AI Field Manual
FM-MK-01 · Guide

Marketing  ·  Intermediate

Synthesize customer interviews into evidence-backed themes

Twelve customer interviews, roughly nine hours of conversation, and a readout due on Friday. You remember the vivid ones: the customer who was furious about pricing, the one who loved the new dashboard. Memory favors the loudest voices. This guide gets you from transcripts to themes you can defend, with an exact quote and an interview behind every claim, and counts you have checked yourself.

Status
Published
Reading time
About 14 minutes
You'll need
De-identified interview transcripts, your research questions, and an AI tool approved for customer data
Last reviewed
November 2026
Before you start

What AI is for here, and what it isn't

Interview synthesis is slow because it's careful work: reading every transcript against your questions, tagging what each person said, grouping tags into themes, and counting how widely each theme holds. Done by memory, it drifts toward the stories that were most vivid in the room.

AI is good at the mechanical part. It reads every transcript with the same attention, applies the same codes throughout, and attaches a quote to each one. That makes it easier to see the pattern rather than the anecdote.

It is also capable of producing a quote nobody said: a tidy paraphrase presented in quotation marks, or a sentence blended from two people. And its counts are often wrong. So the method here has one rule above all: every theme carries an exact quote and the interview it came from, and you check both before the readout. Which themes matter to the business, and what to recommend, stay with you.

Step 1

Prepare transcripts and questions

Give each transcript a participant ID and a short profile line that matters for analysis, without identifying anyone: "P04 · operations manager · mid-size customer · 2 years on product". Remove names, company names, email addresses and anything else that could identify a person.

Write down your research questions, the three to five things the study was meant to answer. "What do customers think?" isn't one. "What stops new customers from completing setup in the first week?" is. The questions are what keeps the coding focused. Without them, the AI will code everything, and you'll get forty themes.

Step 2

Code each transcript

Code one transcript at a time, or a few short ones together. Long batches lose detail, the same way one-pass summaries do.

The coding prompt
You are helping a researcher code customer interview transcripts.
Code only what the participant said, not the interviewer.

RESEARCH QUESTIONS:
RQ1: [question]
RQ2: [question]
RQ3: [question]

FOR THIS TRANSCRIPT ([participant ID]), list every passage relevant to
a research question. For each:
- RQ: which question it relates to
- CODE: a short label (3–6 words) describing what they said
- QUOTE: the participant's exact words, copied character for character.
  Do not paraphrase, shorten inside the quote, or join separate remarks.
  If you need to shorten, use [...] and keep each part exact.
- LOCATION: timestamp or line number

Then list anything notable that doesn't fit a research question, under
OTHER, with the same fields.

RULES: If a passage is ambiguous, code it and mark it [AMBIGUOUS].
Do not summarize the participant's views in your own words.

TRANSCRIPT:
[paste one transcript]

The instruction against joining separate remarks matters more than it looks. A quote stitched together from two moments in an interview can say something the participant never said.

Step 3

Group codes into themes

When every transcript is coded, combine the code lists and ask for themes.

The theme prompt
Below are coded passages from [n] interviews. Group the codes into
themes for each research question.

FOR EACH THEME:
1. THEME: one sentence stating the finding, not a topic label.
   ("New customers stall at data import", not "Data import")
2. PARTICIPANTS: the list of participant IDs whose passages support it.
3. EVIDENCE: the two or three strongest quotes, each with participant ID
   and location, copied exactly from the coded passages.
4. COUNTER-EVIDENCE: any passages that contradict or complicate it.
5. STRENGTH: how many participants, and whether they raised it
   unprompted or only when asked.

Also list codes that didn't fit any theme, and themes supported by only
one participant (label these SINGLE SOURCE).

CODED PASSAGES:
[paste]

Ask for counter-evidence every time. A theme with nothing against it is either very strong or not yet examined.

Step 4

Verify quotes and counts

This step is the reason the readout will hold up.

  • Check every quote you'll use by searching the transcript for it. If the exact words aren't there, the quote isn't usable, however reasonable it sounds. Find what the participant actually said, or drop it.
  • Recount the participants for each theme from the participant list, not from the AI's number. Counting is a common failure.
  • Check context. Read a few lines either side of each key quote. "I'd never pay that" means something different when the next sentence is "…for the old version."
  • Separate frequency from importance. Two participants who can't get through setup may matter more than nine who'd like a darker theme. That judgment is yours.
  • Describe counts as what they are. "Six of the twelve people we interviewed" is accurate. "Half of our customers" isn't: a dozen interviews show that a problem exists and what it's like, not how common it is across your customer base. If you need a prevalence figure, test the finding with a survey or product data.
Step 5

Worked example: why new customers stall

Fictional example. A product marketing team interviewed twelve customers who signed up in the past quarter. Research question 1: what stops new customers from completing setup in the first week?

The coding and theme prompts produce four themes for RQ1. The strongest: "New customers stall at data import because their spreadsheets don't match the required format", supported by P02, P04, P05, P07, P09, P10 and P12. Seven of twelve.

Checking it, the marketer finds two problems:

  • One quote doesn't exist. The readout's lead quote, attributed to P07, is "The import was the hardest part of the whole setup." Searching P07's transcript, the closest words are: "Honestly the import took longest. The rest was fine, but the import, yeah." Similar meaning, different words. The real quote goes in; it's arguably better.
  • The count is off. P10's passage, read in context, is about importing users from another system, not spreadsheets. The theme holds for six of twelve, not seven.

A second theme, "Pricing is too high", felt like the biggest story in the room because P03 was so emphatic. The theme prompt labels it with two participants, both raised only when asked. It goes in the readout as a minor finding, with the count.

The readout leads with data import, six of twelve interviewees, with two verified quotes and a recommendation the marketer writes: a downloadable import template and a sample file, tested with the next five sign-ups.

The readout is also careful about what "six of twelve" means. It describes these twelve interviews; it isn't an estimate that half of all new customers stall at import. Interview participants are chosen to explore a problem, not sampled to measure how common it is. So the readout adds a next step: check import completion in the product data for the whole quarter's sign-ups, which will show how widespread the problem actually is.

Step 6

Record the evidence

Theme evidence record
Study: [name], [n] interviews, [dates]     Analyst: [your name]
Research question: [RQ]

THEME: [finding, as a sentence]
Participants: [IDs] — [n] of [total] (recounted by hand)
Raised unprompted by: [IDs]

EVIDENCE (each quote verified in the transcript)
"[exact quote]" — [ID], [timestamp/line]  ✓ verified

COUNTER-EVIDENCE
"[exact quote]" — [ID], [location]

SO WHAT (analyst's judgment)
[why it matters to the business; recommendation]

AI assistance: coded with [tool]; quotes and counts verified by [name], [date]

What stays with you: which themes matter to the business, and what you recommend.

Esc

Saved in this browser

Reading list

Contents

Reference

Keyboard shortcuts

/ or Ctrl K
Search the manual
↑ ↓
Move through search results and index entries
Enter
Open the selected result
← →
Switch discipline tabs; previous / next entry in a workflow card
Esc
Close any dialog
?
Show this list