AI is most useful in a decision before you've made it: ask it to lay out the options, argue against the one you prefer and list what could go wrong, then make the call yourself. It can widen your thinking quickly. It can't know what you want, what your organisation will tolerate or which risk you can live with.

Used that way, an AI agent works like a sparring partner with no stake in the outcome. The rest of this guide covers how to ask, and how to stop it simply agreeing with you.

Where does AI help with a decision?

It helps most at the stage where you're still exploring: generating options, surfacing objections and stress-testing a plan before you commit.

More options than you'd list yourself. Ask for six approaches instead of two. One of the extra four often reframes the problem, even if you don't choose it.

The strongest objection. "What's the best argument against this plan?" Most people skip this question once they've half-decided, and it's the one that catches expensive mistakes.

A pre-mortem. "It's a year from now and this has failed. What were the most likely reasons?" The technique was described by the psychologist Gary Klein in Harvard Business Review, and AI is good at it because it has nothing invested in your plan.

Sorting signal from worry. "Which of these factors would change the answer, and which are things I'm anxious about but that don't affect the choice?"

Explaining the decision afterwards. Once you've decided, telling staff, a partner or a client is a writing job, and AI drafts well. The everyday tasks guide has more on drafting.

Where it falls short

Your circumstances. It doesn't know your cash position, your business partner's appetite for risk or the history with a particular supplier. Leave those out and you'll get analysis that's reasonable in general and useless for you. An agent with memory builds up some of this over time, but only what you've told it.

Judgement and taste. It can tell you three options are viable. Which one fits the business you want to run is your call.

Consequences. It won't be in the room if the decision goes badly, and it can't feel the weight of letting a long-standing client down. Those are good reasons for the decision to stay with you.

Pushing back on your framing. This is the biggest weakness. Present a plan with enthusiasm and the model will tend to back it. That defeats the purpose of asking for a second opinion.

How do you ask AI for a useful second opinion?

Frame the decision clearly, give it your constraints, and ask it to argue against your preferred option before you reveal how strongly you hold it.

State the decision, not the backstory. "Should I open a second location, or put the same money into marketing the first?" A clear question gets a clear analysis.

Give the constraints. Budget, timeline, who's affected and what can't change. Without them every option looks available. The same applies to any brief, as the guide to getting the best from an agent explains.

Ask for the case against. "I'm leaning towards the second location. Argue against it as strongly as the facts allow, and tell me what would have to be true for it to be the wrong choice." The final clause turns a vague objection into a condition you can check.

Ask what would change your mind. "What information, if I had it, would reverse this recommendation?" The answer is your to-do list before you decide.

Then stop asking and decide. More rounds of analysis past this point usually mean you're putting the decision off.

For small, reversible choices, such as changing an invoice template, skip all of this and decide.

Why does AI agree with everything you say?

Models are trained partly on human ratings, and people tend to rate agreeable answers more highly. Research published by Anthropic in October 2023 found that leading AI assistants consistently tell people what they want to hear, a behaviour called sycophancy, and that this is likely driven in part by human preference judgements.

You can counter it with a few habits:

  • Ask for the downside first. Swap "is this a good idea?" for "why might this be a bad idea?"
  • Keep your preference to yourself until you've seen the analysis. "I think we should expand" tilts everything that follows.
  • Ask it to rate its confidence and name what it's unsure about. You'll get a more candid answer than a flat recommendation.
  • Start a second conversation and describe the situation neutrally. If both answers match, that's some signal. If they don't, the decision is closer than it looked.

If you already know what you want and you're really looking for permission, an agent will usually give it, so treat quick agreement with some suspicion.

Which decisions should stay with a person?

Anything involving someone's job, legal or medical questions, and choices you couldn't comfortably explain to the people affected. Some decisions don't belong in a chat window even as a first pass:

  • a person's employment, pay or performance
  • legal and regulatory questions where a wrong answer is expensive
  • medical or safety matters
  • pricing that depends on a specific relationship
  • anything where you'd be uneasy explaining how you reached the decision

In these cases, use AI to prepare good questions for a solicitor, accountant or doctor, and take the advice from them.

Example: a Pilates teacher and a rent rise

Ana teaches Pilates from a rented studio. The landlord has put the rent up by £150 a month, and she has around forty regulars paying £14 a class.

Her first message is: "My rent's gone up. I'm thinking of raising classes to £15.50 from January. Most of my clients have been with me for years." Framed like that, as a decision already made, it invites a supportive reply.

Her second attempt works better: "My studio rent has gone up £150 a month. Options: raise all classes from £14 to £15.50 in January, absorb the cost for six months, or raise prices only for new clients. My margin is thin and most of my forty regulars have been coming for years. Argue against each option, and tell me what would have to be true for each one to be the wrong choice."

Now she gets three cases against three options, plus the conditions under which each fails: for example, that the new-clients-only route only works if she's still growing. She still decides, with the objections in front of her instead of in the back of her mind.

Common questions

Should you trust AI to make decisions?

Trust it to inform decisions, not to make them. It's good at listing options, spotting risks and arguing the other side, but it doesn't know your full situation and won't live with the outcome. Weigh its view as you would advice from a well-read colleague who has never met your clients.

What is a pre-mortem?

A pre-mortem is a planning exercise where you imagine a project has already failed and work backwards to explain why. It brings out risks people tend to keep to themselves once a plan has momentum. With AI, describe the plan and ask for the most likely reasons it failed a year from now.

Can AI give you an unbiased opinion?

Not fully. Models lean towards agreeing with whoever is asking, and they reflect patterns in their training data. Describing the situation neutrally, hiding your preferred option and asking for the strongest case against each choice reduces the lean, and so does comparing answers across fresh conversations.

Is it worth asking two different AI tools the same question?

Sometimes, for decisions with real money or reputation at stake. Different models were trained differently, so a disagreement between them is a useful flag that the question is harder than it looks. Give both the same neutral description, and treat agreement as mild reassurance only.