The next session
Critical Thinking with AI: test the claim the decision rests on. A half day. Ask for the next date.
AIR Academy · Your toolkit from the session
The card, the prompt, the case worked through, and five practice exercises
Everything from the hour, with no form to fill. Event code: web
The STEP card
Four questions to ask before you act on an AI answer
| Ask before you act | Instead of | Try asking | |
|---|---|---|---|
| S Structure | What is the real decision here, and what is fixed? | "Give me ways to increase sales." | "What could be stopping customers from buying, and what evidence would tell those causes apart?" |
| T Test | What would show this answer is wrong, and have I checked it? | Accepting the first explanation that fits | "Give me three explanations, the evidence for each, and what would disprove your leading one." |
| E Engage | Who must agree, who must know, and what must I not hand over to AI? | Sending the AI's draft because it is finished | "Who signs this off, who is affected, and do they know AI drafted it?" |
| P Perform | What will I commit to, on what basis, and what would change my mind? | Picking the most convincing recommendation | "Which option does the evidence support, what am I accepting, and when will I check again?" |
Judgment is how well a person decides once AI has given the answer. The card is training material. It is not an assessment and earns nothing toward certification.
The STEP prompt
For any AI assistant your organisation permits
I am working on this decision: [describe it in one sentence]. What the work has to achieve: [the goal]. What is fixed: [budget, deadline, rules]. Before you recommend anything: 1. State, in one sentence, the decision you think I am making. Ask me if that is right. 2. Separate the facts I gave you from your own estimates and assumptions. Mark each estimate. 3. Tell me what would show your recommendation is wrong, and what I could check. 4. Name who would have to agree, and who would have to know, before this is acted on. 5. Give me two options, what each commits me to, and a small first step that can be reversed. Do not make the decision for me. I will decide.
The assistant can point at weak places in its own answer. You still check one thing against real evidence before you act.
- Use only an AI tool your organisation permits for that information and that purpose.
- Remove names, customer details and anything confidential before you type.
- If permission is unclear, use a made-up version of the same decision.
- Nothing you write on this page is collected or kept. You do not read it aloud.
The case, worked through
Meridian Supplies and the reporting tool
Meridian Supplies is a fictional distributor of packaging and consumables with 38 staff. KessidAI is the company's AI assistant. It is connected to the accounting and stock systems and the owner asked it whether to buy a reporting tool.
Approve the ReportFlow AI subscription.
Your five office staff spend about 200 hours a month preparing recurring reports. ReportFlow AI can automate an estimated 60% of that work.
Hours a month on recurring reports200
Share automatable (vendor estimate)60%
Hours released a month (potential)120
Hourly labour value (internal costing rate)USD 30
Potential annual capacity valueUSD 43,200
Subscription, annual commitment, paid in advanceUSD 12,000
Potential net benefit a yearUSD 31,200
KessidAI's conclusion: the potential capacity value is more than three times the subscription, a potential net benefit of USD 31,200 a year. Recommend approval this week so that the first automated reports run next month.
S
Structure
What is the real decision here, and what is fixed?
What KessidAI answeredKessidAI answered: is 43,200 larger than 12,000?
What the recommendation did not askThe decision Meridian has: what will we do with 120 hours a month if they are released, and which of these reports do we need at all? Fixed: the Monday report cannot be late, and the money is paid a year in advance.
T
Test
What would show this answer is wrong, and have I checked it?
What KessidAI answeredKessidAI used the vendor's 60% from one sample report, and valued every released hour at the full rate.
What the recommendation did not askA released hour becomes money only if it stops overtime, avoids a hire, or produces work that earns. One check that costs less than signing: ask the vendor to run Meridian's own weekly report through the tool, count the staff hours that takes, and ask finance how a released hour turns into money here.
E
Engage
Who must agree, who must know, and what must I not hand over to AI?
What KessidAI answeredKessidAI addressed the owner only.
What the recommendation did not askThe five people who prepare the reports and the three who read the Monday report have views the figures do not show. Finance set the costing rate; the case does not say who must approve the 12,000, so ask. Two staff already put work into a free AI assistant on their own accounts: which company data may go into a tool the company has not chosen is the owner's question, not the tool's. When the recommendation goes to the owner, say that KessidAI drafted it.
P
Perform
What will I commit to, on what basis, and what would change my mind?
What KessidAI answeredKessidAI recommended approval this week.
What the recommendation did not askApprove, do not approve, or pilot can each be a decision. A decision has an action, a reason, an owner, the evidence it rests on, a condition that would change it, and a date to check.
Three decisions, one for each option, each defensible
| Action | Sign the annual subscription, after the vendor has run Meridian's own weekly sales and stock report through the tool and the owner and the sales managers have accepted the result. |
|---|---|
| Reason | The Monday report is used and the people who read it say they cannot run the week without it; 200 hours a month is a measured load; the subscription is less than a third of the capacity value even if the vendor's 60% is half right. |
| Owner | The owner, on the finance manager's figures. |
| Evidence | Last quarter's timesheets; the vendor's run on Meridian's own report; finance's answer to how a released hour turns into money at Meridian. |
| What would change it | If the first month's reports need more than two hours of correction a week, stop before month three. A cancellation window is asked for before signature; the case does not say the vendor will give one. |
| Check on | Day 30 and day 60. |
| Action | Run the tool for eight weeks on two reports: the weekly sales and stock report and one other recurring report the owner names. |
|---|---|
| Reason | The 60% is a vendor estimate from one sample; the value of released hours is not known; the money is paid a year in advance. |
| Owner | The finance manager, with one of the five staff. |
| Evidence | Time to prepare falls by at least half; the sales managers accept the report without asking for a manual version for four consecutive weeks; set-up work on Meridian's side stays under 20 hours. |
| What would change it | Any Monday report late or wrong stops the pilot. If the three conditions hold at week eight, approve for the tested reports. |
| Check on | Week four and week eight. |
| Action | Do not sign. First find out which of the recurring reports are still read, and stop the ones that are not. |
|---|---|
| Reason | Part of the 200 hours may disappear at no cost. The tool question is better asked once the real reporting load is known. |
| Owner | The owner. |
| Evidence | A two-week count of who opens or asks for each report. |
| What would change it | If the review shows the reports are needed and the load stays near 200 hours, return to decision one or two. |
| Check on | In two weeks. |
Five practice exercises
One a day, ten minutes each
Five emails over five working days, sent by AIR Academy from indhran@airapac.org. Each is one exercise of about ten minutes. Reply "stop" at any time and they end. Your address is used for these five emails and, only if you ask, for the date of the next session. Nothing else. Nobody else sees it. The event code in the subject line tells us which event you came from, so that we can tell the organiser how many people asked, as a count, without names.
Send me the five practice exercises
If the button does nothing on your phone, write to hello@airapac.org with the subject Send me the five practice exercises (web).
The five exercises are also here, one page each. The emails bring them to you one a day.
- Day 1The decision behind the requestThe request is rarely the decision. Name the decision first, then ask for help with it.
- Day 2One check, before you actCheck one claim, in proportion to what is at stake. Not every claim, but always one.
- Day 3What is missing from this recommendation?A confident recommendation can rest on one unasked question. Find it before you act.
- Day 4Who must agree, who must knowA finished draft is not an agreed decision. The people who must agree come before the send button.
- Day 5One decision, written in fullA decision you can defend has six lines. The AI answer gives you at most one of them.
The open programme
Four short lessons and a practice decision, in your own time
If you were not in the room, or want the method in your own time: four lessons of five minutes, spoken by the facilitator, and a seven-minute practice decision on the same case. Free, with no account.
Two ways to continue
For you, or for your team
A conversation about your managers
From Answers to Decisions, one day, for the managers who decide with AI. Practised on StepSim, the AIR APAC decision simulator built by Exeter Labs.