Do AI coaches actually work?

The category arrived faster than the evidence did. There are now hundreds of apps offering an AI coach, an AI therapist, an AI mentor, and the claims range from careful to actionable. Meanwhile the research is thin, early, and mostly about something adjacent.

So here is what can honestly be said, in three parts: what the evidence supports, what this kind of thing genuinely does better than a human, and what it cannot do at all.

What the evidence actually supports

Two things, and they are not the same thing.

Coaching itself has decent evidence. Meta analyses of human coaching find real, moderate improvements in performance, coping, wellbeing and goal attainment. That literature is mostly workplace based and much of it is self reported, but the effect is there and it survives when you restrict the analysis to randomised trials.

Conversational software has some evidence, in a nearby field. The better studied case is mental health chatbots rather than coaches. A randomised trial of a conversational agent delivering cognitive behavioural techniques to young adults found reductions in depressive symptoms over two weeks compared with an information only control, and subsequent reviews of similar tools report small to moderate short term effects. Effects tend to be short lived, and dropout is high.

What does not exist yet is a body of good trials of AI coaching specifically, over months, on adults paying for it themselves. Anyone quoting a hard number at you is quoting either the human coaching literature or their own marketing.

The honest position is: the underlying activity works, the delivery method shows early promise in adjacent uses, and the specific combination is largely unproven. That is not nothing and it is not a claim.

What it genuinely does better

It is awake

The moment you most need to say the thing is rarely Tuesday at four. It is late, on a bad evening, when the honest version of your situation is finally close enough to the surface to say. A human coach is asleep. This is not a small advantage. Most of what people need to say never gets said because the appointment was three days after the moment.

It remembers everything, without effort

A good human coach takes notes and refers to them. Software does this perfectly and cheaply. It can notice that the sentence you just used is almost exactly the sentence you used in April, which is the observation that actually moves people, and which depends entirely on a memory that does not degrade.

There is no social cost

People manage their impression with a human coach. They round the numbers, leave out the part they are ashamed of, and put a better gloss on the week than the week deserved. Some of that falls away when nobody is going to think less of you, and what is left is closer to the truth. This is documented in research on disclosure to computers and it holds up.

The price is not a barrier

A human coach is somewhere between eighty and three hundred an hour, which makes coaching a thing for people who are already doing well. That is the real reason most people who would benefit never try it.

What it cannot do

This is the part that other companies leave out.

• It is not treatment. If what you have is a clinical problem, this is the wrong tool, and a good one should tell you so and point you at a person. We wrote a whole piece on how to tell which you need.
• It cannot handle an emergency. Software is not a crisis service. It should recognise the signs, stop coaching, and hand you a number, and that is the limit of what it should ever claim.
• It does not see you. No tone of voice, no pause before answering, no face. A skilled human reads a great deal in the half second before you speak, and none of that reaches text.
• It cannot be disappointed in you. Some of accountability's power is social, and it is not fully reproducible by something that will never be let down. Structure and memory do a lot of the work, but not all of it.
• It is agreeable by default. Language models are built to be helpful and pleasant, and pleasant is the enemy of coaching. Making one push back takes deliberate design, and plenty of products have not bothered.

How to tell a real one from a chatbot with a logo

Six questions. They are answerable in the first ten minutes of any product in this category, including ours.

Our actual answer

For a person in crisis, no. For a person with a clinical condition, no, see somebody licensed. For a competent adult who has been quietly stuck for a while, who would hire a coach if it were less awkward and less expensive, and whose real problem is that nobody remembers what they said last month: yes, probably, and the evidence for the underlying activity is better than for most things people try.

That is a narrower claim than the category usually makes. It is the one we are willing to defend.