Synthesize Interviews and Workshops into Findings That Cite Their Sources

Workflow area: Elicitation & Discovery · Tier: Foundation · Hands-on time: ~25 minutes

The synthesis step is where elicitation quietly fails. Analysts either drown in transcripts and summarize from memory, or they hand the pile to an AI and inherit whatever the model decided stakeholders "probably meant." The fix is not avoiding AI — it is forcing citation discipline: no theme, no need, no pain point enters your requirements pack without a named speaker, a timestamp, and a verbatim quote you can check. The model does the sorting; you do the verifying; the stakeholder does the confirming.

When to use — and when not to

• Use it after any multi-session elicitation rounddiscovery interviews, requirements workshops, process walkthroughs — where you have transcripts or detailed notes from three or more sessions and need one coherent findings document.
• Use it when findings will justify requirementsa synthesis with citations lets any reviewer trace a requirement back to who asked for it and why, which is exactly what a challenge in sign-off review demands.
• Use it to reconcile sessions that disagreethe model is good at surfacing that operations said X on Tuesday and finance said not-X on Thursday, provided you ask it to look.
• Skip it for a single short conversationone call produces one page of notes; synthesis machinery adds ceremony, not value.
• Skip it when you have no transcriptsynthesis from your own recollection just launders memory through a model. Get recordings or contemporaneous notes first.
• Skip it for sessions recorded without consentif a participant did not agree to recording, the transcript should not exist, and it must not enter an AI tool.

The recipe

1. 1.

   Assemble the evidence pack. Collect every transcript from the elicitation round — meeting-AI exports from Teams, Zoom, or Fireflies work fine. Label each file with session name, date, and attendee roles. Strip anything the sessions touched that is out of scope for this analysis (HR matters, unrelated projects). Five minutes of curation here saves an hour of confusion later.

2. 2.

   Run the themed synthesis with citation rules. Paste the transcripts (or attach the files) into Claude or ChatGPT with this prompt:

   Copy

   You are a business analyst synthesizing elicitation sessions.
   Project context: [ONE LINE — e.g. "replacing manual vendor
   onboarding at a 900-person manufacturer"]
   Sessions: [LIST — e.g. "S1: ops team interview 12 May; S2: finance
   workshop 14 May; S3: IT walkthrough 15 May"]
   
   Produce a synthesis with these sections:
   
   1. THEMES — group what participants said into 4-8 themes. Under
      each theme, list every stated need or pain as one line in the
      form: [SESSION, SPEAKER ROLE, TIMESTAMP] verbatim quote —
      one-line interpretation.
   2. CONFLICTS — statements from different sessions or speakers
      that disagree, quoted side by side with citations.
   3. OPEN QUESTIONS — things mentioned but never resolved, with
      the citation of where they surfaced.
   4. NOT SAID — topics in our scope that no session covered.
   
   Rules: use ONLY the transcripts. Never merge two speakers into
   one quote. Never paraphrase inside quotation marks. If you are
   not certain a quote is verbatim, tag it [CHECK]. If a theme has
   thin evidence (one mention), tag it [THIN]. Do not infer needs
   that nobody stated.
   
   Transcripts:
   [PASTE STAKEHOLDER NOTES / TRANSCRIPTS]
   

3. 3.

   Read the output as a skeptic, not a customer. Good output looks like this:

   Copy

   THEME 2: Approval delays dominate onboarding pain
   - [S1, Ops lead, 14:32] "a new vendor sits for nine days waiting
     on the second signature" — cycle time driven by approvals, not
     data entry.
   - [S2, Finance manager, 08:17] "we won't drop the second approval,
     that's an audit finding waiting to happen" — control is
     non-negotiable; speed must come from routing, not removal.
   
   CONFLICTS
   - S1 Ops lead (22:10) wants approval thresholds raised to $50k;
     S2 Finance manager (31:44) states thresholds are set by treasury
     policy and are out of project scope.
   

4. 4.

   Spot-verify before anything ships. Pick at least three citations per theme — prioritize [CHECK] tags, quotes with numbers, and anything that would change scope — and open the transcript at that timestamp. You are checking three failure modes: quotes that were "cleaned up" into something nobody said, quotes attributed to the wrong speaker, and interpretations that stretch past what the words support. Delete or fix anything that fails. A synthesis you have not spot-checked is a draft, not a finding.

5. 5.

   Close the loop with the stakeholders. Send each session's participants the themes their session contributed to, with their quotes visible, and ask one question: "did we hear you correctly?" Corrections come back fast when people see their own words. Only after this confirmation do findings graduate into the requirements pack.

Make it yours

• Claude Project as a standing synthesizer: create a Project per engagement, put the project context and this prompt in the instructions, and add transcripts as you go. By week three you can ask "what has anyone said about exception handling?" and get cited answers across every session.
• ChatGPT with a reusable custom instruction: save the prompt as a reusable snippet or custom GPT so every analyst on the team produces the same section structure — comparable syntheses are what make program-level rollups possible.
• NotebookLM for a large transcript backlog: load all session transcripts as sources and use its grounded Q&A to chase specific topics with inline citations. Weaker at producing your full themed structure, strong at "where exactly did anyone mention SAP?"
• Dovetail-class repositories at team scale: if your org runs a research repository, file the verified synthesis there with tags for stakeholder group and theme, and link each finding to its transcript — the citation discipline carries straight into the tool.

Quality check

• Every finding traces to a sourcespeaker role, session, timestamp. A finding without a citation is an opinion wearing a costume.
• Spot-checks passedyou opened the transcript for at least three citations per theme and the quotes were verbatim and correctly attributed.
• No invented process steps or needssearch the synthesis for anything you cannot remember hearing; models fill silence with plausibility.
• Conflicts survivedif every stakeholder appears to agree, the synthesis is probably smoothing; real elicitation rounds contain friction.
• Stakeholders confirmedat least one participant per session has seen their quotes in context and said "yes, that's what I meant."
• [THIN] tags resolvedsingle-mention themes either gained corroboration in follow-up or are explicitly marked as unvalidated.