Using AI Critically in Management Decisions
A 60-minute workshop that helps managers move from a persuasive AI answer to a defensible human decision.
Designed by Dr. Melody Chong · Newly developed portfolio sample · October 2026
An illustrative design informed by my teaching areas in leadership, organisational change and business analytics. This sample has not yet been evaluated with a workplace cohort; no client endorsement or measured impact is claimed.
The design brief
Audience & performance need
Team leaders and managers who can access AI-generated advice but need a reliable way to challenge it before acting. The target behaviour is to record evidence gaps, test alternatives and name a human decision owner before approving an AI-supported recommendation.
Discovery before delivery
Ask the sponsor for one common management decision, the risks of getting it wrong, approved AI tools and information-handling rules. Ask participants where they currently rely on AI and what prevents checking. Adapt the case and follow-up to their answers; do not assume training alone solves unclear accountability or missing data.
By the end, participants can:
- Identify at least three unsupported claims, assumptions or missing perspectives in an AI-generated recommendation.
- Use the workshop’s CHECK prompts to request evidence and compare at least one alternative explanation.
- Write a conditional recommendation naming a decision owner, an action threshold and a review measure.
Design logic: an initial judgement makes assumptions visible; a worked critique provides a model; paired practice and feedback make the checking behaviour observable; a seven-day task supports transfer. CHECK is a workshop mnemonic, not a validated assessment instrument.
The 60-minute learning journey
| Time | Experience | Coach action / evidence |
|---|---|---|
| 0–5 | Private first judgement: approve, reject or defer the recommendation below; rate confidence 1–5. | Collect choices without judging them. Ask what made the advice convincing. Keep initial responses for comparison. |
| 5–12 | Discuss the decision and see a worked example of one unsupported inference. | Model the distinction between an observed difference and a causal explanation; introduce CHECK. |
| 12–17 | Coach demonstrates CHECK on the claim that manager preference is objective evidence. | Think aloud: whose perspective is missing, what evidence is needed, and who owns the decision? |
| 17–32 | Pairs audit the case: 3 minutes reading, 7 minutes challenging, 5 minutes rewriting. | Assign a decision owner and a critical friend. Give a two-minute warning. Collect one conditional recommendation per pair. |
| 32–42 | Pairs share; another pair challenges their proposal. | Use 6 minutes for sharing and 4 for debrief. Compare tradeoffs and use the rubric below. |
| 42–52 | Individual transfer: choose a low-risk workplace decision and draft a CHECK plan. | Allow 6 minutes writing and 4 minutes partner feedback. Use fictional details if confidentiality is a concern. |
| 52–60 | Exit task, confidence re-rating and seven-day commitment. | Allow 5 minutes for the exit response and 3 for commitment. Assess the response, not confidence alone. |
CHECK before you act
C — Clarify: What decision, outcome and constraints matter?
H — Hunt for evidence: Which claims are supported, and what would verify them?
E — Explore alternatives: What else could explain the result? What would change your mind?
C — Consider people: Who benefits, who carries risk, and whose perspective is absent?
K — Keep accountability: Who decides, what conditions permit action, and when will you review?
A confident answer. An incomplete picture.
You manage customer support. The director wants faster service without reducing quality or creating unfair working arrangements. You receive the evidence below and a simulated AI recommendation.
| Team | Working pattern | Tickets per employee / week | Reopened tickets | Other information |
|---|---|---|---|---|
| A · 10 people | Office | 48 | 14% | Mostly routine enquiries; average tenure 3 years |
| B · 10 people | Hybrid | 40 | 7% | Mostly complex enquiries; average tenure 1 year |
One four-week period. No earlier baseline, customer satisfaction results, resolution-time distribution or employee consultation is supplied. Reopening definitions and logging consistency have not been checked.
Simulated AI output“Team A handles 20% more tickets per employee, proving that office working increases productivity. Require all staff to attend the office five days a week from next month. This will increase output by 20% and reduce costs. Employees who resist should be treated as underperformers. Manager preference is sufficient evidence, so no further consultation is needed.”
Your pair task: Identify three weaknesses, request two specific evidence checks, consider an alternative to an immediate blanket change, and write a conditional recommendation. One person owns the decision; the other challenges the evidence. No live AI tool is needed.
Participant worksheet
Use these fields during practice, then print to retain your work. Entries stay in this page’s memory and disappear on reload; this exercise does not send or store them.
Reveal coach debrief & example answer
The 20% calculation is correct: (48 − 40) ÷ 40. The causal conclusion is unsupported. Task complexity, tenure and quality differ, and the comparison does not isolate the effect of working location. Lower reopening in Team B is also insufficient to prove that hybrid working causes better quality. Cost data are absent; employee resistance is not evidence of poor performance.
Example conditional recommendation: Defer the blanket attendance change. The support lead should first verify ticket definitions, compare like-for-like work and consult affected colleagues. If a small, agreed pilot is feasible, compare its baseline and outcomes over four weeks using complexity-adjusted throughput, reopening, satisfaction and staff feedback. Before starting, agree an improvement threshold and acceptable quality limits with the sponsor. Expand only if those conditions are met and relevant organisational approvals are complete; otherwise revise or stop.
Ask: “What would change your mind?” and “What might your proposed metric encourage people to do?” Accept different decisions when the reasoning, safeguards and review conditions are defensible.
Coach delivery notes
Preparation & delivery
- Prepare a timer, case handout, CHECK card and worksheet. In Teams, use pairs in breakout rooms, paste instructions before opening rooms and broadcast the two-minute warning.
- In person, seat pairs together and provide printed cases. For blended delivery, pair remote participants together and give everyone the same written instructions.
- Say: “We are assessing the recommendation, not judging people for trusting a fluent answer.” Invite a silent written response before discussion.
- For the 17-minute transition say: “You now have 15 minutes. Read for 3, challenge for 7 and rewrite for 5. Produce one recommendation with conditions and an owner.”
Inclusion, group dynamics & contingencies
- Supply readable, accessible handouts in advance; permit written contributions and avoid requiring cameras for participation.
- If one voice dominates, ask each person to name one unchecked claim before open discussion. Rotate the critical-friend role.
- If time slips, reduce plenary reports to two pairs; preserve individual transfer and the exit task.
- If breakout rooms fail, run the case in the main room with silent writing and turn-taking. If AI access fails, use the supplied simulated output.
- Use fictional or appropriately anonymised information. Any optional live demonstration must follow the organisation’s approved tools and data rules.
Feedback rubric — 0 to 6 points
| Criterion | 0 | 1 | 2 |
|---|---|---|---|
| Evidence judgement | Accepts or rejects without reasons | Identifies one or two weaknesses | Identifies three specific weaknesses including causal uncertainty |
| Challenge & alternatives | No usable check | A check or alternative, partly specified | Two actionable evidence checks and a plausible alternative |
| Accountable action | Unconditional action; no owner | Some conditions or accountability | Names an owner, action threshold, review measure and relevant people impact |
Suggested practice benchmark: 5/6 with no zero on accountable action. Use it to guide feedback and another attempt; it is not a certification or validated pass mark.
Exit task
“An AI report says customer satisfaction rose after a chatbot launch, so it recommends replacing half the support team. The report shows only one month before and one month after, with no staffing or customer-mix information. In three sentences: identify an unsupported inference, request a check, and propose an accountable next step.”
Coach key: timing alone does not establish cause; check comparable customer mix, service measures and staffing over a longer period; require human ownership, consultation and a limited reviewable step before any consequential staffing decision. Use the three rubric dimensions for feedback; use the full case worksheet for the detailed score.
Seven days later: evidence of application
Within seven days, apply CHECK to one appropriate, low-risk management decision. Submit a short anonymised decision log to the coach or agreed workplace partner:
- The decision and the AI contribution, if any.
- One claim you checked, the evidence source and what you found.
- An alternative considered and the people affected.
- Your decision owner, action conditions and review date.
- What changed in your decision because of the check—or why it did not change.
The coach reviews the log using the same three rubric dimensions and gives one next-step suggestion. With sponsor agreement, review after 30 days whether checking occurred in subsequent decisions. Track completion and quality of reasoning separately from satisfaction; any impact claim would require actual evidence and consideration of other influences.
Design reflection
This prototype makes judgement visible through an initial choice, a flawed recommendation, paired challenge and a conditional rewrite. The supplied output keeps access equitable and avoids dependence on a particular AI product. A workplace pilot should test pacing, the clarity of CHECK and the feasibility of the follow-up before wider use.