People
Who owns this decision, who is affected by it and who has to live with the result? If the answer is a team rather than a person, who speaks for the team when the decision is questioned?
Learn
Tools for better decisions with AI
I use these questions when an AI system has produced a confident answer and someone still has to decide what to do with it. They are for a working meeting, not for a policy document. You can copy them into your own notes and change the wording to suit your team.
People, process and data
Three questions that work together
I treat people, process and data as complementary questions. They are not stages in a method, and you do not finish one before starting the next. A clear answer on one can hide a gap in another, which is why I ask all three about the same decision.
Process
How does the decision move from a recommendation to an approval? Who checks the work before it is acted on, and what would make someone stop it?
Data
What evidence does the recommendation rest on, what is missing and what has been assumed? A number can be accurate and still not answer the question you asked.
For further reading, AIR APAC, the institution I founded, publishes articles and a field guide for managers. Use them alongside the prompts below.
Worksheet
Prompts you can copy
Each block below can be selected and pasted into a document or a chat. Answer the prompts about one real decision rather than a hypothetical one.
Decision and owner
What is the decision, in one sentence? Who owns it? Who is affected, and who has to live with the result? Who can stop it?
Recommendation and alternatives
What is the recommendation, in one sentence? What else was considered? Why was each alternative rejected? What would change the recommendation?
Source and assumptions
Where did the figures come from? Which assumptions does the conclusion depend on? What is missing or unverified? How confident are we, and why?
Workflow and approval
Who prepares the recommendation? Who checks it? Who approves it? What is recorded, and who is accountable if it is wrong?
Measure and revisit
What will we measure? When will we review it? What result would tell us to reverse the decision? Who reviews, and by when?
Workshop exercise
A fictional example: replacing a supplier
This exercise is fictional and the figures are illustrative. They are not client results. Try it with a colleague before comparing your answers.
An AI system recommends replacing a supplier to save US$12,000 a year. The transition is expected to cost US$8,000. Review and support are expected to cost US$6,000 a year.
Before
Replace the supplier. The AI estimates a saving of US$12,000 a year.
After: a conditional decision
Consider the replacement only after checking service quality and continuity, and if the transition cost stays at US$8,000, review and support stay at US$6,000 a year, and we accept that the first year costs US$2,000 more than we spend now. Revisit at twelve months, when the saving would be US$6,000 a year if the assumptions hold.
In the first year the cost is US$14,000 against a saving of US$12,000, so the first-year cost rises by US$2,000. In later years the saving would be US$6,000 a year if the assumptions hold. The point of the exercise is not the arithmetic. It is the conditions attached to the approval.
Ask the room
- What do you check before agreeing to the conditional decision?
- Who approves the extra US$2,000 in the first year?
- Who owns the review at twelve months?
- Which assumption would have to fail before you reversed the decision?
For boards
Four questions for a board
- Which decisions are we willing to delegate to AI, and which stay with us?
- Who is accountable when an AI-supported decision turns out to be wrong?
- What do we require before a recommendation reaches this board: the source, the assumptions and the alternatives?
- What will we measure, when will we review it, and what would make us change course?
I wrote about these questions in what a board should ask and in when it is right to automate a decision.
Process and data tools
Follow one decision through the workflow.
Use these alongside the worksheet. Bring the people who prepare, check and approve the work into the same conversation.
A workflow canvas
Request and intended result: Inputs and their sources: Steps, handoffs and waiting time: Where AI contributes: Human check and reason to stop: Approval owner: Result to measure and review date:
Mark the bottleneck. Ask whether a faster AI step changes the whole workflow or simply moves the queue.
Seven questions for the data
- What is the original source?
- How current is it?
- What does each measure mean?
- Which records, groups or periods are missing?
- What assumptions or estimates were added?
- Can another person reproduce the key calculation?
- What evidence would change the decision?
See these questions in context in the four applied work examples.
Continue reading
Choose the question you are working on.
People: who carries the change?
Start with what a leader owes the people who must change, then what happens when a team stops checking.
Process: did the work improve?
Start with why speed is not the same as progress, then a pilot that worked.
Data: what does the number establish?
Start with reports on AI use and how people decide, then ask how people decided.
Leadership: what do we remain responsible for?
Start with what a board should ask, then when it is right to automate a decision.
If this is useful, we can work through it together
Bring a decision your team is facing and we can test it against these prompts. The aim is to leave with the conditions written down.