Best AI for Consulting Work: From Interview Notes to Findings Deck

A consultant can turn client interview notes into a findings deck by getting consent, transcribing each interview, asking AI to extract structured summaries, grouping them into themes in an evidence table where every theme shows who said it and a verified quote, then drafting one slide per finding from that table. Check every quote before it goes on a slide.

The order matters more than the tool. Consultants who start by asking an assistant for "the key findings" from a pile of notes get fluent themes with no evidence behind them. Building the evidence table first means each finding on a slide can be traced to people and words, which is what a client's sceptical director will ask for.

Follow me on Instagram@sagnikteaches

Step 1: consent and confidentiality before the first recording

Interviews are people talking candidly about their employer, so the consent step carries more weight than any later technique. Before recording, say three things and write down that you said them: that the interview will be recorded and transcribed, which tools will process the audio (naming the transcription service and any AI assistant), and who will see the transcript and in what form (for example, only the consulting team; quotes unattributed). Recording consent rules differ between places, so follow your engagement terms and the law that applies to the people you are speaking to, and do not assume one rule covers every call.

Connect on LinkedInSagnik Bhattacharya

An illustrative opening line: "With your permission I'd like to record this so I can listen properly instead of typing. The recording is transcribed by a service we use, summarised with an AI tool, and only my team will see it. Anything in the report will be unattributed unless you say otherwise. Is that all right?" Keep the answer in the interview log.

Subscribe on YouTube@sagnikteaches

Next, decide where the audio and transcripts live. A shared drive with restricted access is a better home than the consumer app on someone's phone. If the client's data protection terms restrict third-party tools, ask for written approval of the specific tools, or transcribe on a plan whose terms exclude training on your content. If an interviewee asks for no recording, take notes by hand and treat those notes as the transcript, with a lighter summary step.

Step 2: get a transcript you can trust

Use a transcription service or the transcript feature of your meeting tool. List prices on the vendors' pages in November 2026: Otter Pro $8.33 a user a month on annual billing or $16.99 monthly, with 1,200 transcription minutes a user a month and a 90-minute limit per conversation; Fireflies Pro $10 a seat a month on annual billing or $18 monthly. Otter's page sometimes shows regional promotions on top, so compare the list price before buying. For twelve interviews of 50 minutes each, that is 600 minutes, inside the Otter Pro monthly allowance, and each 50-minute interview sits under the 90-minute cap; a two-hour workshop would not.

Whichever you use, run a quality check on the first transcript before trusting the rest: listen to five minutes of the recording while reading the text, and mark errors in names, numbers and technical terms. Fix a glossary of the client's terms (product names, acronyms, people) and add it to the tool's custom vocabulary if it has one, or put it at the top of each summarising prompt. A realistic failure: the transcript renders a product code as a common word, and the finding that "customers prefer the X-200" becomes "customers prefer the next 200". Spot-check any number, any product name and any person named, since those are where speech-to-text slips and where a deck embarrasses you.

For when a human transcript is worth paying for, AI against human transcription sets out the trade-off.

Step 3: one structured summary per interview

Process interviews one at a time, using the same prompt each time so the outputs line up. The structure is the point.

Below is a transcript of one interview for a consulting project on
[topic]. The interviewee's role is [role]. Using only what is in the
transcript, produce:
1. Three to five main points, each in one sentence.
2. For each point, the best supporting quote, copied exactly, with the
   approximate position in the transcript (minute or paragraph).
3. Anything the person said that contradicts another point they made.
4. Things they said they do not know or that someone else would know.
5. Any number or date they mentioned, exactly as said.
If something is not in the transcript, write "not stated". Do not infer.

[paste transcript and glossary]

An illustrative output for one interview with a store manager in a retail client: point 1, "Stock arrives in the afternoon so shelves are empty during the morning rush"; supporting quote "we get the delivery at two, the rush is at nine, so we are always running half empty"; point 2, "Staff do not trust the stock figures on the system"; numbers: "twelve staff", "three deliveries a week". What you fix: the position markers (models often estimate them badly, so search the transcript for the quote text), and any point that paraphrases too loosely. The "not stated" instruction is what makes the output usable, since otherwise models tend to fill the blanks.

Save each summary in a folder, named by interview ID, not by name. Use "INT-04, store manager" rather than a person's name, and keep the key to names in a separate document with restricted access.

Step 4: themes and the evidence table

With twelve summaries, ask the assistant to propose themes. Give it the structured summaries, not the raw transcripts, so it works from the exact quotes you have already pulled.

Here are 12 structured interview summaries (INT-01 to INT-12). Propose
6 to 8 themes. For each theme: a one-line statement, the interview IDs
that support it, one verbatim quote from each supporting interview
(copied from the summaries), and any interview that disagrees. Do not
create a theme supported by fewer than three interviews; list those
separately as "single-voice points".

Turn the output into an evidence table:

The rows and counts here are illustrative. The "Strength" column is a simple rule you set before you look: for example, Strong is half the interviewees or more, Moderate is a third to a half, and anything under three is a single-voice point. Clients like the rule because it shows how you weighted evidence, and you like it because it stops the loudest anecdote becoming a finding. Remember that counts describe the people you interviewed, not the client's whole workforce, so word findings accordingly ("six of the twelve managers interviewed said...").

Now check the table. Open each "best quote" in its transcript, confirm it is verbatim and in context, and look for the quote's neighbours. A realistic miss: a theme's quote begins "We sometimes struggle with deliveries", but the next sentence in the transcript is "though it's better since March", which changes the weight. This check is the central quality step in the whole workflow, and it is what keeps the deck honest. For prompts aimed at consulting interviews specifically, AI prompts for management consultants: interviews and synthesis goes further.

Step 5: from evidence table to slide draft

A findings deck gets one slide per finding, each with an action title, the evidence and a "so what". Draft them from the table, not from memory.

Using only the evidence table below, draft one slide per theme rated
Strong or Moderate. For each slide give: an action title under 15 words
stating the finding; up to three bullets of evidence with interview
counts ("6 of 12 managers"); the best quote with its interview ID; and a
"so what" line suggesting the implication for the client, marked as
"for discussion". Do not add any fact that is not in the table.

[paste evidence table]

Illustrative slide draft: title "Stock arrives after the morning rush in half of the stores interviewed"; bullets "6 of 12 managers described afternoon deliveries"; "4 of those said shelves are half empty at opening"; quote "We get the delivery at two, the rush is at nine, so we are always running half empty."; so what, for discussion: "Test an earlier delivery window at two pilot stores". What you fix: change "half of the stores" to the wording the numbers support, remove any adjectives, and replace the so-what with the recommendation you actually hold. The slide wording from the assistant is a scaffold. The findings are yours.

Then build the deck in your usual tool. If you want to compare tools for that stage, see Gamma, Copilot and ChatGPT for consulting decks. Add a backup slide for each finding with the full set of supporting quotes; it is where the client's questions will land.

When you only have typed notes, not recordings

Not every interview is recorded. Some interviewees refuse, some rooms are not suitable, and some clients bar recording outright. The same routine works with notes, with one change: the quality of the structured summary depends on how good your notes are, so write them to be machine-readable. Use short lines, one point per line, mark direct quotes with quotation marks only when you are sure they are word for word, and write numbers in full as said. After the interview, spend ten minutes expanding shorthand while it is fresh. Tell the assistant in the prompt that the source is the consultant's own notes, not a transcript, so that "not stated" covers what you did not write down and quotes are labelled "from notes, not verbatim". In the deck those lines should read as paraphrase.

A small before-and-after of a note. Raw: "del 2pm, rush 9, half empty, sys wrong". Expanded for the prompt: "Delivery arrives at 2pm; the morning rush is at 9am; the manager said shelves are half empty by midday; the manager said the system's stock figures are wrong. Not verbatim." The second can be summarised without invention, and the first cannot.

Checking one quote, end to end

The verification step is easier to describe with a worked case. The evidence table says the best quote for a theme is "We always run half empty on Saturdays" (INT-06). Open the transcript, search for "half empty", and read a minute either side. Three outcomes are common. The quote is exact and in context: keep it. The quote is close but the speaker said "we sometimes run half empty on Saturdays": change the table, because "always" and "sometimes" are different findings, and a manager who reads the deck will notice. Or the phrase appears only in the assistant's summary and nowhere in the transcript: drop the quote and look for a real one, and note the interview ID in the log as a place where the summary prompt drifted. Make this a ritual: for every quote that will appear on a slide, put a tick and the transcript position in the table. It takes about two minutes a quote, and a twelve-slide deck with fifteen quotes needs half an hour.

Contradictions, outliers and what not to smooth over

Models like harmony: asked for themes, they tend to produce tidy agreement and bury the person who disagreed. Ask explicitly in the theme prompt for disagreement, as above, and then read the dissent yourself. In consulting work the outlier is often the most valuable finding. Two stores out of twelve with a different returns routine may be the best practice the client wants to spread, or the source of the problem. Keep single-voice points in a short list in the deck's appendix, with a line on whether to follow them up, and ask the partner whether any deserves a second interview.

Contradiction inside one interview is also information. A manager who says the system is reliable and then describes counting shelves by hand has told you about their trust in practice, which is more useful than a one-line summary. The structured summary prompt's third item (contradictions) is there to catch exactly that. Quote both statements side by side on a backup slide, with the minute markers, if the point matters.

A second illustration: five senior interviews

The routine scales down. A consultant for a sports club interviews five senior people (the chair, the treasurer, two coaches and the volunteer coordinator) for 45 minutes each on why membership renewals dropped. Five interviews are too few for percentages, so the evidence table changes: instead of "6 of 12" the strength column says "all five", "three of five" or "one". The deck states plainly that it reflects five senior voices and no members. Findings are framed as hypotheses for the next stage, with a recommendation to survey members. The assistant helps the same way (structured summaries, theme table with quotes, slide scaffold), and the total effort is about eight hours instead of eighteen. The point is that the base is stated on every slide and the conclusions are sized to the evidence, which matters more than the number of interviews.

Presenting anonymity honestly

If you promised anonymity, check each quote for identifying detail, not just for names. Job titles, locations within the business, unusual events and turns of phrase can all identify a person. Replace specifics with generic descriptors ("a store manager", "a senior coach") and take out details that point to one person, and tell the interviewee in advance if that is the rule. When you must change wording, mark the quote as lightly edited in the notes and keep the original in the log. A safe standard is to ask whether the client's HR lead could guess the speaker from the quote alone. If so, rework it or drop it.

A worked example: twelve store interviews, from audio to deck

An illustration with invented figures, using a pet-shop chain as the client. The consultant interviews twelve store managers for 50 minutes each, aiming at a deck on why stock availability drops at weekends.

The AI column adds up to 16 hours (1 + 3 + 4 + 4 + 4), against 39 for the manual route (1 + 10 + 12 + 8 + 8). Treat the exact figures as an example of the shape, not a benchmark: the saving comes from stages that were mechanical, while the consent step and quote checking stay human. At an internal rate of $120 an hour, 23 hours saved is $2,760 of capacity, against a tool cost of perhaps $20 to $50 for the month. The risk is not the cost; it is skipping the checking time to keep the saving.

Where this goes wrong, and how it shows up

• Merged voices. Pasting several transcripts into one prompt can attribute one person's quote to another. Process separately and use IDs.
• Tidy quotes. The assistant "cleans up" a quote and changes its meaning or removes a hedge. Always copy the exact text and verify against the transcript.
• Counting the wrong thing. "Most people said" when most of the interviewed people said. State the base.
• A theme the model liked. A neat theme with two supporting interviews and four adjectives. The three-interview rule catches it.
• Over-identifying. A quote about "the only night manager in the east store" identifies its speaker even without a name. Check each quote for identifying detail against what you promised.
• Silent drift in the deck. The quote is edited for length on a slide and no longer matches. Mark any trim with an ellipsis, and re-check before sending.

Tools and a sensible setup

A lean setup is a transcription service you have checked for data terms, one paid general assistant (the individual paid plans of ChatGPT, Claude and Gemini sit around $20 a month; Claude Pro is $17 a month on annual billing), a spreadsheet for the interview log and evidence table, and your usual deck tool. A team plan makes sense when several consultants share prompts and projects: Claude Team lists a standard seat at $20 annual or $25 monthly. Keep the glossary, the prompts and the evidence-table template in one project folder so each new engagement starts from them.

For related reading on interviews outside consulting, writing case studies from customer interviews uses a similar transcript-to-structure approach, and analysing staff surveys and exit interviews covers the sensitivity of people-related data. If you would like to build the routine on your own interviews with someone, that is what my AI implementation consultation covers.

Questions consultants ask about interview synthesis with AI

Do I need consent to record and transcribe interviews?

Yes, ask every interviewee before recording and say which tools will process the audio and who will see the transcript. Recording consent rules differ between places, so follow your engagement terms and local law, and record the consent itself in the interview log.

Can AI find the themes for me?

It can propose themes and group quotes under them quickly. You decide which themes are real by checking that each is supported by several interviewees, and by reading the quotes it chose in context rather than trusting the grouping.

How many interviews can one prompt handle?

Process each interview separately first, then combine the structured summaries in a second prompt. Pasting twelve full transcripts into one prompt invites missed detail, mixed-up attribution and quotes that quietly change between the transcript and the output.

How should I quote people in the deck?

Use verbatim quotes only after checking them against the transcript or recording, and anonymise according to what you promised. An AI-polished quote is a paraphrase, and should not be presented as the person's words.

Further reads

• How to Summarise Bundles and Transcripts With AI Safely — Summarise long transcripts safely.
• How to Review AI Call Transcripts for Quality and Compliance — Review transcripts for quality and compliance.
• How to Build Interview Questions and Scorecards With AI — Build interview questions and scorecards.
• How Consultants Use AI for Client Research Before Discovery Calls — Prepare for interviews with client research.
• Best AI for Consulting Research: Market Scans With Source Checks — Add desk research with a source log.
• AI Tools for Strategy Consultants: Desk Research and Market Sizing — Add market numbers to the story.
• Best AI Tools for Management Consultants: Seven From Research to Decks — Seven AI tools for independent consultants from research to decks, with a confidentiality tier rule, costs and a worked two-week diagnostic.
• Gamma AI Presentation vs Copilot in PowerPoint for Client Decks: Costs and Fit — A repeatable test for Gamma against Copilot in PowerPoint on one client deck: the brief, five edits, a weighted scorecard, export checks and what each costs.
• AI Brand Audit: A Template Branding Studios Can Sell — A ten-day brand audit for a music school, step by step, with the 12-page client deliverable a branding studio can reuse and sell.
• AI Brand Strategy: Positioning Workshops and Messaging for Agencies — A care-agency strategy project in stages: twelve stakeholder interviews, three scored positioning options, a message house and the client workshop.
• AI Consulting Report Writing: From Findings to a Client-Ready Draft — Turn a findings log into a client-ready report with evidence tags, a number check on every figure and an executive summary written last.
• Using AI in Management Consulting: Synthesise Stakeholder Interviews — A practical interview-to-theme workflow with a coded example, evidence table and contradiction check.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: otter.ai/pricing, fireflies.ai/pricing and claude.com/pricing, all read 4 November 2026.