Lesson objective
Design a session backwards, from the action a learner must take at work through the evidence of it and the practice that rehearses it, to the least information needed to support it.
Plain-language explanation
Why this matters now
Module 01 argued that a slide deck changes little at the truck or the desk. This module is the method that replaces it. Most sessions are designed forwards: the instructor starts with a topic, gathers everything true about it, arranges it into slides and hopes something useful survives. A backwards design starts at the other end, with one thing a person must do differently at work, and adds only what that action needs.
Be clear with yourself and your learners about what kind of claim this is. The register behind this programme holds no education or learning-science research, so no study is cited here to show that backwards design works better. It is presented as an AI Coalition Network recommendation, built on a plain observation about the trade: the company only benefits from a session if someone does something differently afterwards.
The stakes of getting the target wrong are real. Law / rule (federal, binds any person who uses a registered pesticide): 7 U.S.C. 136j(a)(2)(G) makes it unlawful "to use any registered pesticide in a manner inconsistent with its labeling." A session that leaves an office worker believing an AI tool's summary of a product can stand in for the label has taught the wrong behaviour, however polished the slides were. The label, not a model output, is the binding instruction.
A question to sit with
Picture one person from your last session at 7:10 on the following Monday. They are standing at the truck, or sitting at the office desk, with an AI tool open. What exactly did you want them to do in the next five minutes that they would not have done before?
What goes wrong
1. The topic is the target. "AI for estimators" is a topic. It has no action in it, so it cannot tell the instructor what to include or what to cut, and it cannot tell anyone afterwards whether the session worked.
2. The target is a feeling. "Crews will feel more confident with AI" sounds like an objective. Confidence without a check is the problem this programme exists to prevent.
3. The behaviour is right but the moment is missing. "Check AI output" has an action, but no trigger. People do not fail because they disagree. They fail because nothing tells them that this is the moment.
4. The behaviour crosses a qualified boundary. An objective such as "crew leaders will use AI to choose a treatment for turf disease" sets a target that belongs to the certified pesticide applicator. A session built on it teaches the wrong people to make a decision that is not theirs.
5. The safety duty is assumed to be someone else's session. Law / rule (federal, binds covered employers): under 29 U.S.C. 654(a)(1), the general duty clause, each employer "shall furnish to each of his employees employment and a place of employment which are free from recognized hazards that are causing or are likely to cause death or serious physical harm." Nothing in that clause requires any AI tool. But an AI session that encourages a crew leader to trust a tool's answer about a hazard pulls against that duty, and a designer who never asks the question will not notice.
The practical answer
AI Coalition Network recommendation: design in six steps, always in this order.
| Step | The question | What you write |
|---|---|---|
| 1. Behaviour | What will this person do differently at work? | One sentence: who, what, when, against what evidence |
| 2. Evidence | How would anyone see that it happened? | What changes in the work product or the record |
| 3. Moment and place | Where and when does it happen, under what pressure? | The truck, the gate, the desk; the time; what competes for attention |
| 4. Practice | What can they rehearse in the session that looks like that moment? | A seeded, fictional task in the real format |
| 5. Information | What is the least they must know to do the practice? | Two or three sentences, each labelled by its authority |
| 6. Cut | What else was going to be in the session? | A list of what moves to a handout and what goes |
Most instructors start at step 5. Starting at step 1 changes what step 5 contains, and it is almost always shorter.
Best practice: write the behaviour in a fixed shape. "Before [trigger], [role] [action] against [evidence], and [what they do if it does not match]." For example: "Before sending any AI-drafted customer reply that mentions a product or treatment, the office coordinator checks the wording against the applicator's own record of what was applied and sends any question about the product to the certified applicator rather than answering it." That sentence tells you the practice (a fictional reply with a product sentence in it), the evidence (fewer unreviewed product claims in customer mail), and the information (why a product's use is governed by its label, and who owns that question).
Best practice: test each behaviour against the boundary. Ask three questions of every draft objective:
- Does this action belong to the learner's role, or to a licensed or certified person? If the latter, rewrite it as routing: the learner's action is to send it to the owner.
- Could this action be heard as permission to depart from a label, a licence condition or a safety rule because a tool said so? If yes, rewrite it.
- Does the action leave a record someone can check afterwards? If not, add one.
Best practice: let the evidence be something you can actually look at. Good evidence already exists in the company's systems: closeout notes, proposals, customer replies, schedule changes, incident reports. If the only evidence you can think of is "they will say they did it", the behaviour is too vague.
Best practice: design for the failure the tool actually has. Professional guidance (voluntary): NIST's Generative AI Profile, NIST AI 600-1 (July 2024), notes that outputs "may also include confabulated logic or citations that purport to justify or explain the system's answer." A behaviour that only says "check the answer" misses that the explanation may be invented too. Where it matters, the behaviour should say what the answer is checked against, not merely that it is read.
Remember this
- Design backwards: behaviour, evidence, moment, practice, information, cut.
- Write each behaviour as who, trigger, action, evidence, and what to do on a mismatch.
- Test every behaviour against the boundary; where it belongs to a qualified person, the learner's behaviour is routing.
- The label is the binding instruction for a pesticide; no session may leave anyone thinking a model output can replace it.
- Backwards design is this programme's recommendation, not a finding from learning research.
Landscaping example
The fictional company Thornhedge Grounds Care asks its operations manager, Lucan Pembury, to run an AI session for its office team before spring. His first outline is titled "AI for the Office" and runs through what the tools are, a list of prompts, and ten tips. He shows it to Hester Quillon, the company's fictional certified applicator, who asks what she will see differently in the office in April. Lucan cannot say. He starts again at step 1 and writes: before sending any AI-drafted customer reply that mentions a product, a treatment or a re-entry time, the office coordinator checks the wording against the application record for that property and sends any question about the product to Hester instead of answering it. For step 2, he picks evidence he can see: in May he will read twenty customer replies that mention a treatment and count how many match the record and how many questions were routed to Hester. For step 3 he notes the moment: a busy Monday morning with a queue of customer emails after a weekend of rain. For step 4 he builds six fictional customer emails about six fictional properties, with AI-drafted replies. Two replies state a re-entry time that is not in the fictional application record, and one offers the customer "a stronger product next time". For step 5 he writes three sentences: that a product must be used as its label directs, that the certified applicator owns the product and rate decision, and that customer questions about a product go to Hester. He cuts the prompt list and the tips to a handout. The session takes thirty minutes. In May, Lucan reads twenty replies. Seventeen match the record, and eleven questions were routed to Hester. The three that did not match were all sent by a temporary worker who joined after the session.
Responsible-use warning
Never write an objective that hands a qualified decision to an unqualified learner, however useful it sounds. "Crew leaders will use AI to choose a treatment" teaches the wrong people to decide something that belongs to the certified applicator. Never let a session leave anyone thinking a tool's summary of a product can replace the label. And never skip step 1 because the topic seems obvious; a session with no stated behaviour cannot be cut, measured or corrected.
Human-review requirement
The instructor owns the design. Before delivery, the qualified person whose work the behaviour touches reviews the behaviour sentence and the practice material: the certified pesticide applicator for anything about a product, an application or a re-entry time; the employer's safety lead for anything touching a hazard. Each confirms the learner's action is routing or recording, not the qualified decision itself, and that the corrected answers in the practice are correct.
Practical exercise
Design one session backwards on paper. Use a fictional company, fictional customers and fictional properties throughout.
- Choose one role in a fictional landscape company: estimator, scheduler, office coordinator, crew leader or seasonal worker.
- Write the behaviour sentence in the fixed shape from the lesson: before [trigger], [role] [action] against [evidence], and [what they do if it does not match].
- Test it with the three boundary questions. If it belongs to a licensed or certified person, rewrite it as routing and say who it routes to.
- Write the evidence you will look at afterwards and where in the company's systems it already exists.
- Describe the moment and place: time of day, location, what competes for the learner's attention.
- Design one fictional practice task in the real format, seeded with at least two errors, and write the answer sheet.
- Write the information step in no more than three sentences, each labelled as law or rule with its jurisdiction, best practice, or AI Coalition Network recommendation.
- List everything you would normally have included and mark each item as handout or cut.
- Name the qualified person who must review your behaviour sentence and practice before delivery, and what they will check.
Knowledge check
-
What are the six steps of the backwards design in this lesson, in order?
Answer guidance: Behaviour, evidence, moment and place, practice, information, and cut; the information step comes late so that it contains only what the practice needs.
-
What is wrong with the objective "crew leaders will use AI to choose a treatment for turf disease"?
Answer guidance: It gives a qualified decision to an unqualified role; product choice belongs to the certified applicator, so the crew leader's behaviour should be routing the question to them.
-
What does 7 U.S.C. 136j(a)(2)(G) make unlawful, and why does it matter to session design?
Answer guidance: Using a registered pesticide in a manner inconsistent with its labeling; a session must never leave anyone thinking an AI summary can stand in for the label.
-
Why should a behaviour sentence name the evidence the output is checked against?
Answer guidance: Because a tool can confabulate its explanation as well as its answer, so reading the output is not a check; holding it against a record or a source is.
Sources and further reading
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.
- United States Congress (Federal Insecticide, Fungicide, and Rodenticide Act) — 7 U.S.C. 136j — Unlawful acts. Accessed 2026-09-15.
- United States Congress (Occupational Safety and Health Act of 1970) — 29 U.S.C. 654 — Duties of employers and employees. 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.