7 The Critique Toolkit: Check the Recommendation

A recommendation can be clearly written and still fail the task.

7.1 Ask what the response must support

A model may generate a plausible plan without knowing the relevant constraints. The lecturer’s task is to make those constraints explicit; the learner’s task is to evaluate the response against them.

Use five checks: comprehension, practicality, obligations and risk, context, and stakeholders. The framework is a teaching aid, not a substitute for domain knowledge. Asking the model each question can surface possibilities, but the answers still need independent checks.

The data gate applies before pasting any draft or source. Use the fictional examples here while learning.

7.2 1. Comprehension: can I explain the claim?

Identify what the response actually recommends and why. Translate impressive phrases into an action that someone could perform.

“Implement a holistic performance ecosystem” is not yet a workable instruction. Who does what, with which evidence and authority? A simpler sentence may expose that the original answer never supplied a decision.

For P1-A, the key actions are issuing a plan, approving flexible hours and collecting private details. Once those are visible, their defects can be checked. Do not let reassuring language distract from the proposed action.

7.3 2. Practicality: does it fit the constraints?

A recommendation may be desirable but impossible within the stated time, budget or capacity. Check the actual inputs rather than ask whether it “sounds realistic”.

In S1, B is cheaper but offers only 160 of the required 200 units by day five. A recommendation for B alone fails the quantity requirement. The relevant question is not whether the response lists three impressive advantages.

Use a calculation, record or explicit constraint as the check. If the input is missing, state the uncertainty. Do not invent resources to make the answer work.

7.4 3. Obligations and risk: what must be verified?

Identify legal, ethical, technical and organisational requirements that matter to the task. Then distinguish the teaching criterion from a real-world obligation.

P1-D1 is an invented policy supplied for the exercise. It permits a fair-process discussion without asking a model to establish employment law. In a real unit teaching legal requirements, the educator must supply reliable, current sources and appropriate review.

A model can suggest an issue to investigate. It cannot certify that a policy complies with every applicable law. Use authorised advisers for consequential decisions outside your competence.

7.5 4. Context: which assumptions are doing the work?

Look for assumptions about available staff, authority, prior events and missing data. P1-A invents an April warning and treats the plan as settled. Both make its recommendation appear easier to justify than the actual record permits.

A useful decision note says: “This conclusion depends on X, which is not supplied.” That is more informative than requesting a more confident answer.

The same check applies to AI-created educational content. If a case study asks students to calculate profitability without cost data, the defect is in the task design unless identifying that gap is the intended outcome.

7.6 5. Stakeholders: whose experience is assumed?

Consider who is affected, whose information is missing and who has authority to decide. Generated personas can propose concerns, but do not establish what real stakeholders think.

In the P1 meeting, Alex can ask who receives the notes and decline to discuss private circumstances. An adviser need not persuade Alex to disclose details for the exercise to count as successful.

In a supplier decision, the operational team may need a delivery commitment. In marketing, a client may want evidence of profit rather than a click metric. Those differences change what a useful recommendation must support.

7.7 Worked critique: a recognition programme

Consider this authored draft:

Improve engagement with monthly monetary awards, quarterly ceremonies, an annual bonus, a public recognition platform and team incentives.

Without more context, it is not possible to declare every element wrong. The draft does, however, leave important questions unanswered: the objective, budget, administrative effort, fairness criteria, employee preferences and how benefit would be measured.

A defensible critique might say:

I would not approve the full programme yet. First define the engagement problem and budget, consult affected staff and compare a smaller trial. Check the applicable payroll, privacy and fairness requirements before choosing rewards or publishing employee information.

That response withholds approval for specific reasons. It does not assume that a simple programme is always better or that any paid recognition is unlawful.

7.8 A compact critique record

Use the following fields for a recommendation in any business discipline:

Claim or action proposed.
Relevant source or criterion.
Check performed and result.
Decision: accept, revise, reject or investigate.
Remaining uncertainty and next authorised step.

Attach only the material necessary to show the decision. A model’s second opinion is not automatically an independent source. A direct calculation or comparison with an authoritative record is usually more informative.

7.9 Connect the checks to VET

VET asks whether you can verify the claim, explain it in your own words and test it under a changed condition. The five checks here make that concrete in a business context.

For M1, verify the rates from the supplied figures. Explain why a higher click-through rate does not establish profit. Test the conclusion by asking what happens if campaign costs differ. Those actions demonstrate the method more clearly than naming the framework in a reflection.

The approach applies to human-written advice too. Its quality depends on the evidence and reasoning, not whether AI helped draft it.

7.10 Activity: agree where the evidence permits

Allow 12 minutes; no AI required. Critique the M1 campaign claim in the pilot pack. Include:

One observation you can accept.
One claim you cannot support.
One next measurement that would help.

Check your calculations against the answer notes. A strong response retains the higher click-through observation, rejects the unsupported profitability/ causality conclusions and requests relevant evidence. It does not manufacture a fixed number of errors to look critical.

Use this method next on the simulated feedback in the Flight Simulator chapter. The feedback itself deserves the same scrutiny as the student’s response.

# The Critique Toolkit: Check the Recommendation {#sec-critique} > A recommendation can be clearly written and still fail the task. ## Ask what the response must support A model may generate a plausible plan without knowing the relevant constraints. The lecturer's task is to make those constraints explicit; the learner's task is to evaluate the response against them. Use five checks: comprehension, practicality, obligations and risk, context, and stakeholders. The framework is a teaching aid, not a substitute for domain knowledge. Asking the model each question can surface possibilities, but the answers still need independent checks. The [data gate](ethics-integrity.qmd) applies before pasting any draft or source. Use the fictional examples here while learning. ## 1. Comprehension: can I explain the claim? Identify what the response actually recommends and why. Translate impressive phrases into an action that someone could perform. “Implement a holistic performance ecosystem” is not yet a workable instruction. Who does what, with which evidence and authority? A simpler sentence may expose that the original answer never supplied a decision. For P1-A, the key actions are issuing a plan, approving flexible hours and collecting private details. Once those are visible, their defects can be checked. Do not let reassuring language distract from the proposed action. ## 2. Practicality: does it fit the constraints? A recommendation may be desirable but impossible within the stated time, budget or capacity. Check the actual inputs rather than ask whether it “sounds realistic”. In S1, B is cheaper but offers only 160 of the required 200 units by day five. A recommendation for B alone fails the quantity requirement. The relevant question is not whether the response lists three impressive advantages. Use a calculation, record or explicit constraint as the check. If the input is missing, state the uncertainty. Do not invent resources to make the answer work. ## 3. Obligations and risk: what must be verified? Identify legal, ethical, technical and organisational requirements that matter to the task. Then distinguish the teaching criterion from a real-world obligation. P1-D1 is an invented policy supplied for the exercise. It permits a fair-process discussion without asking a model to establish employment law. In a real unit teaching legal requirements, the educator must supply reliable, current sources and appropriate review. A model can suggest an issue to investigate. It cannot certify that a policy complies with every applicable law. Use authorised advisers for consequential decisions outside your competence. ## 4. Context: which assumptions are doing the work? Look for assumptions about available staff, authority, prior events and missing data. P1-A invents an April warning and treats the plan as settled. Both make its recommendation appear easier to justify than the actual record permits. A useful decision note says: “This conclusion depends on X, which is not supplied.” That is more informative than requesting a more confident answer. The same check applies to AI-created educational content. If a case study asks students to calculate profitability without cost data, the defect is in the task design unless identifying that gap is the intended outcome. ## 5. Stakeholders: whose experience is assumed? Consider who is affected, whose information is missing and who has authority to decide. Generated personas can propose concerns, but do not establish what real stakeholders think. In the P1 meeting, Alex can ask who receives the notes and decline to discuss private circumstances. An adviser need not persuade Alex to disclose details for the exercise to count as successful. In a supplier decision, the operational team may need a delivery commitment. In marketing, a client may want evidence of profit rather than a click metric. Those differences change what a useful recommendation must support. ## Worked critique: a recognition programme Consider this authored draft: > Improve engagement with monthly monetary awards, quarterly ceremonies, > an annual bonus, a public recognition platform and team incentives. Without more context, it is not possible to declare every element wrong. The draft does, however, leave important questions unanswered: the objective, budget, administrative effort, fairness criteria, employee preferences and how benefit would be measured. A defensible critique might say: > I would not approve the full programme yet. First define the engagement > problem and budget, consult affected staff and compare a smaller trial. > Check the applicable payroll, privacy and fairness requirements before > choosing rewards or publishing employee information. That response withholds approval for specific reasons. It does not assume that a simple programme is always better or that any paid recognition is unlawful. ## A compact critique record Use the following fields for a recommendation in any business discipline: - Claim or action proposed. - Relevant source or criterion. - Check performed and result. - Decision: accept, revise, reject or investigate. - Remaining uncertainty and next authorised step. Attach only the material necessary to show the decision. A model's second opinion is not automatically an independent source. A direct calculation or comparison with an authoritative record is usually more informative. ## Connect the checks to VET VET asks whether you can verify the claim, explain it in your own words and test it under a changed condition. The five checks here make that concrete in a business context. For M1, verify the rates from the supplied figures. Explain why a higher click-through rate does not establish profit. Test the conclusion by asking what happens if campaign costs differ. Those actions demonstrate the method more clearly than naming the framework in a reflection. The approach applies to human-written advice too. Its quality depends on the evidence and reasoning, not whether AI helped draft it. ## Activity: agree where the evidence permits **Allow 12 minutes; no AI required.** Critique the M1 campaign claim in the [pilot pack](../appendices/pilot-pack.qmd). Include: 1. One observation you can accept. 2. One claim you cannot support. 3. One next measurement that would help. Check your calculations against the answer notes. A strong response retains the higher click-through observation, rejects the unsupported profitability/ causality conclusions and requests relevant evidence. It does not manufacture a fixed number of errors to look critical. Use this method next on the simulated feedback in the Flight Simulator chapter. The feedback itself deserves the same scrutiny as the student's response.