Imagine you’re preparing a recommendation for your team. You ask an AI tool to help, and it gives you a clear, well-structured answer.
It sounds reasonable. It also happens to support the option you already preferred.
What do you do next?
You could tidy up the wording and move on. Or you could pause long enough to ask: “What has actually been checked here?”
That question matters because a convincing explanation and a sound decision are not the same thing. The tool may have helped you organise your thinking. It may also have given you an answer that needs substantial checking. Your job is to work out which parts you can rely on before someone else relies on you.
Start with the decision not the tool
Be clear about what you are asking AI to help with. Drafting possible discussion questions is different from recommending a consequential action affecting a customer or employee.
Name the decision, who it affects and what could happen if the answer is wrong. Use your organisation’s approved tools and data rules. Don’t paste confidential material into a service simply because it would make the prompt easier to write.
Then agree the boundaries. Is the output a starting point for discussion, a draft to review or information being used to support a decision? Who has the knowledge and authority to check it?
“A person will look at it” is a beginning, not a complete review process. Decide what that person needs to examine and when to bring in someone with relevant expertise.
Catch the part you want to believe
Catch, Challenge, Change is a practical way I use to explore assumptions and responses. Applied here, Catch begins with noticing your own reaction.
Are you relieved because the answer saves time? Impressed by how confidently it is written? Pleased that it agrees with your plan?
None of that proves the answer is wrong. It does give you a reason to separate your reaction from the evidence.
Try writing your current view in a sentence before reviewing the output. Then ask, “If this had argued for the opposite option, would I be checking it more carefully?” Treat that as a prompt for reflection, not a test that diagnoses or removes bias.
Challenge the evidence not just the wording
The US National Institute of Standards and Technology identifies confidently presented false or erroneous output as a generative-AI risk. Its July 2024 Generative AI Profile also recommends checking sources and citations. This is risk-management guidance, not a claim that every output is false or a measure of any particular tool’s current accuracy. NIST Generative AI Profile, section 2.2 and action MS-2.5-003.
For your decision, pick out the claims that actually matter. Check them against the original material, not another generated summary. Does the source exist? Does it say what the answer claims? Does it apply to your situation and date?
Consider a hypothetical proposal that says customers want a faster service. You would want to know where that came from. Recent customer feedback? An assumption in the prompt? A general statement with no evidence about your customers at all?
You can ask the tool to offer alternatives or identify weaknesses, but that second answer is still something to examine. It isn't independent proof that the first answer was reliable.
Change the next step when the evidence is thin
You don’t have to choose between accepting the whole answer and abandoning the tool.
Keep the parts that are useful and supported. Mark what remains uncertain. If a missing fact could change the decision, get it before proceeding. Where an appropriate low-risk trial is possible, define what you will test, who will review it and what would make you stop.
The practical question is: “What is the smallest responsible next step, given what we actually know?”
Sometimes that is a small experiment. Sometimes it is a conversation with a qualified colleague. Sometimes it is not acting yet. Catch, Challenge, Change helps structure the reflection; it does not replace technical checks, organisational policy or professional advice.
Listen when your team raises a concern
Now imagine you’re introducing the tool to your team and someone hesitates. It would be easy to label them resistant to change. But what are they concerned about?
They may be unsure what information they can use. They may not know how to check an answer. They may be worried about workload or what the change means for their role. Those are different questions, and enthusiasm on your part doesn’t answer them.
Ask, “Which part feels unclear or risky to you?” Listen before trying to persuade. Explain what has been decided, what is still being worked through and where they can get support. Don’t promise certainty about someone’s future that you don’t have.
You can encourage curiosity while taking a concern seriously. In practice, that might mean working through an approved example together and showing the checks, including where you decide not to trust an output.
Try a short decision review
Before acting on the next AI-assisted recommendation, ask:
- What decision are we making, and who could be affected?
- Which important claims have we checked against original sources?
- What assumptions came from us, and what remains unknown?
- Who is responsible for the decision, and what would make us change course?
This is a practical conversation starter, not a complete AI governance framework. Adapt it to the risk of the work and your organisation’s requirements.
In The Human Advantage, I explore how assumptions shape judgement, teamwork and change. The Leadership Mindset workshop gives leaders room to practise questioning their thinking on real challenges. Neither is technical AI implementation training.
If your team needs a conversation about the human side of change, tell me what they’re working through. A useful starting point is the decision in front of them, not the next piece of technology.
