Deciding well when AI gives you the answer

From Answers to Decisions: the open programme. Deciding well when AI gives you the answer.

From Answers to Decisions is a free programme in four short lessons and one practice decision, the open introduction to the AIR Academy programme of the same name. It is about deciding well when AI gives you the answer. I am Indhran Seghar, founding director of AIR APAC.

Tuesday, 9:40. "Sales are down. Send me ways to lift them by Friday."

Priya took over a regional sales team last month, at a company called Dovrane Supply. This is an invented example, like every case in this programme. At 9:40 on a Tuesday this message arrives from her director. She has a client call at ten.

VaskelAI could give her ten tactics, each with steps, owners and dates, in forty seconds.

She could paste the message into VaskelAI, the planning assistant her company uses. She has seen what comes back: ten tactics in forty seconds, each with steps, owners and dates. It looks like a week of good work.

The habit that worked before AI: take the task as given and start on it.

This is how most of us were trained to work. Someone gives you a task, and you start work at once. People who worked fast were valued for this, and that was fair.

Habits are out of date. People are not.

Priya is good at her job. This habit served her well for fifteen years. What changed is the tool on the other side of it.

Before: a wrong task cost a few hours, and someone noticed. | Now: AI starts at once on whatever it is asked.

Before AI, a task aimed at the wrong thing cost you a few hours, and usually a colleague noticed. Now AI starts at once on whatever it is asked, and nobody is checking the question.

A wrong question now gets a fast, complete, convincing answer.

Priya's ten tactics would be exactly that. Each one would be sensible. Together they would answer a question nobody has checked.

60% of executives said they regularly use AI to support decisions.

Deloitte's Human Capital Trends 2026 found that 60% of executives said they regularly use AI to support decisions. The survey asked more than 9,000 leaders in 89 countries. The number of executives among them is not published.

The new habit: find the real decision before asking for an answer.

The shift is small, and it happens before the prompt. Find the real decision first. Then ask.

Back to Priya. 9:43.

Back to Priya, who still has a call at ten.

"Sales are down." What is the real decision here?

She asks herself one question before she asks VaskelAI anything. What is the real decision here? It is not which tactic to use. First she needs to know why sales fell, and which cause her team should work on.

Fewer new enquiries. Fewer enquiries becoming orders. Lower prices. Customers leaving.

Sales fall for different reasons, and each reason is a different decision. Fewer new customers asking is one. Customers leaving is another. Ten tactics would split her team's effort across all of them.

Two large customers cut their orders after a late delivery in August.

She calls her team leader. It takes four minutes. Most of the fall is in two large customers. They cut their orders after a late delivery in August. This is a decision about keeping two customers, not about selling more.

Priya still had to ask the team who knew the two large customers.

Notice where that answer came from. VaskelAI only had the director's message. Priya had a team who knew the two large customers, and she asked them.

Before she asks, Priya names what is fixed: no new staff, and the two large customers cannot be lost.

Then she names the goal: keep the two large customers. And she names what she may not change: no new staff this quarter. A tactic that ignores either one is not an option, however good it sounds. People call this strategic thinking.

VaskelAI: options to win back the two large customers, with no new staff.

Her request to VaskelAI names the two large customers, the late delivery and the rule of no new staff. It asks for options to win them back, and for a sign that it is working. The answer is now about the right thing, and the forty seconds are well spent.

"A person can write an excellent prompt and still make a poor decision."

Pasting in the director's message would have been a clear request, aimed at the wrong decision. What changed the decision was four minutes with her team leader.

Judgment is how well a person decides once AI has given the answer.

The word we use for this skill is judgment. Judgment is how well a person decides once AI has given the answer. Lesson 1 is about deciding what the answer is for.

The Structure question: What is the real decision here, and what is fixed?

The first of the four questions is called Structure. What is the real decision here, and what is fixed? Ask it before you type anything.

Before you use AI to go faster, decide what the work has to achieve, and what you may not change.

This is the first line of the card you receive at the end of the programme. Before you use AI to go faster, decide what the work has to achieve, and what you may not change.

Asking first feels slower. The question takes about a minute.

Asking first feels slower. The question takes about a minute, and Priya's call took four minutes. If her team leader had confirmed the wording, she would have gone ahead, with a reason. A week of work on the wrong problem costs more.

The answer is not the decision.

AI gives you a fast answer. You still have to decide what the answer is for, and what you are not allowed to change.

Today: before your next AI request, write two short sentences. "This is for deciding ..." and "What I may not change is ..."

One thing to try today. Before your next request to any AI tool, write two short sentences. The first begins: "This is for deciding", and you finish it. The second begins: "What I may not change is", and you finish that. Then ask, and notice how your request changes.

Next: Lesson 2, Check the answer, not the wording.

Once the question is right, the answer comes back looking finished. In lesson 2 a recommendation looks perfect and depends on one claim nobody has checked. Check the answer, not the wording.

Lesson 2. Check the answer, not the wording.

In lesson 1, a manager named Priya found the real decision before she asked an AI tool for anything. Now the answer comes back, and it looks finished.

Thursday, 4:15. Two pages: headings, a table, a risk section, a recommendation.

Daniel is the operations manager at Kelmara Foods, another invented company. He asked TalmereAI, the office assistant his company uses, whether to move the building maintenance contract to a cheaper provider. Two pages come back.

TalmereAI: "Switch provider. Save 22% a year. The current contract ends on 30 days' notice."

The recommendation reads like a careful consultant wrote it. Switch provider. Save 22% a year. The current contract ends on 30 days' notice, so the change can start next quarter. His director's meeting is at 4:30.

The habit that worked before AI: judge the work by how complete and professional it looks.

This habit was a fair one. For years, the look of a piece of work told you something about the care that went into it.

AI output always looks complete and professional, including when it is wrong.

It tells you nothing now. AI output always looks complete and professional, including when it is wrong. A right answer and a wrong one arrive in the same format, with the same confidence.

66% relied on AI output without checking it, at least once. 42% did so sometimes or more often.

In 2025, KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries. Of the employees who use AI at work, 66% said they had relied on AI output without evaluating its accuracy at least once. 42% said they did so sometimes or more often.

You cannot check everything. Checking every line is not the way.

You cannot check two pages line by line before 4:30. Checking everything is not the new habit.

The new habit: test the one claim the decision rests on.

The new habit is narrower. Find the one claim the decision depends on, and test that.

Back to Daniel's two pages. 4:17.

Back to Daniel, with thirteen minutes.

If the contract cannot be ended in 30 days, nothing else on the page matters.

He asks which claim, if it were wrong, would change the decision. Not the 22%. If the saving were 18% he would still switch. The 30 days. If that is wrong, nothing else on the page matters.

What would prove it wrong? What else could explain it?

Then he asks what would prove it wrong, and what else could explain a 30-day clause. A signed contract with a longer notice period would prove it wrong. So would a penalty for leaving early. An old draft could also explain it.

Daniel: "Are you sure about the 30 days?" TalmereAI: "Yes. Clause 14 states 30 days' written notice."

He asks TalmereAI if it is sure. TalmereAI is sure. It quotes the clause and names the file in the shared folder.

An explanation from the AI is not a check.

An explanation from the AI is not a check. The explanation and the answer come from the same source. It can sound confident and detailed, and still repeat the same mistake.

Check outside the AI.

A check happens outside the AI. Go to the source of record: the signed contract, the live system, or the person who owns it. If the AI names a file, ask whether it is the current one.

The signed contract: one year's notice, and a penalty for early exit.

He calls the finance manager, who holds the contracts. The shared folder had the draft from two years ago. The signed version says one year's notice and a penalty for leaving early. TalmereAI read the file it was given, and it was the wrong file.

One short call changed the decision.

It took one short call, and the decision changed. He weighs the saving against the penalty and the notice, and does not switch this year. He will renegotiate at renewal and put the notice date in his calendar. If the 30 days had been right, he would have gone ahead.

Experience is what makes the check possible.

Daniel knew that contracts have notice periods, and that drafts and signed copies differ. That experience told him which claim mattered, and who to call. This programme uses your experience. It does not replace it.

199 US adults. Three "cognitive forcing" designs together reduced over-reliance on one measure. Buçinca, Malaya and Gajos, 2021.

Why do we skip the check? In an online study of 199 US adults in 2021, three "cognitive forcing" designs taken together reduced over-reliance on one measure. The designs were deciding before seeing the AI, seeing it only on request, and a short wait. Participants rated these designs least favourably.

It feels slower. The minute is the check.

It feels slower, because it is slower. The four questions take a minute, and the check takes as long as the evidence needs. That time is the work, not a delay before it.

The Test question: What would show this answer is wrong, and have I checked it?

The second of the four questions is called Test. People call this critical thinking. What would show this answer is wrong, and have I checked it? Ask it of the answers that matter to the decision.

AI shows you a pattern or an answer. Check what else could explain it, and what would prove it wrong.

This is the second line of the card. AI shows you a pattern or an answer. Check what else could explain it, and what would prove it wrong.

The answer is not the decision.

AI gives you an answer that looks finished. You still have to find the claim it depends on, and check it.

Today: take one AI answer you acted on this week. Find the one claim it depends on. Check it at the source of record, or ask the person who owns it.

One thing to try today. Take one AI answer you acted on this week. Find the one claim the decision depended on. Go to the source of record, or ask the person who owns it. If the AI named a file, ask whether it is the current one.

Next: Lesson 3, Bring in the people who must agree.

Once an answer is checked, it looks ready to send. But some of the people it touches must agree before you act, and some must know that AI drafted it. That is lesson 3. Bring in the people who must agree.

Lesson 3. Bring in the people who must agree.

Lesson two was about checking the answer. Now the work is checked and ready to send. Who else does the decision belong to?

The draft is checked and ready to send.

The draft is checked. It reads well. You are about to press send.

The habit that worked before AI: when the work is done, send it.

The habit that worked before AI: when the work is done, send it. The work took long enough that the right people saw it first.

AI makes it easy to act alone.

AI changes that. A draft can be ready before anyone else has seen it, so it is easy to act alone.

You can act beyond your authority.

You can send before the people who must agree have seen it.

You can leave out that AI drafted it.

Nobody notices these until something goes wrong. Then everyone asks who decided.

Here is an example.

Here is an example. It is made up, but you will recognise it.

A customer asks when their order will arrive.

A customer writes to Mira, an account manager: when will our order arrive? Mira asks CordenAI for a polite reply, and gives it only the order status. CordenAI drafts a warm answer with a date, the fourteenth.

Mira checks the date against the record and sends the reply.

Mira checks the date against the order record. It matches, so she sends it.

The person responsible for the date never saw it.

What Mira did not see: the record was out of date. That morning Hafiz, who runs delivery, had changed the plan, but the record still showed the old date. He never saw the reply.

Nobody who needed to check it was told that AI drafted it.

The customer believes Mira chose that date, and nobody who needed to check it was told that a tool wrote the reply. If the customer asks who promised the fourteenth, the answer is that Mira sent a date that nobody responsible for it had seen.

The draft matched the record, but the date was not Mira's to give.

The draft matched the record. But a delivery date is a commitment, and a commitment belongs to the person who has to keep it. Hafiz had not been asked.

The Engage question: Who must agree, who must know, and what must I not hand over to AI?

The third of the four questions is called Engage. Who must agree, who must know, and what must I not hand over to AI? It has three parts and takes about a minute.

Who must agree?

Who must agree? Not the people who would like to see it, but the person with authority over what the draft promises, such as a date, a price or a rule. Find the person responsible for that promise.

Sending everything to a senior person is not the answer either.

If you send every draft to a senior person, you have decided nothing. Decide the parts that are yours. Send only the part that is not yours to the person responsible, with a specific question.

Who must know?

Who must know? The people who will act on the result, such as the team that delivers and the customer who plans around the date. They do not approve it, but they act on it.

What must I not hand over to AI?

What must I not hand over to AI? Three things stay with a person: the commitment itself, the final approval, and anything your organisation does not allow in an AI tool, such as a customer's personal details. Check that rule before you open the tool.

Telling people includes saying that AI drafted it.

Telling people includes one more thing: say that AI drafted it. Tell the people who will check or act on the work, and anyone your organisation's rules say must be told. It is not an apology, but a fact the next person needs.

Two traps in AIR APAC's field guide: sending everything upward for safety, and telling people about AI only when they ask.

These are traps six and seven in AIR APAC's field guide of nine traps. One passes a decision upward too early. The other keeps back a fact the next person needs.

Who must agree? Hafiz, because he is responsible for the date.

Back to Mira, with the Engage question. She sends Hafiz the draft and asks which date he can keep. About ten minutes later, he answers: the eighteenth.

Who must know? The customer, and the delivery team.

Who must know? The customer, who now gets a date that will be kept, and Hafiz's delivery team, who learn that a promise has been made.

What stays with Mira? The message, and the decision to send it.

What stays with Mira? The message to the customer, and the decision to send it. Hafiz keeps the date. Mira reads the final text as her own words, because now it is, and she tells Hafiz that AI helped draft it.

It cost about ten minutes. Without it, the customer would have expected an order on the fourteenth that did not come.

It cost one message and about ten minutes. If Hafiz cannot answer today, Mira tells the customer she is confirming and will write tomorrow, and she does not send a date she cannot keep.

This is how you lead AI use in your own work.

Know what is yours to decide, bring in the person responsible for the rest, and tell the people who will act on it. Your experience tells you who they are.

Before you act on an AI recommendation, confirm who must agree, who must know, and what you may not hand to AI.

Before you act on an AI recommendation, confirm who must agree, who must know, and what you may not hand to AI. It takes a minute, and that minute is where the decision becomes yours.

Today: before you send one AI-drafted piece of work, write down who must agree, who must know, and what stays with you. Then say that AI drafted it.

One thing to try today. Before you send one AI-drafted piece of work, write down who must agree, who must know, and what stays with you. Send it first to the person who must agree, with one question. Tell the others, and say that AI drafted it.

Next: Make the decision and take responsibility.

Next, lesson four: Make the decision and take responsibility. The AI gives you a confident recommendation, and following it feels safe. We'll look at why, and at how to decide.

Lesson 4. Make the decision and take responsibility.

Lesson three was about bringing in the right people. Now they have seen it. Someone still has to decide.

The AI recommends option B, and it sounds very sure.

It recommends option B, in three clear paragraphs. It sounds sure.

The habit that worked before AI: wait until you are certain, or follow the recommendation.

Before AI, there were two common habits: waiting until you were certain, or following the recommendation. AIR APAC's field guide names this as trap eight.

AI always offers a confident recommendation. Following it feels safe.

AI makes the second habit easy: its recommendation always sounds confident. Following it feels safe, because the answer looks complete.

Your name is on the decision, so you will be asked why.

Your name is on the decision, not the tool's. People ask what you decided, and why.

Example: choosing a supplier.

Here is an example. The people are fictional.

Three suppliers. OrmandAI recommends the second.

Tomas needs a new supplier for his team. He gives OrmandAI the three quotes. OrmandAI recommends the second: lower cost over a three-year contract, good references, fast delivery. Tomas agrees.

"Why this one, and not the other two?"

At the review, the finance head asks why this one and not the other two. Tomas reads out the recommendation. She asks what he checked himself. He has no reason of his own.

If the only reason is the AI's, the decision is not yet yours.

This is a borrowed decision. The AI may be right, but Tomas has committed his team to a three-year contract. He cannot say what evidence the choice rests on, or what would change his mind.

The Perform question: What will I commit to, on what basis, and what would change my mind?

The last of the four questions is called Perform. What will I commit to, on what basis, and what would change my mind? Answer its three parts, then take responsibility.

What will I commit to? A decision I can explain.

First, make a decision you can explain in your own words. If the only explanation is the AI's text, you have not decided yet. Try to say: "I chose this because..."

On what basis? The evidence it rests on.

Second, name the evidence the decision rests on. For Tomas, that is the references he contacted himself and the delivery record he saw. It is evidence he checked, not evidence the tool described.

What would change my mind?

Third, say what would change your mind, before you commit. For Tomas: if delivery is late twice in the first three months, we review. You do not need certainty, only enough evidence to commit and a plan to change.

Then take responsibility for what follows.

Last, take responsibility for what follows. Whatever happens, the decision was yours. If the event you named happens, look at the decision again and say what you decided. Blaming the tool is trap nine in the field guide.

Do not judge a decision only by its result.

A good decision can have a bad result, and a poor decision can have a good result, because luck plays a part. So judge a decision by what you checked and why, and still take responsibility for the result.

If you lead a team, there is one more decision.

There is one more decision, if you lead a team.

People who lead teams shape how AI changes the team's work, within what the organisation has decided.

If you lead a team, you shape how AI changes its work, within what the organisation has decided. Decide which tasks AI helps with, which stay with people, and who checks the result. If you do not, each person decides alone.

Nearly 90% of CEOs were confident their employees trust their leadership. 46% of employees agreed.

In October 2026, BCG, a consulting firm, published a survey of 645 CEOs of large companies and, separately, 500 employees. Nearly ninety percent of the CEOs were confident that their employees trust their leadership. Forty-six percent of the employees agreed.

40% of the 500 employees said their CEO had not been transparent enough about how AI will affect their roles and careers.

In the same survey, forty percent of the 500 employees said their CEO had not been transparent enough about how AI will affect their roles and careers. A team leader can open the same gap.

Tell your team what AI will change.

Within what your organisation has decided, tell your team which tasks AI will take over and which stay with people. Where nothing is decided yet, say so. Say how you use AI yourself, then ask how they do.

Ask the team to help shape the change.

Invite people to help shape the change, because they know the daily work best.

Decide what you will commit to, on what evidence, and take responsibility for what follows.

Decide what you will commit to, on what evidence, and take responsibility for what follows. For your own decisions, and for how your team works with AI.

Today: for one decision you made on an AI recommendation, finish two sentences. "I chose this because..." "I would change my mind if..."

One thing to try today. Take a decision you made this week on an AI recommendation. Say: "I chose this because..." Then say: "I would change my mind if..." If either is hard to finish, look there.

S: Before you use AI to go faster, decide what the work has to achieve, and what you may not change.

T: AI shows you a pattern or an answer. Check what else could explain it, and what would prove it wrong.

That is the series. Here are the four sentences, two at a time. Each says what you still have to do before you act.

E: Before you act on an AI recommendation, confirm who must agree, who must know, and what you may not hand to AI.

P: Decide what you will commit to, on what evidence, and take responsibility for what follows.

Here are the other two.

Judgment is how well a person decides once AI has given the answer.

Judgment is how well a person decides once AI has given the answer. The four questions are how you practise it. Habits are out of date. People are not.

Next step: the practice decision.

Your next step is the practice decision: one case, three decision points, seven minutes. You see which evidence you looked at, with no score and no level.

The pack by email is optional. Everything stays open.

If you want the card and the five exercises by email, ask for the pack on the page. It is optional, and everything stays open. The page also says how we use your address.

This card is training material. It is not an assessment and earns nothing toward certification.

This card is training material. It is not an assessment and earns nothing toward certification. This programme gives practice, not a result. AIR APAC runs a separate assessment, described at airapac.org/standard.