AI Prompts for Management Consultants: Interviews, Synthesis, Slides

Run six prompts in order: clean and number each transcript, pull themes with verbatim quotes, merge the interviews into one table, build an issue tree from the merged themes, audit every quote, then draft one slide per finding. Quote IDs link each slide claim to a transcript paragraph, which is what makes the deck checkable.

Confidentiality comes first, because transcripts hold what people said about their own employer. The prompts below use one illustrative project throughout: twelve interviews at a 90-person HVAC contractor whose service-agreement renewals are slipping. Every name, figure and quote is invented, and each prompt ends with the checks that stay with the consultant.

Follow me on Instagram@sagnikteaches

Set the confidentiality rules before the first transcript

Interviewees spoke freely on the understanding that you would handle their words with care. Five rules protect that, and they apply before any prompt runs.

Connect on LinkedInSagnik Bhattacharya

1. Contract first. Read the NDA and the engagement terms. Some clients forbid outside AI tools, and if the contract is silent you should get approval in writing.
2. A plan with sensible data terms. Anthropic's privacy page says it does not use inputs or outputs from its commercial products (Team, Enterprise and the API) for training by default, and the other business tiers you are likely to meet, ChatGPT Business and Enterprise, Gemini in Workspace and Microsoft 365 Copilot, are described by their vendors in similar terms. On an individual plan, switch the model-training setting off; labels change, so read the screen. A team plan on ChatGPT Business or Claude Team needs at least two seats.
3. Pseudonymise before upload. Replace names with role codes (SM-1 for a service manager, DS-2 for a dispatcher, CU-3 for a customer, LD-1 for a leader) and the client and its customers with [CLIENT] and [CUSTOMER A]. Keep the key in a separate file that never enters a chat.
4. Narrow the tool. Leave connected apps (formerly connectors) and web search off in these chats unless you need them, switch memory off, and keep one Project per client.
5. Plan the exit. Delete the chats when the project closes, and record the plan, the date and the prompt versions you used. Ask each vendor what happens to a leaver's chats. OpenAI's help centre says that when a ChatGPT Business member is removed, their chats and files are kept indefinitely and the workspace owner cannot see them, while their Projects pass to an owner, so offboarding needs a plan.

For the wider habits around client material, reusing past client work without leaking client data covers what to keep in your own library, and whether solicitors can use ChatGPT without breaching confidentiality shows how another profession reasons about the same duty.

Subscribe on YouTube@sagnikteaches

A quote ID scheme that makes every claim checkable

Traceability is what separates this workflow from pasting transcripts into a chat and asking for a summary. Every quote carries an ID made of the interview number and the paragraph number: I07-P14 means interview 7, paragraph 14 of the cleaned transcript. The IDs are assigned once, in prompt 1, and never regenerated. Every later table, tree branch and slide cites them, so a partner who asks "where does that come from?" gets a paragraph number in seconds.

The cleaned transcript you paste into your own document becomes the master. The audit in prompt 5 compares everything against that master, never against the model's memory of what it read.

Clean the transcript and number every paragraph

Prompt 1 is housekeeping. Auto-generated transcripts mishear product names, run speakers together and bury the useful sentences in filler. The prompt fixes those and does nothing else: no summarising, no tidying of meaning.

Below is the raw transcript of one stakeholder interview, already
pseudonymised. Return a cleaned version.

Rules:
1. Number every paragraph P1, P2, P3 and so on, in order. Each paragraph
   is one speaker turn.
2. Label speakers with the codes already in the text (INT for the
   interviewer, SM-1, DS-2 and so on). Never merge two speakers into one
   turn and never change a speaker code.
3. Fix obvious transcription errors and remove filler words (um, you
   know) only. Do not paraphrase, shorten or reorder anything.
4. Where you cannot tell what was said, write [UNCLEAR] instead of guessing.
5. Keep numbers and dates exactly as spoken.

INTERVIEW ID: I07
TRANSCRIPT: [paste]

Illustrative output for three turns of interview 7:

P12 INT: What happens when a renewal date comes up?

P13 SM-2: Honestly, it depends who is on shift. If the coordinator is in, they ring the customer two weeks before. Otherwise the customer has to chase us.

P14 SM-2: We lost two sites last spring because the first-year visit got missed, and by the time we noticed, the renewal notice was already with a competitor.

The checks that stay with you: compare word counts, because removing fillers trims a little and a big drop means content was cut; read three paragraphs against the audio or the original; confirm the speaker codes are unchanged; and save the cleaned version as that interview's master. Do not ask the model to renumber later, because a second numbering will not match the first. If anyone said something was off the record, mark that paragraph in the master and leave it out of every later prompt, because the model has no way to know the promise was made.

Themes from one interview, with verbatim quotes

Prompt 2 runs one interview at a time. That keeps each answer inside one speaker's evidence and makes every quote easy to check. The rule that earns its keep is that a theme with no verbatim quote does not get listed.

Below is the cleaned transcript of interview I07 (paragraphs numbered).
Extract up to eight themes relevant to this question: "Why are
service-agreement renewals falling?"

Return a table: Theme (eight words or fewer) | Verbatim quote (copied
exactly) | Quote ID (I07-P#) | Speaker code | Type: observed fact,
opinion, or second-hand report.

Rules:
- Copy quotes word for word. If you cannot find a quote that supports a
  theme, do not list the theme.
- One quote supports one theme row. Do not combine paragraphs.
- Do not count, rank or compare with other interviews.

Illustrative output, three of the rows:

Search for every quote in the master transcript, using the find function; this takes a minute per interview and is the single most valuable check in the workflow. If a quote does not match word for word, delete the row. Models tidy quotes silently, dropping a "got" or swapping a word, and a quote shown as verbatim on a slide has to be exactly that. Then read the Type column. A second-hand report ("I heard customers say...") is weaker evidence than something the speaker saw, and it should be treated that way later.

Merge twelve interviews without losing dissent

Prompt 3, the merge, works on the twelve theme tables, not on the raw transcripts. That keeps the input short, keeps every quote ID in play, and stops the model from re-reading and re-interpreting what people said. Its two protective rules are that the count is the number of IDs it lists, and that a theme raised once is marked rather than dropped.

Below are the theme tables from twelve interviews (I01 to I12). Merge
them into one synthesis table.

Columns: Merged theme | Interviews that raised it (IDs) | Count |
Best quote ID per interview | Dissent or qualification (IDs, what they
said, quote ID)

Rules:
1. Merge only themes that mean the same thing. Use the shortest original
   wording as the label.
2. Count is the number of interview IDs you list. Do not estimate.
3. Do not drop a theme because one interview raised it: mark it
   SINGLE SOURCE.
4. Add no information that is not in the tables.

Illustrative output, abridged to the first rows:

Do the counting yourself. Paste the ID lists into a spreadsheet, one ID per cell, and count them there, because models are unreliable at counting and the count will end up in a slide title. Then use your judgement on three things. Did it merge themes that look alike but have different causes? "Visit missed" and "visit delivered but not valued" are close in wording and opposite in remedy. Is the dissent column complete, or did the model smooth over a disagreement between leadership and the front line? And what do you make of the single-source items: a follow-up call, or noise? Count alone is not importance, either. Two leaders can matter more than six dispatchers, and that weighting is yours to apply.

An issue tree that separates findings from hypotheses

An issue tree breaks the client's governing question into branches that do not overlap and together cover it. Prompt 4 drafts one in seconds, which is exactly the risk: a tidy tree can hide the fact that half of it has no evidence. The prompt forces a label on every leaf, so unsupported branches show up as hypotheses to test instead of findings to report.

Governing question: "Why are service-agreement renewals falling, and
what should we change first?"

Using ONLY the synthesis table below, draft an issue tree:
- Three or four first-level branches that do not overlap and together
  cover the question.
- Up to three leaves under each branch.
- For each leaf, write Evidence: the interview count and quote IDs from
  the table, or NO EVIDENCE YET.
- For each NO EVIDENCE YET leaf, name one data source that would test it
  (for example the renewal log).
Do not add causes that are not in the table.

Illustrative output, shown as a table:

The tree is a draft of your thinking, not a result. Three checks stay with you. Are the branches genuinely separate and complete, or is this just the theme list re-sorted under headings? Which branch answers "what should we change first"? That depends on impact and cost, and the interview count does not tell you either. And keep the hypotheses off the findings slides. Interviews describe what people believe happens; a cause needs data such as the renewal log. The prompt's last rule, naming a data source for each unsupported leaf, is the start of your next work plan.

One more judgement the model cannot make is who is speaking. A leader describing the renewal process is describing intent, and a dispatcher describing last Tuesday is describing practice. When the two disagree, as I11 and I02 do in the table above, the tree records both and you decide which to test first. The renewal log usually settles it, and settling it is worth more to the client than a longer list of themes.

Audit every quote before it goes near a slide

Prompt 5 checks the citations. By this point the synthesis and the tree cite dozens of quote IDs, and a single wrong one can undermine a slide. The audit prompt takes the cleaned transcripts and the finished tables and checks each citation. Run it in a fresh chat, with both pasted in, so the model is not marking work it produced a few minutes earlier in the same conversation.

Below are (A) the cleaned transcripts, numbered by paragraph, and
(B) a synthesis table and issue tree that cite quote IDs.
For every quote ID cited in B, return:
- The paragraph text from A, copied exactly
- Match: EXACT, PARAPHRASE, WRONG SPEAKER or NOT FOUND
- If PARAPHRASE: the words that differ
Then recount every "n of 12" in B from the interview IDs listed and say
whether each matches. Do not rewrite anything.

Illustrative output:

The audit is another model pass, so it can be wrong too. The rule I would use: read the paragraph yourself for every quote that will appear on a slide, and spot-check about one in ten of the rest. When something fails, fix the master tables, not only the slide, or the error returns the next time the tables are used. This is also why the IDs were assigned once: if numbering had drifted between prompts, an audit like this would be impossible.

One finding, one slide

Prompt 6 drafts slides one finding at a time, from verified material only. The prompt asks for an action title, which is a full sentence that states what the audience should take away, and then forces the supporting detail to carry its own evidence. It also bans the wording that quietly overstates a small set of interviews.

Draft one slide from the finding below. Use only the evidence given.

FINDING: First-year service visit missed
EVIDENCE: 7 of 12 interviews (I02, I03, I04, I07, I08, I09, I12);
verified quotes I07-P14, I03-P10; qualification I06-P17

Return:
1. Action title: one sentence, under 20 words, stating the so-what.
2. Three supporting bullets, each ending with its count or quote ID.
3. One verbatim quote for a pull-out, attributed by role code only.
4. Source line: "Source: 12 stakeholder interviews, [dates]".
5. Speaker notes: the caveats, the qualification, and the data that
   would confirm this finding.
Say "interviewees". Never say "customers" or "most".

Illustrative output:

Title: Seven of 12 interviewees say a missed first-year service visit puts the renewal at risk

Bullets: Seven interviewees across service, dispatch and leadership roles report first-year visits missed or late (I02, I03, I04, I07, I08, I09, I12). Two service managers link a missed visit directly to a lost site (I07-P14, I03-P10). One customer says the visit happened on time but did not explain what the agreement covers (I06-P17).

Pull-out quote: "We lost two sites last spring because the first-year visit got missed" (SM-2).

Source: 12 stakeholder interviews, [dates]; counts of interviewees, not survey percentages.

Speaker notes: These are interviewees' accounts, not measured data. Test against the renewal log: the share of sites with no first-year visit recorded before the renewal notice. Qualification: I06-P17.

Compare that title with the one a tired draft produces: "Customer feedback summary". The first tells the audience what to conclude and carries its own count. The checks stay with you. Does every number trace to the synthesis? Does the title claim a cause or a prevalence the evidence cannot bear? Is the quote verbatim and attributed by role only? And does the slide hold one finding, with the caveat in the notes where a presenter will see it? Once the text is right, which tool lays it out is a separate decision, covered in Gamma, ChatGPT and Claude for slide decks.

Three ways the pipeline goes wrong

The failures below are the ones to expect, described as illustrations.

• Merged speakers. A raw transcript runs a service manager's and a dispatcher's words together, the cleaning prompt keeps them in one paragraph, and the quote is later credited to the wrong role. The audit flags it as WRONG SPEAKER, and the fix goes into the master transcript.
• Wrong counts. A draft title says "9 of 12" while the synthesis lists eight interview IDs. A client who knows the interviewees will spot the discrepancy, and the recount line in prompt 5 is there for exactly that.
• The majority illusion. Three of the twelve interviews were with customers, and the draft says "customers say". Three conversations cannot speak for a customer base, and the ban on "customers" and "most" in prompt 6 exists to stop it.

None of these is exotic. All three produce text that reads well, which is why the checks cannot be optional.

Keeping the prompts, the masters and an audit trail

Models do not give the same answer twice, so a saved prompt alone cannot reproduce a result. Keep the outputs as well: the cleaned transcripts, the theme tables, the synthesis, the tree, the audit results and the slide drafts, each in a folder named for the project. Number the prompt versions (prompt 3, version 2) and write down the plan, the model name shown in the tool and the date of each run. If a partner or the client later asks how a number got onto a slide, you can walk back through the audit to the paragraph, the interview and the version of the prompt that produced it. Hold the pack in a Project (ChatGPT Projects and Claude Projects both keep instructions and files together) so the rules travel with the work and a colleague starts from the same version.

Hours saved across twelve interviews

These timings are illustrative, for twelve interviews of about 45 minutes. Time your own first.

A saving of about 16 hours is plausible, and a good part of the 13 that remain is checking, which is what lets you put your name on the deck. For the same job seen as a process, synthesising stakeholder interviews in management consulting goes through the stages, and the best AI for consulting work, from interview notes to findings deck compares the tools. Neither replaces the habit this tutorial rests on: nothing reaches a slide unless you can point to the paragraph it came from.

Consultants' questions about AI, interviews and confidentiality

Can I put client interview transcripts into ChatGPT or Claude?

Only when three things hold: the client's contract allows outside AI tools, you are on a plan whose terms do not use your content for training, and you have replaced names with role codes first. If you cannot show all three, do not paste the transcript, however convenient it would be.

How many interviews can one prompt handle?

Treat the context window as a limit on trust, not only on size. Run one transcript at a time for extraction, then merge the extracted theme tables rather than the raw text. Long, mixed inputs are where quotes drift from what was said and counts stop matching the evidence.

Should the slides say customers said it after twelve interviews?

Say interviewees, and give the count. Twelve conversations show what was said, not how common it is across all customers. Avoid percentages, and if a slide's claim matters, test the pattern against data such as renewal logs before the deck goes to the client.

Do I tell the client which steps used AI?

Check your contract first, and say so if it asks about tools or if their material passes through a third-party service. A short line in the method note, naming the steps that used AI and the checks that were human, often settles the question before it is asked.

Further reads

• Best AI Tools for Management Consultants: Seven From Research to Decks — Which tools suit research, synthesis and deck building.
• Best AI Model for Business Consulting: Pick by Task, Not by Hype — Choosing a model by task rather than by hype.
• How Consultants Use AI for Client Research Before Discovery Calls — Use the same facts-first habit before the first client call.
• How Consultants Use AI to Write Proposals in Under an Hour — Win the work these interviews belong to.
• How to Summarise Bundles and Transcripts With AI Safely — Safer ways to summarise long transcripts and bundles.
• 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.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic's privacy centre on commercial data use; OpenAI, Google Workspace and Microsoft privacy pages (business-plan defaults); OpenAI's help article on data retention when a member is removed; ChatGPT Business and Claude pricing pages (seat minimums); OpenAI and Claude help pages on Projects.