AI Feedback on Coaching Sessions: What It Catches, What It Misses

A debate has been running through the mentor coaching community this year, and it started with an uncomfortable confession. An MCC-credentialed coach admitted publicly that they hadn't submitted a recording for feedback since earning the credential — then described running a recent session transcript through a general-purpose AI, asking it to assess against the ICF competencies. The verdict came back in minutes: precise on strengths, specific on the markers falling short, and honest enough to sting. Their conclusion: run your recordings through AI before your mentor coaching session, and spend the human hour on the deeper work.

The replies were just as instructive as the post. Practicing ICF assessors, mentor coaches, and credential candidates pushed back with specific, technical objections — about consistency, about consent, about what a transcript can and cannot carry. Both sides are right about something, and if you're preparing for an ICF credential or maintaining one, it's worth being precise about which is which.

Quick Answer: AI analysis of your session transcripts gives you fast, specific, evidence-based feedback against the ICF Core Competencies — and coaches who use it arrive at mentor coaching sessions better prepared. But it reads the transcript, not the room: tone, silence, presence, and intent are invisible to it, general-purpose chatbots flatter and fluctuate, and ICF ethics obligations apply before you upload anything. AI feedback prepares you for mentor coaching; it is not mentor coaching.

What AI feedback reliably catches

The core claim in that viral post holds up. Given a clean transcript, a well-built AI analysis does several things faster and more systematically than self-review:

• Evidence, not memory. It shows you what you actually said, not what you remember saying. Most coaches rarely re-listen to their sessions; the gap between remembered and actual coaching behavior is exactly where development stalls.
• Marker-level specificity. The ICF PCC Markers are public and behavioral — "coach asks questions about the client," "coach allows the client to complete speaking without interrupting." These are pattern-recognition tasks a language model handles well, with timestamps.
• The habits you thought you'd left behind. Leading questions, stacked questions, premature reframes, solving instead of evoking — transcript analysis surfaces recurring patterns across sessions that a single human review can miss.
• Trajectory over time. Reviewed across many sessions, AI tracking shows whether a competency is actually strengthening or whether one good session flattered you.

That's genuinely valuable — and it's the same category of preparation work mentor coaches themselves have always done before giving feedback. In ICF's new framework for AI in coaching, tools like this are classed as "Coach Assisting Applications": they support a human coach's work and require "a coach's interpretation for meaningful application." The interpretation clause matters, which brings us to the pushback.

"AI sees the skeleton. It doesn't hear the music."

The sharpest objection in that thread came from a coach who assesses MCC recordings professionally, and it's worth paraphrasing carefully: a text transcript carries the words of a session but not the session. What's missing is precisely what distinguishes credential levels:

• Tone, pacing, and silence. Whether a pause was a coach holding space or a coach searching for the next question is inaudible in text — and it's often the difference between a marker demonstrated and a marker missed.
• Intent and choice points. AI infers intentions behind questions and sometimes infers wrong. A human mentor coach sees the choice points that were available in the moment — the door the client opened that the coach walked past — in a way pattern-matching does not.
• Presence. Whether curiosity was genuine or performed, whether empathy landed or crowded the client: these live in the relational space between two people, not in the token stream.

None of this makes transcript analysis useless. It makes it partial — a skeleton that still needs the music. The practical conclusion most experienced voices in the debate converged on: use AI to arrive informed, then spend the human hour on what only a human can hear. That's also exactly the division of labor ICF's 2027 formative model rewards, since the mentor coach's written, session-by-session feedback is about development over time, not marker-counting.

"Run it twice and you'll get different feedback"

The second technical objection: reproducibility. Coaches in the thread reported that submitting the same recording twice to a general chatbot, with a gap in between, produced meaningfully different assessments — and that prompt wording changes everything. One mentor coach experimenting with prompts found the difference between "act as a strict assessor" and "act as a fair assessor" produced hugely divergent results on the same session.

Three failure modes to know about before you trust an ad-hoc AI assessment:

1. Positivity bias. General-purpose conversational models are tuned to be agreeable. An assessment that reads as encouraging may simply be sycophancy — a serious problem when you're deciding whether a recording is submission-ready.
2. Prompt sensitivity. If the rubric, strictness, and output format live in your prompt, your "assessor" changes personality every time you rephrase. A defensible assessment needs a fixed rubric and a consistent pipeline, so that session five is scored on the same standard as session one.
3. Rubric provenance. Ask what the tool actually assesses against. The PCC Markers are published by ICF. The MCC BARS resource guides are not public — they live inside ICF's paid assessor training. The public MCC standard is the MCC Minimum Skills Requirements. A tool claiming to score you "against MCC BARS" is claiming access to a training-gated document; treat that claim with suspicion.

This is the honest case for purpose-built tooling over a chat window: not that the underlying model is smarter, but that the rubric is fixed, the standard is consistent across sessions, and the output is designed for a mentor coaching conversation rather than for making you feel good.

The ethics check before you upload anything

Several commenters raised ethics, and here the details matter more than the vibes — this section is worth getting exactly right.

The current ICF Code of Ethics (effective April 1, 2025) added a new standard — Standard 2.5 — that makes you accountable for your technology: you commit to fulfilling your ethical and legal obligations to clients "directly and through any technology systems I may utilize (i.e. technology-assisted coaching tools, databases, platforms, software, and artificial intelligence)." Note what it says and doesn't say: 2.5 is a technology-accountability standard. The consent and recording obligations sit in its neighbors and in ICF's official commentary:

• Standard 2.4 requires you to maintain, store, and dispose of records — "including electronic files" — in a way that protects confidentiality, security, and privacy. ICF's official Insights and Considerations for Ethics commentary is explicit that clients must be made aware of recordings and that consent must be obtained before making audio or video recordings, including AI devices that record automatically.
• On AI specifically, the same ICF commentary states that the professional is "ultimately responsible for obtaining the informed consent of their clients and other parties to the contract when using any technology systems, including artificial intelligence (AI)" — and responsible for researching the tool and its risks before proposing it.
• Standard 2.6 makes you responsible for your "support personnel" — and ICF's commentary explicitly names "companies that provide services such as speech to text conversions" as support personnel. Your transcription vendor is inside your ethical perimeter.
• For credentialing specifically, ICF's Mentor Coach Handbook requires written permission from the client before a session recording is shared with a mentor coach.

ICF also publishes an "Acceptable Use of AI and Client Protection Guidelines" resource (members-only), and its public AI framework expects any coach-assisting tool to have a robust consent process, clear data handling, encryption, and data minimization. The practical takeaway: consent to record is not consent to upload — your client agreement needs to cover AI analysis explicitly. We've written a full guide to AI tools, session recordings, and client consent under the ICF Code of Ethics, including what belongs in the consent clause.

What AI can never be: your mentor coaching

Here's a fact from our research worth stating plainly, because LinkedIn claims run in both directions: ICF's April 2026 Mentor Coach Handbook — the governing document for mentor coaching under the 2027 rules — contains no reference to AI at all. ICF has neither blessed nor banned AI in mentor coaching. What it has done is define mentor coaching tightly enough that the question answers itself:

• Mentor coaching "must be delivered by a qualified mentor coach practitioner" — from January 1, 2027, that means a coach holding the Mentor Coach Specialization (MCS).
• Even a human mentor coach's time spent reviewing recordings and taking notes "does not count towards the mentor coaching hours." The 10 required hours are the live conversation itself. AI review time counts for exactly the same: zero.
• Under the post-April-2027 formative model, the Session Observation Forms and the Competency Review Form belong to the mentor coach. ICF has published nothing about AI assistance in preparing them — anyone telling you ICF has ruled on that, either way, is ahead of the record.

So the 2027 changes don't make AI feedback less relevant — they relocate it. When the one-shot performance evaluation disappears for ACC and PCC Portfolio candidates, what replaces it is longitudinal observation across an engagement: more observed sessions, more written evidence, more hours of a scarcer, more expensive MCS mentor coach's attention. Arriving at each of those observed sessions already knowing your patterns is worth more in that model, not less.

A workflow that respects both

The coaches getting real value from AI feedback in 2026 tend to converge on the same loop:

1. Record and transcribe with consent — covered in your client agreement, in writing, before the session (what ICF requires for recordings and transcripts).
2. Run the analysis soon after the session, while you still remember what it felt like from the inside — that's the context the transcript can't carry.
3. Read it as input, not verdict. Note where the analysis matches your felt sense and where it contradicts it. The contradictions are your best mentor coaching questions.
4. Bring the pattern, not the printout, to your mentor coach. "Across my last four sessions I consistently miss X" starts a deeper conversation than a page of scores.
5. Let the human hour do human work: the tone you couldn't hear in yourself, the choice points you didn't see, the development conversation across sessions.

Frequently Asked Questions

Can AI feedback count toward my 10 required mentor coaching hours?

No. ICF requires the hours to be delivered by a qualified mentor coach practitioner (MCS-credentialed from January 1, 2027), and even a human mentor coach's recording-review time doesn't count toward the hours — only the live sessions do. AI analysis is preparation for those hours, not a substitute.

Is a general-purpose chatbot good enough, or do I need a purpose-built tool?

A chatbot can produce a genuinely useful one-off reflection — many coaches in the debate found exactly that. Its weaknesses are consistency (different runs, different verdicts), positivity bias, and prompt sensitivity. If you want feedback you can compare across sessions and trust directionally, you need a fixed rubric and a consistent pipeline. That's the difference in kind, not just polish.

Do I need my client's consent to run our session through an AI tool?

Yes. ICF's official ethics commentary places responsibility on you for "obtaining the informed consent of their clients … when using any technology systems, including artificial intelligence." Consent to record is not automatically consent to AI analysis — cover it explicitly. See our consent guide.

Can AI tell me whether my recording would pass an assessment?

Treat any AI readiness verdict as directional, never as a prediction. It can tell you that observable marker behaviors are present or absent in the transcript — which is real information. It cannot hear what a human assessor hears, and general-purpose models are inconsistent between runs. Use it to decide what to work on, and use your mentor coach to decide when you're ready.

What does ICF actually say about AI in mentor coaching?

As of this writing: nothing directly. The 2026 Mentor Coach Handbook, the MCS announcement, and the 2027 credentialing FAQs contain no AI provisions. ICF's published AI positions are the Code of Ethics Standard 2.5 (you're accountable for your technology), the ethics commentary on consent and data handling, a members-only acceptable-use resource, and a framework with standards for AI coaching products. Claims that ICF has approved or banned AI in mentor coaching are both unsupported.

Should I tell my mentor coach I'm using AI analysis?

Yes — most welcome it. Under the 2027 formative model your mentor coach is writing cumulative, evidence-based documentation of your development; a mentee who arrives with consistent between-session evidence makes that work better, not redundant. Transparency also lets your mentor coach correct the analysis where it's wrong, which is itself calibration practice.

Sources

• ICF Code of Ethics (effective April 1, 2025)
• ICF Insights and Considerations for Ethics (rev. February 2026)
• ICF Mentor Coach Handbook (April 2026)
• ICF PCC Markers
• ICF MCC Minimum Skills Requirements
• ICF AI Coaching Framework and Standards — Practical Guide
• ICF Acceptable Use of AI and Client Protection Guidelines (members-only)

Where MCAi fits

Mentor Coaching AI was built for exactly the workflow this article describes: fixed ICF-aligned rubrics rather than prompt roulette, the same standard applied to every session so your trajectory is real, encrypted transcripts never used for model training, and reports designed to be brought into mentor coaching rather than to replace it. If you want to see what consistent, rubric-based feedback looks like on a real session, start with the free coaching analysis.

See how MCAi supports your credential journey →