Using AI as a Speaking Coach: Feedback Sessions for Your Clients

A presentation coach can record a client's talk, transcribe it, then use AI to count words per minute, tally filler words and map the argument's structure against a rubric. The coach then spends the session on delivery, audience and next steps. AI measures what the transcript holds; the coach judges what it cannot.

The market for this is crowded with consumer apps that score a solo practice run and show a dashboard. They suit a client rehearsing alone at midnight. What they lack is your rubric, your knowledge of the room the client will face and the conversation after the numbers. That is the part clients pay a coach for, and the workflow below is built to protect it.

Follow me on Instagram@sagnikteaches

Everything here uses general tools, not one product: a recording, a transcript, and a general AI assistant. Because transcription and assistant features change often, I name no prices and no versions; check your tool's current terms and settings before you start.

Connect on LinkedInSagnik Bhattacharya

The workflow in five steps

1. Consent and set-up. Written agreement, tool chosen, recording method tested.
2. Record. A real or rehearsed talk, with the audience context noted.
3. Transcribe. Verbatim settings on, then a quick accuracy check.
4. Analyse. Pace, fillers, structure, language patterns, scored against your rubric.
5. Coach. You review the numbers before the client sees them, add what only a person can judge, and set one or two targets.

As a realistic estimate, preparation for a ten-minute talk takes 20 to 30 minutes once the routine is set: five for upload and transcription, ten for analysis and checking, ten for your own notes. The first time will take longer.

Subscribe on YouTube@sagnikteaches

Consent first: wording and limits

A client's voice and words are personal data, and a talk may contain commercial secrets or other people's details. Get consent before the first recording, in writing, and make it specific. A short clause you can adapt (illustrative; have your own adviser check it):

Recording and analysis. With your agreement, I will record your practice talks and coaching sessions and create transcripts. I may use [named tool types, e.g. a transcription service and an AI assistant] to count pace and filler words and summarise structure. Recordings and transcripts are kept for [period] in [location], are visible only to me, and are deleted on request or at the end of our work. Please do not include confidential client or colleague details in practice talks unless you have permission to share them.

Three practical points follow from it. Name the tool categories so the client knows data leaves your laptop. Choose a tool whose terms keep recordings out of model training if that option exists, and check how long it stores audio. And agree a rule for talks that contain material from the client's employer: practise with a sanitised version, or get the employer's permission. For the wider question, what to check in an AI tool's privacy policy is a useful companion, and mentioning AI use in client contracts covers the agreement side.

Recording: get clean audio before anything else

Garbage audio makes everything downstream unreliable. A phone placed on a table a metre from the speaker works for most practice talks. In an online session, ask the client to record locally if the platform's recording is poor. Use the same set-up each time so changes in your numbers reflect the client, not the microphone.

• Note the audience: five colleagues, a board, a conference hall. Feedback depends on it.
• Note the time limit the client faces. A talk that overruns by two minutes needs structure feedback, not only pace feedback.
• Note whether it was a first run or a polished one. Early runs have more fillers; compare like with like.

Transcription: the filler-word trap

Many transcription tools produce clean text by dropping "um", "uh", false starts and repeats. That is helpful for minutes and useless for filler analysis. Before you rely on counts, run a test:

1. Record 60 seconds of yourself deliberately saying "um" six times and "you know" three times.
2. Transcribe with the default settings, then with any verbatim or "include filler words" option.
3. Count what appears. If the tool keeps four of six, you know its recall is partial and you can say so to clients.

I cannot give you a universal accuracy figure; it varies by tool, accent and audio. What matters is that you know your own tool's behaviour and say "approximate" when you report the numbers. Use them as trends: 14 fillers in the first run and 6 in the third is a meaningful change even if the tool missed a few each time.

Some editing tools do the counting for you. Descript, for example, detects common fillers such as "um" and "uh" in English transcripts, underlines them in the script and lists every one with its timestamp in a Remove filler words sidebar. That list is a ready-made count with time marks, which is exactly what you want for a by-minute chart. Two cautions: work on a copy, because the sidebar's Delete option removes the audio as well as the text, and you will want the original to replay in the session; and for a client presenting in another language, check whether your tool detects fillers in it at all.

Analysis: what a transcript can tell you

Pace

Words per minute is simple arithmetic: transcript word count divided by speaking minutes. An illustrative worked example. A client delivers a talk lasting 9 minutes 30 seconds, and the transcript runs to 1,425 words.

Speaking time: 9.5 minutes
Words: 1,425
Pace: 1,425 / 9.5 = 150 words per minute

Many coaches treat roughly 130 to 160 words per minute as a comfortable conversational range for English; that is a rule of thumb from practice, not a published standard, so adjust for your clients and rooms. The more useful number is variation: ask the assistant to report pace per minute of the talk. If minutes 1 to 3 run at 128 and minutes 6 to 8 at 176, the client rushes the middle, which usually means they are nervous about their weakest section.

Filler words and verbal habits

Ask for counts by type and by minute, and a list of the sentence openers the client leans on ("so", "basically", "I think"). Seeing "basically" nine times lands harder than "you say it a lot".

Structure

This is where AI is most useful and most often over-trusted. Ask it to read the transcript and report what it finds against a structure you define, then you verify.

Below is a transcript of a 10-minute talk. Without rewriting it, report:
1. The single main message in one sentence, as you read it.
2. The three or four main sections, with the minute each starts.
3. Where the talk states what the audience should do next, quoted exactly, or "not stated".
4. Any section longer than 3 minutes with no example, story or number.
5. Hedging phrases (maybe, sort of, I guess, I think) with counts.
Quote the transcript for every point. If something is unclear, say so instead of inferring.
[transcript]

An illustrative reply: "Main message: the team should move invoicing to a monthly cycle. Sections: problem (0:00), current process (2:10), proposal (5:20), objections (8:40). Call to action: not stated. Longest section without an example: current process, 3 min 10 s. Hedging: 'I think' x 11, 'sort of' x 4."

What to check: does the stated main message match what the client meant? If the AI says the main message is "monthly invoicing" and the client says "cash flow", the talk has a clarity problem, and that mismatch is a coaching moment, not an AI error. Do not skip verification, though: tools sometimes misattribute minute marks if the transcript lacks timestamps, so use a transcript with them.

A rubric that stays yours

The AI should score against your standards, not its own. Write a short rubric once, with anchors, and use it every time. An example for a business presentation (illustrative; edit freely):

The last column matters. Marking who judges stops the AI scoring something it cannot see. Give the assistant the rubric and ask it to score only the rows labelled "AI proposes" or "AI counts", leaving the rest blank for you.

What AI cannot judge

A transcript discards most of what makes speaking land. Be explicit with clients about the gaps, and fill them yourself.

• Tone and energy. A flat delivery of a well-structured talk scores well on text and badly in a room.
• Body language and eye contact. Only video and your own eyes show these.
• Audience fit. Whether jargon suits these listeners is a judgement about people, not words.
• Humour and warmth. A joke in text can look like filler, or the reverse.
• Cultural and accent bias. Speech tools can transcribe some accents less accurately. A low-accuracy transcript can make a client look to have more errors than they do; check the transcript against the audio for any client where the text looks off.

The honest line to clients: the numbers describe the words, I describe the performance, and we decide together what to practise.

Main worked example: a manager rehearsing a budget pitch

Say a client, a finance manager, rehearses a ten-minute budget pitch for a board. She records it on her phone and sends the file. You transcribe it with verbatim settings.

The assistant's structure read for Run 1 says the talk spends five minutes on last year's overspend before reaching the request. Your human read adds what the transcript cannot show: she looked at the floor during the overspend section, and her voice dropped at the figures. Your coaching targets for the next session are therefore two: move the request to minute three, and practise the figures standing up, looking at the person she most needs to convince. Notice that the AI produced the data, but the targets came from you.

Run 2's improvement is real, but treat the exact filler count carefully: the tool may have missed some in Run 1 as well. The direction is clear; the precise figures are not.

Further concrete examples of feedback sessions

Example: a conference speaker with an overlong intro

The transcript shows the first 4 minutes of a 20-minute slot are biography and thanks. Prompt: "Quote the first sentence that states what the audience will gain." It appears at 3:50. Target: reopen with the audience benefit and move the biography into a one-line introduction.

Example: a sales lead who hedges

The assistant counts 23 hedges in a 6-minute pitch, clustered in the pricing section ("I think it's around", "sort of in the range of"). Your coaching: rehearse the price sentence until it can be said without a hedge, and provide a script, such as "The programme is [amount] a month, and the first review is in week four."

Example: a nervous first-time presenter

Numbers matter less here. The transcript shows pace falling from 150 to 95 words per minute in the last two minutes. You show her the pace line, and ask what happened at minute eight. She says she lost her place. The fix is a slide-to-notes change, not an AI score.

Example: a team of four rehearsing a joint pitch

Ask the assistant to report each speaker's share of total words and any handover sentences. If one person speaks for 55 per cent of the time in a talk meant to be even, that is a concrete discussion point, as long as the transcript labels speakers correctly. Check the labels before sharing.

A realistic mistake

A coach over-trusts a "confidence score" from an app built for consumers, and tells a client her confidence is 62 out of 100. The client takes the number literally and worries. Prevent it by never sharing a score you cannot explain: report counts, quotes and your own view, and explain every number in a sentence.

Sharing feedback with the client

Send a one-page summary within a day, with the same sections each time so the client can compare runs: the numbers (pace, fillers, hedges), two things that worked with quotes, two targets with a practice method each, and the date of the next recording. AI can draft the structure of the summary from your notes; you write the targets yourself and check each quote against the transcript. For longer-term tracking of session notes, AI session notes and follow-ups for coaches offers a weekly routine, and how business coaches keep the personal touch is useful on tone.

Turning the numbers into a client summary

The summary is where the workflow pays off, because it is what the client keeps. Give the assistant your own notes and the measured figures, and ask it to arrange them, not to invent praise or advice.

Here are my coaching notes and the measured figures for a client's second run of a budget pitch. Write a one-page summary in plain, warm language with these headings: What changed, Two things that worked (with a short quote each), Two targets for next time, Practice between sessions. Use only my notes and the figures. Do not add advice I have not written. Mark any figure that came from automatic counting as approximate.
[notes and figures]

An illustrative extract of what comes back: "What changed: your pitch is two and a half minutes shorter and the request now arrives at 9:30. Filler words fell from about 41 to about 19 (approximate, from automatic counting)." Check each quote against the transcript and each target against your notes. A typical fix is to remove a stray encouragement such as "great energy", which you never wrote and the transcript cannot show.

Before and after: rewriting a hedged opening

Transcripts make rewriting easy because you can lay the client's words beside a stronger version. An illustrative opening from a run-through, as transcribed:

So, um, I think what I wanted to talk about today is, sort of, our invoicing, and basically I guess I'm going to suggest we maybe move to monthly.

Ask the assistant to flag every hedge and filler, then write the stronger version yourself or with the client in the room:

I'm asking you to move invoicing to a monthly cycle. It brings cash in about three weeks sooner, and I'll show the figures.

The second version contains a number you must confirm with the client; never let a rewrite introduce a claim they cannot support. The exercise works because the client hears their own habit in text, then says the cleaner version aloud until it feels natural.

A four-week plan for one client

Notice the focus narrows. Four targets at once is too many; two per session is a good limit, and the transcript lets you check each one with a count or a quote rather than an impression.

Remote sessions and video: what changes

If the client sends video as well as audio, you gain a second source, but the same rule applies: use it for your own observation, not for automated judgement of faces or emotion. Tools that claim to read confidence from facial expressions are measuring something shaky, and a client could reasonably object to being scored on it. Watch the video yourself, note two or three physical habits (swaying, hands in pockets, looking at the screen rather than the camera), and raise them in conversation.

For online talks, add two checks the transcript cannot make: did the client's slides and voice stay in step, and did their audio drop out at any point? A glitch at minute six can look like a pause in the transcript and mislead the analysis, so listen to any unusual gap before you coach on it.

Pricing and packaging, briefly

A transcript-based review can sit as an add-on to a session package: for example a written review within 48 hours of each recorded run. Price it by the time it takes you (preparation 25 minutes, summary 15), not by the tools' cost. Be careful not to suggest that AI does the coaching; clients are paying for your judgement, and the transcript simply gives it evidence.

Checking the workflow works

1. Test your transcription's filler-word recall on a recording with a known count.
2. Run the same talk twice through your analysis prompt and compare answers; large differences mean the prompt is too loose.
3. Ask one client whether the one-page summary changed what they practised.
4. After three sessions, compare two runs and see if the trends match what you heard in the room.

If the numbers and your ear disagree, trust your ear and investigate the audio. If you want help turning this into a routine for your own practice, my AI implementation consultation covers mapping the steps, though most coaches can start with the test in the first item above.

Questions coaches ask before using transcripts

Will transcription tools count filler words correctly?

Not by default. Many tools tidy away "um" and "uh" to produce clean text. Test with a short recording where you know the count, switch on any verbatim option, and treat the figure as a trend between sessions rather than an exact score.

Do I need client consent to record and analyse a talk?

Treat it as essential: get it in writing before the first recording. State what is recorded, which tools process it, how long you keep it and who can see it. Add a line allowing the client to withdraw and have recordings deleted. Recording rules differ between jurisdictions, so take advice if clients are in regulated sectors.

Can AI tell if a client sounds confident?

No. A transcript carries words and timing, not tone, posture, eye contact or energy. AI can flag hedging language and long pauses, but whether the delivery lands with a room is your judgement as the coach.

Is this a replacement for consumer speech-practice apps?

It is a different job. Practice apps give clients solo drills between sessions. Your transcript workflow adds your rubric, your reading of the audience and a plan for the next session, which an app cannot supply.

Further reads

• Are AI Tools Built for Coaches Worth Paying For Over ChatGPT? — Whether paid coaching tools beat a general assistant.
• How to Turn Sales Call Recordings Into Coaching Notes With AI — Same recording-to-notes idea applied to sales calls.
• How Coaches Use AI to Build Workbooks, Exercises and Programmes — Turning feedback into exercises the client can repeat.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — A readiness check before adding AI to a practice.
• How Consultants Use AI to Write Proposals in Under an Hour — Packaging coaching programmes into proposals quickly.
• How to Set Up Human Review for AI Work Without Slowing Down — Keeping a person checking every AI-made summary.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Descript's help article on removing filler words (checked 4 November 2026). Verbatim and filler settings vary by transcription product, so check your own tool's help page; no third-party statistics used.