Lesson objective
Cut session content that does not serve a behaviour, while keeping the data screen, the named approver and the verification step in every plant AI session however short.
Plain-language explanation
Why this matters now
Designing backwards produces a short list of behaviours and a plan that fits the window. It also produces a painful discovery: most of what the instructor knows about AI does not serve any behaviour on the list. The history of the technology, the comparison of models, the vocabulary slide, the survey of use cases. All of it may be true and interesting. Very little of it changes what someone does at the bench next shift.
Cutting is where many instructors lose their nerve and put things back. This lesson gives a test for what goes, and names three things that never go, whatever the window: the data screen, the named approver and the verification step. Those three are this programme's recommendation for the minimum any plant AI session carries, because each one protects against a failure a single session can prevent.
A question to sit with
Your window has just shrunk from thirty minutes to ten. You can keep one fictional document, one check and three sentences. Which three sentences would still protect the plant if they were the only thing anyone remembered?
What goes wrong
1. Content is kept because it is interesting. The history slide stays because the instructor enjoys it. It serves no behaviour.
2. The three essentials are cut because they seem obvious. "Everyone knows not to paste customer prints into a chatbot." The one person who does not is the one who pastes one.
3. The data screen is taught as a slogan. "Don't put confidential stuff in AI" gives nobody a decision. It does not say what counts, which tools are approved, or whom to ask.
4. The approver is "quality" or "management". A department is not a person. The learner needs a name, or a named role they can find on the shift.
5. The verification step is taught without a source. "Double-check it" without saying against what lets the learner check the output against the tool's own explanation.
The practical answer
Best practice: apply a three-question cutting test. For every element of the plan, ask:
| Question | If no |
|---|---|
| Does it rehearse or directly support one of the behaviour statements? | Cut it, or move it to an optional handout |
| Would an unchecked mistake in this area reach a customer, a controlled document, a shipment or a machine? | Cut it from the live session |
| Can the room do something with it in the time available? | Cut it, or turn it into a reference card |
Most of the tool tour, the history and the vocabulary fail all three. Keep a short "further reading" handout if you cannot let them go.
The three things that never get cut. Whatever the window, every plant AI session carries these, in the plant's own words, with a name attached.
1. The data screen. Before anything goes into a tool, the learner asks what it is and where it would go. A print, a process sheet, a customer file, a supplier's pricing or an employee record may be customer-confidential, the plant's own trade secret, or controlled technical data.
In practice the screen is three questions a learner can ask in ten seconds: Is this approved for this tool? Could it be technical data, customer-confidential or the plant's own process knowledge? If I am not sure, who do I ask? The answer to the last question is a named person, usually whoever the company has designated for export and data decisions. Module 9 teaches the export, trade-secret and data rules in depth; this lesson only insists that the screen is never cut.
2. The named approver. For each output the session covers, the learner can name who owns it. The person the quality system names dispositions a part and signs a certificate of conformance; the maintenance lead decides a machine returns to service; the employer, through the people its safety programme names, owns lockout and guarding; the officer or owner signs an export representation.
Professional guidance (voluntary): NIST AI RMF 1.0 (January 2023), Appendix C, says human roles in decision making and oversight "need to be clearly defined and differentiated", across configurations "from fully autonomous to fully manual." The Framework is voluntary, is not certifiable, and is being revised; date the reference. AI Coalition Network recommendation: a named, qualified person reviews and owns every AI-assisted deliverable before it reaches a customer, a regulator or a filing, and the review is recorded. That is our educational standard, not a law.
3. The verification step. The learner checks the output against a source record: the drawing at the revision on the traveller, the order, the log, the certificate from the mill. Not against the tool's explanation.
Best practice: make the three essentials portable. Put them on one card the learner keeps: the three data-screen questions and the name of the person to ask; the list of outputs and their owners in this plant; and "check against the source record, never the tool's explanation". In a ten-minute window, the card and one practised check are the session.
| Essential | What the learner does | What makes it concrete |
|---|---|---|
| Data screen | Stops before pasting and asks three questions | The name of the export and data contact |
| Named approver | Knows who owns the output | A list of outputs and owners in this plant |
| Verification step | Checks against the source record | The record named for each behaviour |
Best practice: cut honestly. When you cut something that matters but will not fit, write it down as a gap and tell the plant manager. A behaviour that needs its own session should get one, not a line on a slide.
Remember this
- Cut anything that does not serve a behaviour, would not reach a customer, document, shipment or machine, or cannot be used in the time.
- Never cut the data screen, the named approver or the verification step.
- The deemed-export definition means data can be exported without leaving the country; teach the screen and the escalation, never the determination.
- Verification is against a source record; NIST AI 600-1 notes a tool can give logical-looking steps for a wrong answer.
- Record what you had to cut as a gap and tell the plant manager.
Manufacturing example
Ardent Gear Works, a fictional gear manufacturer supplying a fictional defence prime, gives its trainer, Sunniva Obi, ten minutes at each shift's start-of-week meeting for AI. Her content from a previous ninety-minute session includes a history of AI, a comparison of three tools, a vocabulary slide, six use cases and a practice check. She applies the three-question cutting test. The history, the comparison and the vocabulary fail all three questions and go to an optional handout. Five of the six use cases fail the second question, because an unchecked mistake would not reach a customer, document, shipment or machine; she keeps the sixth, an AI-drafted inspection summary. She then writes the card. The data screen: is this file approved for this tool; could it be technical data, customer-confidential or our own process knowledge; if not sure, ask the fictional export and data contact, Marek Oyelowo, before pasting anything. The named approvers: the quality manager for dispositions and certificates of conformance, the maintenance lead for return to service, the EHS lead for lockout and guarding, and the fictional president for any export representation. The verification step: check against the drawing at the revision on the traveller, never against the tool's explanation. The ten minutes become two minutes on the card, six minutes for each person to check a fictional AI inspection summary containing one value from a superseded revision, and two minutes of debrief. She records one gap for the plant manager: the six use cases, cut from the live sessions, are now in the optional handout, and she asks whether any of them needs its own session.
Responsible-use warning
Never cut the data screen, the named approver or the verification step to save time. Never teach the data screen as a slogan without the approved tools and a named person to ask. Never tell learners whether a particular file is an export or needs a licence; that determination belongs to the company's named export authority. Never name a department where a person is needed. And never let a tool's explanation count as verification.
Human-review requirement
The company's designated export and data contact confirms the data-screen questions and is the named person learners are told to ask; the officer or owner owns any export representation. The quality manager, maintenance lead and EHS lead each confirm the outputs listed against their names. The plant manager receives the record of what was cut and decides whether any gap needs its own session. The instructor updates the card whenever a named person changes.
Practical exercise
Cut a session to a short window without losing the essentials. Use a fictional plant and fictional documents throughout.
- Take a fictional sixty-minute plant AI session outline of at least ten elements, including a history slide, a tool comparison and several use cases.
- Apply the three-question cutting test to every element and record the answer to each question.
- Move everything that fails to an optional handout or a gap list.
- Write the data-screen questions in your plant's own words and name the fictional person learners must ask, without telling them whether any file is an export.
- List each output your session covers and the named person or role who owns it in your fictional plant.
- Name the source record for each verification step and state that the tool's explanation is not one.
- Design the single card that carries all three essentials.
- Rebuild the session for a ten-minute window, with a minute budget that adds up to exactly ten and includes one practised check.
- Write the note to the plant manager recording what you cut and which gap, if any, needs its own session.
- Ask the fictional export and data contact and each named owner to confirm the card before first use, and record the date each confirmed it.
Knowledge check
-
What are the three things this programme recommends no plant AI session ever cuts?
Answer guidance: The data screen before anything goes into a tool, the named approver who owns each output, and the verification step against a source record rather than the tool's explanation.
-
What does 15 CFR 734.13(a)(2) add to the ordinary idea of an export?
Answer guidance: It treats releasing technology or source code, but not object code, to a foreign person inside the United States as a deemed export; whether a licence is needed depends on classification and destination.
-
Why is a tool's explanation of its answer not a verification step?
Answer guidance: Because NIST AI 600-1 notes that outputs can include confabulated logic and that models sometimes give logical steps even when the answer is wrong; verification needs the source record.
-
What should an instructor do with content that matters but will not fit?
Answer guidance: Record it as a gap and tell the plant manager in writing, so a behaviour that needs its own session gets one, rather than squeezing it onto a slide or quietly dropping it.
Sources and further reading
- Bureau of Industry and Security (U.S. Department of Commerce) — 15 CFR 734.13 - Export. Accessed 2026-09-15.
- 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.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.