What would change
your mind?

The difficult change may be in the assumptions.

A leader can understand an AI tool, see a convincing demonstration and still run the business in much the same way. That may be entirely reasonable: a demonstration is not a production system, and responsibility for customers, people and performance does not disappear because a task looks easier.

There is another possibility worth examining. The organisation may still be making decisions around limits that are changing: the time needed to understand a market, the cost of preparing a proposal, the availability of specialist knowledge or the effort required to build and test something. If those limits move, familiar choices about staffing, pricing, structure and investment deserve another look.

For TIP, this is central to absorbing AI. The tools become part of how people question, decide and work together. Changing how we think can be more demanding than learning how to use them, especially when our existing approach helped make us successful.

More evidence does not automatically settle an argument.

In a classic experiment, Charles Lord, Lee Ross and Mark Lepper gave people with opposing views on capital punishment conflicting research findings. Participants tended to judge evidence supporting their existing position more favourably. The study illustrates how prior beliefs can shape evaluation; it does not show that every disagreement is irrational, or that the same response inevitably occurs in a boardroom. [1]

Our practical concern is the unequal standard we may apply. Do we interrogate an inconvenient result while accepting a reassuring anecdote? An AI enthusiast can do this as readily as a sceptic. A supplier can present selected successes; a leader can use one failure to dismiss an entire possibility. Neither gives the organisation a sound basis for a decision.

Disagreement can improve the proposition.

A CFO asking whether a time saving survives the cost of checking is doing useful work. So is a service leader asking whether customers have to call back, or a technology leader asking who will maintain the generated code. These questions should change the design of an experiment.

We apply the same scrutiny to our own functional capacity evidence. Each visual identifies the work studied and the result reported. A task-level saving is not a measured result for the whole function. A survey, a reported case and a controlled trial do not become equivalent evidence because they each contain an attractive number. The relevant question is what the source measured, under what conditions, and what that allows us to conclude here.

Where the evidence is weak, improve it. Where it is relevant and robust, examine the assumption it challenges. Calling a sceptic resistant before doing that work is likely to close the conversation we need to open.

Make changing your mind a leadership practice.

Amazon’s principle of being right a lot asks leaders to seek different perspectives and “work to disconfirm their beliefs”. That is a more useful discipline than trying to win every argument. [2] In his 2016 shareholder letter, Jeff Bezos connects decision speed with reversible choices, course correction and a willingness to support a decision despite disagreement. [3]

Andy Jassy’s 2024 letter makes the distinction explicit: leadership is about reaching the best answer for customers, rather than ensuring that your original idea prevails. He describes changing a view when compelling new information arrives. [4] These are leadership practices we can learn from, not evidence that Amazon or any other organisation is immune to bias.

For a team, the useful habit is to state the current belief, the evidence behind it and the observation that would justify revising it. A leader can then explain a changed decision as a response to learning, with a clear account of what changed and what did not.

The evidence can require us to revise the optimistic story too.

METR’s early-2025 experiment found that experienced open-source developers took 19% longer on their familiar tasks when allowed to use AI, despite believing it had made them faster. [5] Its February 2026 update suggests performance may since have improved, but explains why selection effects make the new estimate unreliable. [6]

The lesson is to keep both the result and its limits in view. We should neither hide an inconvenient finding nor turn an older, narrowly scoped result into a permanent verdict. The tools, work and people matter, and our confidence should change as the evidence changes.

Give the disagreement a fair test.

Before running an experiment, bring the workflow owner, a user of its output and someone willing to challenge the proposal together. Write a short decision record using the questions alongside this paper. Agree what counts as success, what would cause a stop, and when the result will be reviewed.

For example, the current belief might be that a proposal requires several days because expert input is scarce. Test whether AI-assisted preparation lets those experts spend less time assembling material and more time judging the recommendation. Count the complete human effort, including corrections, and have the output assessed against the same quality standard. A faster draft that creates more review is not yet a capacity gain.

Agree the next action in advance. Good evidence might justify a larger trial; mixed evidence might require a narrower use; a material failure might end the experiment. Repeatedly moving the threshold after the result protects a position rather than improving a decision.

Make it possible to act on what people learn.

A changed belief will not necessarily change the work if budgets, targets or rewards still favour the old approach. If a manager expects every hour released to trigger a headcount reduction, reluctance to reveal capacity may reflect the incentive structure. Another presentation about AI will not resolve that tension.

Leaders need to explain what the capacity is for, which decisions people can change and how learning will be recognised. Some capacity may strengthen the function; some may support a shared enterprise priority. Those choices belong to the organisation, along with responsibility for carrying them through.

TIP helps people create time, test assumptions and examine the work together. Where participants agree to recording, transcripts can help revisit the questions and commitments in a discussion. They do not let us diagnose someone’s motives. The outcome we want is a better decision, supported by evidence that people can inspect and a team prepared to act on it.