Leading in a Language That Isn’t Yours
18. September 2026Where AI helps a leader think, and where two randomised trials show it stops.
An interest to declare first, because it runs against what follows: I sell coaching, and what I am about to describe is a set of situations in which software does part of that work well, at a scale and price no coach can match.
Reflection fails more reliably than any other part of leadership work, and for a structural reason rather than a personal one. It has no deadline, nobody is left waiting on it, and no commitment breaks when it does not happen, so it loses every calendar contest it enters.
What the evidence shows, and why it is unflattering
In a randomised trial over ten months, people working with a goal-focused coaching chatbot increased goal attainment by around fifty-five per cent against roughly twenty-four per cent in a control group, and the effect was comparable to human coaching on that task. The researchers are candid about the reason, and it is not insight: after the initial goal-setting conversation, most of the interaction was monitoring and asking again what had happened since. The system never tired of asking and never forgot what someone had committed to four weeks earlier, which is exactly the resource senior calendars lack.
The availability matters most at the top. At six in the morning before a board meeting, a chief executive can ask the question she would not ask a colleague, because asking a colleague would make her look underinformed on something she is expected to have settled. The number of people a senior leader can think unguardedly in front of is usually in single figures, and it falls with every promotion.
Then the same researchers found the opposite
A second randomised study, published in 2026 with a hundred and fourteen coachees in a global organisation and validated psychometrics across goals, motivation, resilience and wellbeing, compared accredited human coaches against automated AI coaching directly. The human arm moved every outcome substantially. The AI arm did not beat the group receiving no coaching at all on a single primary measure. The technology had not got worse; the task had changed, from executing a structured protocol to occupying the seat where a person normally sits on open-ended development questions.
The finding that makes this a leadership decision
Success in the second trial was predicted most strongly by how much self-efficacy and hope someone walked in with. Those who already believed they could develop gained the most, which means the people who most need developmental attention are the ones an unsupervised system serves worst. Deciding who gets a person and who gets a platform is therefore not a procurement question, whatever the licence cost suggests.
Move what matters most.
Talk to me. 15 minutes. Tell me about the decision you keep making the same way, and what you would need someone to notice.
