AI coaching practice vs human peer practice: what each is for

What you can do is match the format to what you're trying to learn this week.

No direct efficacy comparison was located in the research reviewed here. The closest randomised study of simulated-client training says outright that without direct comparisons to established training methods, "we cannot determine the relative effectiveness" of the AI approach (CARE, CHI 2026). This study cannot establish which format is better for coaches.

What you can do is match the format to what you're trying to learn this week. The two arrangements make different things possible, and the differences are concrete enough to choose between.

First, be clear which arrangement you mean

"Practice" covers at least five things, and they are not interchangeable:

An AI practice client. Software plays a fictional client; you coach.
A human roleplayer. A person plays a character you've agreed on; you coach.
A peer bringing a real topic. Another coach brings something they actually want to think about; you coach. Usually you swap.
Mentor coaching. Feedback and reflection on observed coaching against competency standards; check the mentor’s qualifications for your route.
Supervision. A reflective space for your practice as a whole, including what it's doing to you.

This guide is about the first and third. Mentor coaching and supervision serve different development and oversight needs. Neither AI rehearsal nor ordinary peer practice replaces them when you need qualified oversight.

What each arrangement gives you

	AI practice client	Peer with a real topic
Availability	Subject to account access and usage limits	Whenever you both can, which is the constraint
Repeatability	Revisit a scenario within the tool’s limits; responses may change	Their topic moves on, as topics do
Control of the scenario	Some tools accept your brief; others provide preset scenarios	You get what they bring
Stakes	No real client's problem is at stake; learning and privacy risks remain	A real person’s time, trust and topic are involved
Feedback	Automated feedback if the tool offers it; quality varies	A person telling you what your questions felt like
Privacy	Use fictional material only; account and conversation data still need protection	Their material is real, which needs a proper agreement
Credential hours	Do not assume simulation qualifies as client experience	Can qualify, depending on the body and the arrangement

Rehearsing a specific moment

Repetition without using another person's practice time. Where the tool allows restarts, you can repeat a difficult opening and stop when you hear yourself doing the thing you're trying to stop doing. Agree any repetition with a human partner rather than assuming they want to repeat a scene.

Counsellor trainees who ran text sessions with ChatGPT as a fictional client described a flexible, low-pressure setting and reduced performance anxiety before working with real clients (Behavioral Sciences, 2025). One valued having "time to really think about the methodology and the theory I chose". That is four people describing their experience in a counselling programme, so hold it lightly, but it describes a use for the format, not evidence of better coaching.

The participants also named limits. The client was "very agreeable... it just went with what I was saying". Emotional depth was thin: "it doesn't replicate the emotional inflection of real clients." Sessions resolved too neatly. And the persona stayed culturally blank unless someone wrote culture into the prompt: "Naturally, it's very neutral. If you want culture or stereotypes, you have to be very specific."

In the CARE experiment, the practice-only group used significantly less empathy after practice, while their confidence in exploration skills rose. They also made fewer inappropriate suggestions, with no statistically significant improvement in reflections or questions. The authors propose two mechanisms: adapting to patients who resisted suggestions but did not respond differently to empathy, and a less emotionally forthcoming post-test patient. This is a caution from one immediate simulated counselling experiment, not an established general effect on coaches.

What a peer with a real topic gives you

A peer brings something they want to think about, and can tell you when you missed it. Agree to ask for that feedback rather than expecting it automatically.

That covers the parts of coaching that only exist when another person's stake is real: contracting at the start and meaning it, following an agenda that turns out not to be the one they opened with, sitting through a silence that is genuinely uncomfortable, noticing that a question landed badly and repairing it. A simulation can portray disappointment, but only a person can tell you how your coaching affected them.

Peer feedback is not automatically better than automated feedback; it's differently fallible. Partners may be polite or comment on what they'd have done rather than what you did, and a peer at your stage may not see what a mentor coach would. Asking about a specific behaviour rather than for a general verdict gets you further: "did you feel I checked what you wanted from the session?" rather than "how did I do?"

The hours question, which is separate

If you're building experience towards a credential, the format matters for a reason that has nothing to do with learning.

The ICF counts one client coaching hour as a full 60-minute session "with a client who has hired you specifically as their coach", in real time. It treats barter, "such as peer-to-peer coaching or coaching in exchange for goods or services", as payment, and it excludes coaching inside a training programme or with students, trainers and mentors unless it happens outside the classroom with a real client and separately from training credits (ICF experience requirements, checked 11 September 2026).

So a properly set-up peer exchange can count for ICF purposes. Not every peer conversation does, and other bodies set their own rules, so check the ones you're applying under rather than carrying an ICF answer across. The ICF experience page describes coaching real clients and does not establish eligibility for AI simulation. That does not settle other bodies’ policies. Keep rehearsal out of your hours log and describe it as development if you record it at all.

Choosing this week

Choose according to what you need:

For repetition, use a rehearsal to try a specific moment different ways.

For a person's actual response, arrange practice with a consenting partner who brings a real topic.

For eligible hours, check the real client relationship and session against your body's rules.

For qualified oversight, seek mentor coaching or supervision rather than substituting either practice format.

One suggested arrangement is to rehearse, coach a consenting person, get specific feedback and choose what to work on next. The evidence reviewed here does not establish an optimal mix.

Sources

Can LLM-Simulated Practice and Feedback Upskill Human Counselors? A Randomized Study with 90+ Novice Counselors, Proceedings of CHI 2026, DOI 10.1145/3772318.3791821. 94 novice counsellors; no human-training comparison group; counselling, not coaching.
Learning Through Simulation: Counselor Trainees' Interactions with ChatGPT as a Client, Behavioral Sciences, 2025. Nine transcripts, four focus-group participants; self-reported experience, not a comparison of formats.
ICF coaching experience requirements, checked 11 September 2026. Applies to ICF credentials only; shorter sessions count proportionately.