Practising by Talking to an AI

Why talking to an AI works at all

The gap most people have is not knowledge, it is retrieval under load. You know the answer, and something about being asked out loud, in real time, by someone who is waiting, makes it come out worse than it does in your head. Closing that gap requires reps of the exact thing - being asked and answering - and reps are what an AI partner supplies without limit.

It also removes the social cost. A colleague will do this twice before it becomes an imposition. An AI will do it forty times, at midnight, and ask the same uncomfortable question again after you fumbled it, which is precisely the repetition the fumble needed.

And it produces the one thing rehearsing in your head cannot: an actual spoken answer, with a real length, real pace and real hesitations. Almost everyone underestimates how long their answers run, usually by a factor of two, and no amount of mental rehearsal reveals that.

Typing to a chatbot is not speaking practice

Text chat trains writing. It gives you unlimited time to compose, a backspace key, and no pressure to fill silence - the three things that are absent from every real conversation you were trying to prepare for.

If a tool is going to help your speaking, the practice has to be spoken out loud, in one take, with the clock running. That is the only version that engages the machinery that fails you in the room.

This is worth checking before you commit to a tool, because a lot of products described as speaking partners are text products with a play button. The test is simple: can you answer without touching the keyboard, and does it hold you to a length.

What to look for in a free one

It should ask one question at a time and then wait. A partner that delivers three questions in a paragraph has turned a conversation into a reading exercise, and you will answer the last one and forget the first two - which is also what happens with human interviewers, but you cannot practise recovering from it if the format never creates it.

It should follow up on the weakest part of your answer rather than moving on politely. The value is concentrated in the second question, because the first one you were ready for.

It should measure something. A partner that only talks back leaves you with an impression of how it went, and impressions are exactly what recording exists to replace. Pace, length and filler rate for each answer are the minimum worth having.

It should let you pick the pressure. Practising a friendly standup and practising a hostile investor are different exercises, and doing only the friendly one builds a confidence that does not survive contact.

What it cannot give you

Stakes. Nothing is at risk, and part of what breaks people in real meetings is knowing that something is. Practice narrows the gap without closing it.

A face. Most of what you read from a listener - the moment they stop following, the small frown at the number you glossed - comes through an expression, and answering into a text box or a voice with no face trains none of that.

Genuine standards. An AI will accept an answer that a person would have pushed back on, so you have to supply the discipline it will not. Set your own rules before the session and hold to them: thirty seconds a question, no restarting a sentence, no answering a question you were not asked.

Silence that costs something. A human pause is uncomfortable and that discomfort is the thing you are training against. An AI pause is free, which makes pause practice easier than it will be on the day.

How to run a twenty-minute session

Pick one scenario and stay in it. Mixing an interview, a pitch and a standup in one session gives you three shallow reps instead of six useful ones.

Answer in about thirty seconds: a claim, one piece of evidence, then stop. The stopping is the part being practised. Most people keep talking past the point because the silence afterwards feels like a failure, and it is not.

Do not restart. When an answer goes wrong, finish it anyway, because finishing badly is a skill and it is the one you will need. Restarting trains you to restart, and you cannot restart in a real meeting.

Take the worst two answers and do those questions again at the end. That is the whole session: fifteen minutes of reps and five minutes on the two that failed.

When to stop talking and start recording

A conversational partner is the right tool for retrieval under pressure. It is the wrong tool for measuring a habit, because every session has different questions and different lengths, so nothing is comparable week to week.

For habit work you want the opposite: the same prompt, the same length, recorded solo, so the only variable is you. That is the exercise that shows a filler rate falling from nine per minute to four across a fortnight.

Most people need both, in roughly a two to one ratio - daily solo recording to establish and track the habit, a conversational session twice a week to make sure it survives being interrupted. Doing only the second is how you end up comfortable talking and unable to say why nothing changed.