-article inspection report, the inspector checks it against the fictional drawing at the revision on the traveller and initials the new verification column; the person the quality system names owns the report. The second: when an AI-drafted customer reply states a part number, revision or ship date, customer service checks each against the order and the schedule before sending; the customer service lead owns the reply. The third: when about to paste any print or customer file into a tool, the user stops and asks the named export and data contact, the fictional operations director Lena Haug, unless the file is already on the cleared list; Lena owns the determination. Anneke drops the fourth, because Carl could not name a moment on his shift where it applied. She adds one line to each: what the tool may do and where its help ends. Carl reads the new list and says he could watch for the first two on any shift. Anneke records the three statements at the top of her session design template before building a single slide.

Responsible-use warning

Never write a session objective as a state of mind such as "understand" or "be aware of", and never write one about a tool feature instead of a job. Never leave out the trigger, the source record or the owner. Never cite NIST AI RMF 1.0 as a requirement, a certifiable standard or the current final word; it is voluntary and under revision. And never present backwards design as proven by research; it is this programme's practice.

Human-review requirement

The process owner for each behaviour approves its statement before the session is built: the quality manager or the person the quality system names for inspection records and dispositions, the customer service lead for customer communications, the maintenance lead for work planning and return to service, and the company's named export or data contact for any data-screening step. The plant manager approves the session as a whole. The instructor confirms every statement names a real owner in this plant.

Practical exercise

Write behaviour statements for a real-looking session. Use a fictional plant, fictional roles and fictional documents throughout.

  1. Choose a fictional plant and one team, such as inspection, planning, maintenance or customer service, and describe what AI-drafted material reaches that team today.
  2. Write the four objectives you would have written before this lesson, in the usual "understand" or "be aware of" style.
  3. Rewrite each as a five-part behaviour statement: when, who, does what, against what, owned by whom.
  4. Read each rewritten statement to a fictional supervisor and write down whether they could see it happen on a shift; rewrite any they could not.
  5. Drop any behaviour where an unchecked AI output would not reach a customer, a controlled document, a shipment or a machine, and note why.
  6. For each remaining behaviour, write one line saying what the tool may do and where its help ends.
  7. Check the final list has three to five behaviours, and that each names a source record, not the tool's own explanation.
  8. Write the one sentence you will say about NIST AI RMF 1.0 if a manager asks why every objective names an owner, including that it is voluntary and under revision.
  9. Send the final statements, in writing, to each fictional process owner named in them and note any owner who disagrees with the wording.

Knowledge check

  1. What are the five parts of a behaviour statement in this programme's design practice?

    Answer guidance: When (the trigger), who (the role), does what (an observable action), against what (the source record) and owned by whom (the named person); it is this programme's practice, not a requirement.

  2. Why is "participants will understand the risks of AI" a weak objective?

    Answer guidance: Because it describes a state of mind that nobody can observe, practise or check on the next shift, and it names no trigger, source record or owner.

  3. What does NIST AI RMF 1.0 Appendix C say that supports naming an owner in each objective?

    Answer guidance: That human roles in decision making and oversight need to be clearly defined and differentiated; the Framework is voluntary, not certifiable, and is being revised, so cite it as 1.0.

  4. Why does each behaviour statement name a source record rather than the tool's explanation?

    Answer guidance: Because NIST AI 600-1 describes confabulation as confidently stated but false content, so a tool's own explanation can be wrong; the drawing, order or log is what the output is checked against.

Sources and further reading