I keep coming back to one study. Researchers ran a pre-registered experiment with 758 consultants at BCG. Some worked with an AI model and some worked without one. On tasks the AI could handle well, the consultants with AI finished more of the work and finished it faster.

Then the researchers gave 373 of them a business case designed to fall outside what the AI could do. Without AI, 84.5% of the consultants reached the correct recommendation. In the two groups with AI, it was 60% and 70.6%.

The part that stays with me came next. The graders rated the AI-assisted recommendations as more persuasive, and that included the wrong ones.

I want to be fair to the study. It was one task, with early-career consultants, using GPT-4 in 2023. Fabrizio Dell'Acqua and his colleagues published it in Organization Science in March 2026. I take it as a warning about where our attention goes when an answer looks good.

The answer now arrives finished

For most of my working life, a weak answer looked weak. A first draft from a colleague had gaps you could see. A figure was missing, or the argument stopped half way.

That warning sign is often missing now. An AI answer arrives finished, with clean sentences and a confident summary at the top. In my experience it sounds just as sure whether it's right or wrong.

That's what I mean when I say sounding right is a style, not a signal. The polish tells you the writing is fluent. It does not tell you the answer is true.

This matters for leaders, because a leader is often the last person to see an answer before it becomes a decision. By then it may have passed through other people, and each of them may have trusted the polish too.

What I learned at Electrolux

At twenty-nine I recorded 241 per cent growth at Electrolux. I was dismissed within the year. Results are not progress. A result that looks finished can hide the question nobody asked. An AI answer can do the same thing, and it can do it in seconds.

Two findings I keep in mind

Two findings sit behind my concern.

The first is about habits. KPMG's co-published 2025 global study of trust, attitudes and use of AI surveyed more than 48,000 people in 47 countries. Fieldwork ran from November 2024 to January 2025. Of the employees who use AI at work, 42% said they had relied on AI output without evaluating its accuracy sometimes or more often.

I cannot tell from the survey why they did not check.

The second finding is about what happens when a person and an AI decide together. Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone reviewed 106 experiments in 2024. People working with AI did better than people working alone. But the pair did worse than the better of the two working on its own, and most of all on decision tasks.

Deciding is the part of the job that leaders keep, so I pay attention to that finding. The studies did not test leaders, and I am not saying they did.

The question I want leaders to ask

I want leaders to ask one question. Do you know how your people are deciding with AI?

Knowing which tools your people use is the easy part. The hard part is what happens between the answer arriving and the decision being made. Who checked the figure, and against what? Who would have noticed if it was wrong?

AI can draft the plan. It cannot decide whether the plan is right. That part stays with a person, and the person needs a reason to slow down at the right moment.

In practice I think this changes two things for a leader. First, you have to ask about the checking as well as the answer. "Where did this figure come from?" is a better question than "Is this ready?" Second, you have to make it safe for someone to say they haven't checked something yet. If people are praised only for polished answers, expect polished answers.

What AIR APAC has written down

AIR APAC, which I founded, has published a standard that names nine things a person does when deciding with AI. You can read them at airapac.org/standard/dimensions/.

These are the nine, in the Standard's own names:

  • Problem Framing
  • Scope and Constraint Definition
  • Evidence Verification
  • Assumption and Alternative Testing
  • Risk and Failure Anticipation
  • Decision Rights and Escalation
  • Stakeholder Alignment and Disclosure
  • Proportionate Commitment Under Uncertainty
  • Consequence Ownership and Correction

What I like about the list is that none of the nine is about knowing a tool. They describe what a person does around the answer. Evidence Verification, for example, means tracing a claim to its source and checking in proportion to the consequence. It is the habit I would want in the room when an answer looks this good.

What I won't claim

I don't measure anyone, and nothing on this site scores or assesses people. The views in this piece are mine. The research is other people's work, and I've tried to describe it in their terms and no further. AI helps people on many tasks, and the BCG study and the review of 106 experiments both found that. My point is narrower: a finished answer asks more of the reader.

Where to start

I give keynotes, briefings and workshops on this.

If you're planning an event and want this conversation in the room, tell me about the event.