How to Synthesize Stakeholder Interviews With AI

Prepare observations, not instant themes

Gather consented transcripts, interviewer notes, research questions, participant context, and relevant artifacts. Apply the organization’s privacy and retention rules before using AI. Give each interview and excerpt a stable identifier. Separate what a participant said or did from the analyst’s interpretation.

The UK Government Service Manual advises researchers to record observations rather than interpretations, sort observations into themes, derive findings, and decide actions. It also recommends making recordings available so observations can be confirmed (analyse a research session). Its interview guidance emphasizes open, neutral questions and real examples rather than generalities (using in-depth interviews).

Ask AI to extract candidate observations with excerpt IDs. A researcher should compare each extraction with the transcript before coding begins.

Sources: [1] [2]

Code in two passes

First, use descriptive codes close to the source, such as “copies data into spreadsheet” or “waits for manager approval.” Second, apply analytical codes tied to the decision, such as “manual handoff” or “unclear authority.” Keep a codebook with definition, inclusion rule, exclusion rule, and example.

Create a matrix with participants as rows and themes as columns. Each cell contains excerpt IDs, not merely a check mark. Add columns for stakeholder role, exposure to the process, and relevant constraint. Counts can show distribution in this interview set, but they do not turn a purposive qualitative sample into a population estimate.

Worked synthesis and dissent register

Imagine six synthetic interviews about approving AI-assisted client documents. Four contributors describe waiting for manager review, one manager says review is optional for low-risk drafts, and one risk lead says all external claims require source checks. The evidence suggests inconsistent interpretation, not a single agreed policy.

The synthesis artifact states: “Contributors described manager review as the practical gate, while the manager and risk lead described different conditions. The interviews indicate an unresolved distinction between approval of the document and verification of its claims.” It links the statement to excerpts I01-E4, I02-E7, I04-E2, I05-E9, I03-E6, and I06-E3.

The dissent register records the manager’s low-risk exception and the risk lead’s claim-level requirement. The next action is to map the written policy and observe one real review workflow.

A failed AI summary says, “Stakeholders agree that manager approval is required.” It overweights four similar accounts and erases two perspectives that change the decision. Repair it by reopening the excerpt matrix and stating the contradiction.

Move from themes to decisions

For each finding, record implication, confidence, contrary evidence, affected decision, and next evidence needed. Distinguish frequency from importance. A rare security constraint may matter more than a common inconvenience. Distinguish participant interpretation from documented process.

Use AI to test rival explanations: could the theme reflect role, tenure, business unit, or how the question was asked? These are hypotheses for follow-up, not conclusions.

Verification checklist

The output is a bounded stakeholder synthesis, not proof that the interview set represents an entire organization.

• Every finding links to exact excerpts from more than one perspective where possible.
• Observations and analyst interpretations remain separate.
• The codebook has clear inclusion and exclusion rules.
• Contradictory and minority evidence appears in a dissent register.
• Counts are described only within the reviewed interview set.
• Personal data and recordings follow consent, access, and retention rules.
• A second reviewer challenges high-impact findings and proposed actions.