Mentavo
The debate about AI coaching tends to run along two extremes. On one side, the expectation that a language model can answer personal questions that normally take years. On the other, the dismissal that it's a chatbot which agrees with everything.
Both are partly true, and that's exactly why neither helps. The more interesting question sits in the middle: which concrete jobs do these systems do reliably in everyday life — and which don't they?
This piece describes both, without promising effects and without blurring the line.
The short version
- The useful part of AI coaching is sorting and retaining, not uncovering deeper truths.
- A system without context from your everyday life can only answer generically — however good the model is.
- Agreement is the most common weakness of these systems and the hardest to notice.
- What an AI coach cannot do: carry responsibility, know you, replace treatment.
- The benefit usually isn't created during the conversation but in what's left afterwards.
- A system becomes checkable when it names what a statement is based on.
What these systems genuinely do
1. Translate the vague into language
The most common and most underrated contribution. Describe something unformed and you get it back in sentences more precise than the starting state. That sounds trivial, but it's the point at which many conversations first move at all.
The effect depends less on the system than on the act of formulating itself. Why things clarify the moment you say them out loud — even to a listener who doesn't need to understand them — is covered in the piece on saying it out loud.
2. Ask questions you don't ask yourself
Not because the system is smarter, but because it isn't involved. It has no stake in how the matter turns out and doesn't have to live with the relationship afterwards.
The same phenomenon explains why it's easier to give a friend good advice than to take it — covered in the piece on why you give better advice than you take.
3. Make connections across longer periods visible
This is the capability a conversation between people specifically lacks: a system holding entries from several months can point out that something is appearing for the fourth time.
The precondition is that it has that context at all. A general chat window starting from zero each time cannot do this in principle — regardless of model quality.
4. Turn a conversation into a next step
Practically the most valuable part and the least impressive: recording what was discussed and turning it into something that reappears on the right day.
Without that step a coaching conversation stays inconsequential — with people just as much as with systems. Why insight alone changes nothing is covered in the piece on self-awareness without change.
What they can't do
Disagree when it gets uncomfortable
Language models are trained to come across as helpful and agreeable. That produces a systematic bias: describe a decision and you'll get confirmation more often than objection — including when objection would be warranted.
This is the most important weakness, because it feels good. How to spot it and what it does over time is covered in the piece on when AI always agrees.
Know you
A system knows what was recorded. It doesn't know what sat between two sentences, what you deliberately left out, or how you come across in a room.
That's not a technical shortcoming waiting to be fixed; it's a property of the data. Confuse the two and a well-phrased answer gets mistaken for understanding.
Carry responsibility
A suggestion remains a suggestion. Whether a conversation gets had, a job gets quit or a relationship gets ended is entirely yours — and the system doesn't bear the consequences with you.
Replace treatment
This is the clearest line and the one most often blurred. An AI coach is not therapy, not a diagnosis and not crisis support. With persistent distress, suicidal thoughts or an acute crisis, support belongs with a professional — not in a chat window.
Why context makes the difference
Most disappointments with AI coaching share a cause: the system knows only the current sentence. So it answers generically, and generic answers to personal questions read like a self-help column.
With context the kind of answer changes. Not because the model is better, but because the data is different: if it's known that the same question already came up in July, and that the last three weeks ran differently from usual, you get a follow-up question instead of advice.
That's what the coach in Mentavo is built for: the conversation knows journal, check-ins, goals and routines as far as they exist, and can create an appointment or a task from it on request — each time only after confirmation. Calendar and Apple Health are separate switches and off by default.
How to recognise a usable system
Four tests you can apply within a few conversations:
- Does it name the basis? A statement about you should refer to something concrete and traceable — not to an assessment arriving from nowhere.
- Does it ask back? Systems that deliver a complete answer immediately have usually simplified the question.
- Does it occasionally disagree? If no objection appears in twenty conversations, that isn't a good sign.
- Is anything left? After closing the window something should exist — a note, an appointment, a sentence. Otherwise it was entertainment.
The realistic expectation
An AI coach is neither a counterpart that understands you nor a tool without use. It sits between: a system that's good at bringing the unformed into language, asking follow-up questions, and holding onto what would otherwise be lost.
Expect that, and you'll rarely be disappointed. Expect understanding, and you will be regularly.
What it means to entrust something personal to such a system in the first place is covered in the piece on telling an AI something personal.
What a usable conversation looks like
Most conversations with these systems go badly because they open with a question that demands an answer: "what should I do?" That reliably produces advice, and advice is the least useful part.
A different order works better and takes about ten minutes:
- Describe rather than ask. Set out the situation first, with no question attached. That alone usually produces a more precise version than the one you had in your head.
- Ask for questions, not answers. "Ask me three questions about this" leads somewhere different from "what do you think?"
- Request objection explicitly. "What argues against it?" at least partly offsets the built-in agreement.
- Close with one sentence. Not a summary of the conversation, but what happens next — with a date.
The last point decides whether the conversation has consequences. Without it you're left with a good feeling and a memory that fades in three days.
One practical note: if an answer sounds conspicuously good, it's worth testing. These systems phrase things convincingly even when the basis is thin — the question "what are you basing that on?" separates the two quickly.
What this article doesn't claim
There is no robust evidence that AI coaching relieves symptoms, changes behaviour or outperforms another method. This piece describes functions, not effects.
It also doesn't claim these systems suit everyone. Some people sort their thinking in conversation with others, while walking, or in writing with no counterpart at all — and get further that way.
With psychological distress lasting weeks, with suicidal thoughts, or in an acute crisis, professional help is the right route. In the UK, Samaritans can be reached free on 116 123, around the clock; in the US and Canada, the 988 Suicide and Crisis Lifeline is available by call or text.
In closing: good at sorting, poor at understanding
The useful band of AI coaching is narrower than the marketing suggests and wider than the dismissal allows.
It sits where language, follow-up questions and continuation are involved: getting something diffuse into sentences, being asked the question you've been avoiding, and ending up with a sentence that's still there tomorrow.
Everything beyond that — understanding, judgement, responsibility — stays with people. That isn't a temporary limitation. It's the division of labour.
