Using AI in Management Consulting: Synthesise Stakeholder Interviews

A management consultant can use AI to turn consented interview transcripts into provisional codes, group those codes into themes and attach each claim to exact source passages. Check speaker labels and quotations against raw notes, record dissenting accounts, and keep the number and role of supporting interviews visible. The consultant decides which findings the evidence actually supports.

The first useful output is an evidence table, not a persuasive slide. A model can make six inconsistent accounts sound like one coherent story. The table should retain interview ID, role, passage location, verbatim wording, code, proposed theme and reviewer decision. That structure lets a colleague challenge a theme before it hardens into a recommendation.

Follow me on Instagram@sagnikteaches

Frame the question before collecting accounts

Write the decision the engagement needs to inform. For example, a client may want to know why order approvals take longer than the agreed service target. That is narrower than asking AI to "summarise stakeholder sentiment". Set the unit of analysis: one interview, one incident, one process step or one hand-off. The choice affects how you count evidence. If one manager repeats the same complaint seven times, that remains one stakeholder account, not seven independent confirmations.

Connect on LinkedInSagnik Bhattacharya

Draft an interview guide with a common core and room for follow-up. Ask for the last concrete instance, what happened before and after, which system or person was involved, and what record could corroborate the account. Avoid leading prompts such as "How much does the approval bottleneck frustrate you?" if the bottleneck itself is unproven. A neutral version asks where approval work waits and what evidence shows the wait. Use the broader consulting prompt guide for prompt design; the workflow here concentrates on evidence after interviews are collected.

Subscribe on YouTube@sagnikteaches

An illustrative engagement starts with twelve interviews across sales, operations, finance and service. The sponsor believes approvals are slow because finance is understaffed. That belief becomes a hypothesis to test, not the theme heading. The consultant plans to compare statements about queue times, missing information, authority limits and exceptions with actual ticket timestamps. If interviewees disagree, the disagreement may reveal a split between teams rather than poor data quality.

Agree the recording and quotation rules

Before recording, check the engagement contract, client policy and participant notice. State whether audio will be recorded, who will see transcripts, whether a third-party processor is involved, how long working copies will be kept and whether named quotations may appear in deliverables. The right handling for an interview with a junior employee may differ from a public executive interview. When consent or policy does not permit recording, take structured notes and mark them as notes rather than a verbatim transcript.

Microsoft's Teams support page says meeting transcripts can show speaker names and time stamps, and participants see a notice when transcription starts. By default, organisers and co-organisers can download a transcript as a .docx or .vtt file; whether anyone else can depends on admin and organiser settings. Those product facts are useful only if the client permits transcription and the meeting is configured accordingly. Do not assume that all attendees can download the text or that speaker attribution is correct. Test the client's exact meeting setup before relying on it.

One illustrative interviewee agrees to private analysis but not a direct quote in the board deck. The analyst can code their account internally, then present a de-identified paraphrase if the contract allows. Another interviewee may consent to recording but ask to pause for a sensitive example. Stop recording for that portion and note the boundary. A transcript that is technically available is not automatically authorised for every downstream use, including an external AI tool.

For a tool such as ATLAS.ti, check data terms before upload. Its AI Coding manual says the feature uploads document content to ATLAS.ti and OpenAI servers with explicit consent, then proposes codes for review. That is a specific processing path to evaluate against the engagement terms. A consultant can also use an approved general assistant or work manually; the choice should follow the client's confidentiality requirements and the team's ability to audit output.

Turn recordings into reviewable source material

Assign each interview a stable ID such as INT-01 and keep the mapping to names in a restricted file. Record date, role, interviewer, consent status, file version and whether the source is audio, a transcript or notes. Keep line numbers or time stamps. Remove or mask identifiers that are irrelevant to analysis, but preserve the context needed to understand a statement. A sentence about "the system" is useless if the transcript no longer indicates which step or application the speaker meant.

Listen to a sample of every machine transcript, then review all passages likely to enter a finding. Check terminology, numbers, negations and speaker turns. A transcript can turn "we do not approve on Friday" into "we approve on Friday"; a model will build an elegant but false pattern on top. Correct the source or append a correction note with the audio time. Keep a clean copy for coding and the original for audit. The transcript review tutorial offers a fuller quality check if audio is difficult.

An illustrative transcript fragment says, "The request sat with procurement for two days." On the recording, the speaker says it sat with project control, not procurement. The distinction changes the process owner. The analyst corrects the transcript, records the time stamp and marks the first AI code as invalid. The model's plausible wording cannot outrank the recording. If no audio exists, the consultant should describe the source as an interviewer note and avoid presenting its wording as a verified quotation.

Format the clean text so each question and answer has a clear boundary. ATLAS.ti's manual notes that its AI Coding works by paragraph and that clear paragraph structure helps with interviews. That advice generalises: separate interviewer questions from participant answers, because a recurring question can otherwise masquerade as a recurring participant theme. Preserve any pauses, uncertainty or corrections that materially change the meaning. You may remove filler words for readability in an internal paraphrase, but never call the result verbatim.

Define codes that a second analyst can apply

A code is a short label attached to a passage, such as "missing request details" or "approval authority unclear". A theme is a wider interpretation supported by several coded passages. Begin with a small provisional codebook linked to the engagement question. For each code, write an inclusion rule, an exclusion rule and an example. Allow new codes when the interviews reveal something the guide did not anticipate. Do not ask AI for twenty abstract topics and treat the list as findings.

These are illustrative codes, not a claim about a real client. Give AI only the permitted transcripts and the code definitions. Require it to quote a passage and location for every proposed code, to say "no match" when appropriate, and to flag ambiguity. An analyst should review the suggestions and record changes. If the model labels every complaint "queue wait", the codebook may be too broad or the model may have ignored the exclusion rule. Check both before changing the story.

Use only the supplied interview text and the four code definitions.
For each relevant passage return interview ID, speaker role,
line or time stamp, exact excerpt, proposed code and reason.
Keep interviewer questions out of the evidence. Do not infer a cause
from tone. Mark uncertain passages for human review. List passages
that contradict an emerging pattern. Do not write recommendations.

An illustrative model output codes "We waited because the form lacked a project number" as "queue wait". The analyst changes it to "incomplete request" because the stated delay was rework before submission, not a completed request awaiting approval. That correction matters: a recommendation to hire approvers would target the wrong mechanism. Record the rejected AI label and reason so another reviewer can see why the final code differs from the first pass.

Build themes with excerpts and counterexamples

Group codes only after reading the underlying passages. A theme should answer a question relevant to the engagement, describe a mechanism and indicate its limits. "People dislike the process" is too vague. "Requests often arrive without a project number, so teams return them before approval" is testable against interviews and ticket records. Distinguish what interviewees experienced from a causal explanation the consultant proposes. The second requires corroboration.

The following table is synthetic practice data. Its short quotation cells are fabricated solely to demonstrate the format; they are not real client statements or published evidence. In a live engagement each quote must match the authorised transcript exactly, with an internal reference that the reviewer can open.

The theme table does not need to be pretty. It needs a claim, source, role, passage location, evidence type and status. Add a column for permission to use a direct quote externally if the deliverable includes one. If several excerpts say essentially the same thing, keep enough to show role diversity and context; a long quote collage is not a substitute for analysis. If one person supplies most of the evidence, say so. The interview-notes-to-deck tutorial can help with presentation after the evidence has been tested.

An illustrative contradiction deserves its own row. Eight people say approvals stall in finance, while two finance interviewees say requests sit with operations until required documents arrive. Rather than averaging the accounts into "cross-functional delays", ask for time stamps for a small set of requests. The difference may be a hand-off definition: each group starts its clock at a different point. A useful finding would identify the unowned interval, with documents to support it. If the logs are unavailable, report the competing accounts and the limit.

Work a twelve-interview synthesis from start to finish

Imagine a consulting team engaged to understand a slow approval process. It has twelve interviews: three in sales, four in operations, three in finance and two in service. The client gives permission for restricted analysis of transcripts, but only de-identified excerpts may appear in the final report. The consultant also receives a sample of twenty-five approval tickets and the written policy. These figures are illustrative; no real client or study is being described.

The analyst first checks twelve transcript headers against the interview schedule and samples the audio. Two speaker labels are wrong, and one interview is notes only because recording was declined. The analyst fixes the labels and marks the notes-only source. Next, AI suggests 38 passage-level codes. Human review accepts 27, edits seven and rejects four. The team groups the accepted passages under missing information, hand-off ownership and escalation uncertainty, keeping the original passage IDs. Counting suggestions is a way to track review workload, not a measure of truth.

The first theme draft says finance is the main bottleneck because seven interviews mention finance. The evidence table shows that four of those seven references describe work before a complete request reaches finance. The twenty-five ticket sample has nine returned for missing information, six waiting for a named approver, four paused for client input and six completed without a recorded hold. These simple counts are not a population estimate. They do show that the initial "finance bottleneck" label hides several mechanisms. The consultant revises the theme to separate incomplete submissions from decision queue time.

The team then reads every passage behind the three proposed themes, including the two interviews that report no delay. One operations team uses a prefilled form and rarely returns requests. That is a useful counterexample: it may point to a process difference worth testing. The consultant asks to inspect that form and compare its required fields with the standard template. AI can generate the comparison checklist, but a person checks the actual files. The hypothesis is that clearer intake reduces rework; the interviews alone do not establish the size of any saving.

The final client discussion contains three provisional findings, each with the number of supporting interviews, roles represented, corroborating ticket examples, contrary accounts and an open question. The consultant avoids a claim such as "58% of staff experience the issue" because seven of twelve purposively selected interviewees is not a representative staff sample. They show one permitted, checked excerpt for each theme and keep the detailed evidence ledger internally. Before recommending a workflow change, they confirm that the process owner agrees with the documented hand-offs and that the ticket sample was interpreted correctly.

Budget the review explicitly. In this illustrative engagement, the team spends 35 minutes per interview on transcription and source checks, 20 minutes on code review, then three hours comparing themes with tickets and policy. Those are planning estimates for the example, not universal benchmarks. A client should compare them with a manual baseline on its own material. If AI saves first-pass sorting but creates hours of quote correction, the process needs a different tool, better audio or a narrower prompt.

Keep counts honest and attribution safe

For every theme, state the denominator and what was counted. "Six of twelve interviews mentioned missing details" means something different from "six of twelve employees have this problem". A stakeholder may have been asked directly, volunteered the issue, or referred to someone else's experience. Record those distinctions. You can count passages for workload management, but use unique interviewees and roles when describing breadth. A forceful quote from one senior person should not silently outweigh independent accounts from several frontline staff.

Consider an illustrative split: four of five managers say the process is clear, while five of seven individual contributors describe unclear exceptions. The simple total says five of twelve raised a concern, which hides the role pattern. The consultant should present the split, show the exact exception that caused confusion, and test it against the policy. It may be a communication gap, a policy gap or a sampling effect. AI is good at grouping similar language; the analyst must decide whether groups are comparable.

Do not publish identifying details merely because names were removed. A unique job title, project description or incident date can reveal a speaker in a small team. Replace those details in client-facing excerpts only with the speaker's permission and clearly indicate edits where the contract requires it. Never invent cleaner wording and place it inside quotation marks. If a statement must be shortened, verify that the remaining words keep its meaning and retain the full passage in the internal ledger.

A quick failure test: ask a reviewer who did not run the AI pass to locate three random quotations in the raw source and explain why each supports its theme. If one cannot be found, halt the quotation export and audit the rest. If it can be found but the surrounding sentence reverses its meaning, revise the theme and examine other excerpts from that interview. The test is more revealing than asking the model whether its own citations are accurate.

Challenge the synthesis before it becomes a recommendation

Use a separate challenge pass. Ask what evidence would disprove each theme, whose perspective is absent, whether the interview guide pushed people towards the same answer and whether a document or metric could resolve the ambiguity. A model can propose challenge questions, but the consultant must answer them with sources. The research-synthesis tutorial for design teams gives another perspective on preserving participant differences across qualitative work.

For example, AI may propose "staff resist the new system" because several interviewees mention the old spreadsheet. Read their passages: they may actually prefer the new system but keep the spreadsheet to track requests that disappear between two queues. The correct investigation is about visibility at the hand-off, not resistance to change. A prompt that searches only for objections can manufacture a change-management narrative. A neutral challenge asks for passages showing both adoption and workarounds.

Make the final review a short gate: every claim has at least one source; every direct quote is checked; contradictory evidence is disclosed; counts use the stated denominator; and recommendations are separated from observations. If a finding rests on one interview, call it a hypothesis and specify the next test. If a finding conflicts with ticket data, show the conflict rather than smoothing it away. Your client can then decide what further evidence is worth buying or collecting.

Keep the evidence ledger with the engagement record for the agreed retention period and remove working copies according to the client agreement. Document which tool produced the first pass and which consultant approved the final theme. The successful output is a set of claims another consultant can retrace from the deck to an authorised excerpt and then to the original source. That traceability makes AI's speed useful without giving it the authority to speak for stakeholders.

Questions about interview evidence

How many interviews make a theme credible?

There is no universal count. Record how many people mentioned the issue, which roles they represent, the context in which it arose and any contrary accounts. A theme mentioned by one person may still be material if that person owns the process or has direct evidence. Do not equate frequency with importance or claim that a small interview set represents the entire organisation.

Can AI-generated quotations go into a client deck?

Only use wording that a reviewer has checked against the original transcript or notes and that the confidentiality terms permit you to share. AI may compress several statements into a sentence that nobody actually said. If you cannot verify the words, label the text as a paraphrase or remove it. Keep an internal reference to the interview and passage.

Should stakeholders see the themes before the final report?

A short validation discussion can reveal misread process details and missing exceptions, provided the engagement allows it. Show provisional themes and the questions they raise rather than presenting them as settled conclusions. Record corrections and disagreements. The consultant remains responsible for deciding whether the evidence supports a finding and recommendation.

Further reads

• AI vs Human Transcription: Which Is Worth Paying For? — Choose a transcription method for difficult recordings.
• How to Turn Meeting Notes Into Tasks Automatically With AI — Turn validated meeting decisions into assigned actions.
• How to Check Sources and Citations in AI Research — Strengthen source checks beyond interview evidence.
• AI Workshop Agenda and Facilitation Notes for Consultants — Test emerging themes in a facilitated session.
• ChatGPT Prompts for Strategy Consultants: Issue Trees and Hypotheses — Turn supported themes into testable questions.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support, Start, stop, and download live transcripts in Microsoft Teams meetings; ATLAS.ti 26 Mac user manual, AI Coding; vendor pages checked 4 November 2026.