How to analyze customer interview transcripts with AI, and keep every quote real

The method in eight steps

1. Prepare the transcripts
2. Find who does the job, and which job
3. Map the steps of the job
4. Write the desired outcomes
5. Collect what people use today
6. Check every quote word for word
7. Find the thin steps and plan the next calls
8. Group the outcomes into value themes

The method is jobs to be done, in the style of Outcome-Driven Innovation. If the words are new to you, read jobs to be done, explained first. The examples below come from an invented study of freelance bookkeepers.

1. Prepare the transcripts

• Keep one file per interview, as .md, .txt, .docx or .pdf.
• Keep the speaker labels, so it is always clear who said what. A question from the interviewer is never evidence.
• Add one line at the top with the person's role and setup, for example “Freelance bookkeeper, 25 small-business clients, works alone.”
• Use the raw transcript, not a cleaned-up summary. A summary has already lost the words you will need to quote.

2. Find who does the job, and which job

For each interview, write the person's own job as a verb, an object and a clarifier: “Close a client's books each month.” Name no product and no measure of success. Then ask “why?” of each person's job until the jobs of different people meet. If they only meet at “run a business”, you are looking at two different jobs, and each needs its own map.

3. Map the steps of the job

List the steps people describe and put them in order. Almost every job passes through eight stages: define, locate, prepare, confirm, execute, monitor, modify and conclude. Each step is a verb and an object, such as “Collect the client's documents”. A first batch of interviews gives you part of the map; a mature map has 12 to 20 steps.

4. Write the desired outcomes

A desired outcome says how the person judges success at one step: a direction, a metric, an object and a clarifier. Most start with “Minimize” and measure time, likelihood or number.

Put each outcome under the step where the person can prevent the failure, and check for an existing outcome with the same object before you add a new one.

5. Collect what people use today

Note every product, everyday tool, person and home-made method people use at each step, with the quote that shows it: “I keep a spreadsheet of who still owes me what.” Keep what they stopped using too, with the reason: “We tried a client portal. Nobody logged in.” A solution people gave up on is some of the best evidence you will get.

6. Check every quote word for word

Asked for a summary, an assistant will often reword what people said and still put it in quotation marks. A reworded quote looks like evidence, but nobody said it. Researchers who study the voice of the customer have long insisted that needs be kept “in the customers’ own words, and not in industry or company jargon, in order to not lose meaning.”

By hand, search for each quote in its transcript before you use it. Product Weaver does it for you: a small script checks every quote against a copy of its transcript, and refuses the whole batch if one quote does not match.

7. Find the thin steps and plan the next calls

Count how many people support each step and outcome. Steps that rest on one person, or on nobody, are where your next interviews should go. Write three to five questions about them before the next call. Treat the counts as a signal from interviews, not as a market measure: only a survey can tell you how important a need is across a market.

8. Group the outcomes into value themes

Once you have dozens of outcomes, group them into five to nine needs in the customer's voice, each starting with “Help me”: “Help me get what I need from clients without chasing them.” A good theme is not a step, names no solution, holds one need, and suggests several different ideas. Sort every product idea into a theme.

How many interviews do you need?

Fewer than most people think, if they are deep and one-on-one. In a typical voice-of-the-customer study, 10 to 30 customers are interviewed one-on-one for about an hour each (Gaskin, Griffin, Hauser and others, 2011). In a study of a fairly uniform group, Guest, Bunce and Johnson (2006) found that most themes appeared within the first 12 interviews, and their basic elements within the first 6.

So start with the interviews you have, analyze after each batch, and stop when new interviews add few new outcomes. Product Weaver's map says how sure the findings are: under 3 conversations is a hunch, and a clear pattern needs 8 or more, with the last 2 adding at most 2 new needs.

Doing it with Product Weaver

Install the plugin in Claude, ChatGPT and Codex, or GitHub Copilot. Open the folder with your transcripts and ask in plain words, for example:

• “Process this week's discovery interviews in the Interviews folder.”
• “Add these three calls to the knowledge base and show me what changed.”
• “Group the outcomes into value themes.”

You get a knowledge base that grows with every batch, a review sheet with the decisions that wait for you, and a discovery map in one HTML file. The plugin runs on your machine and sends nothing unless you turn on usage statistics for a study, and even then never your interviews, names, quotes, prompts or file names.

Doing it by hand with any assistant

You can get a first pass from any AI assistant with a prompt like this one:

Help me analyze customer interviews with the jobs-to-be-done method.
For each transcript I give you:
1. Name the person's role and setup.
2. Write the job they are trying to get done as verb + object + clarifier.
   Name no product and no measure of success.
3. List the steps of that job, in order.
4. For each step, write what they want as "Minimize the time / likelihood /
   number of ..." statements.
5. List the tools, people and workarounds they use today, and anything they
   stopped using, with the reason they gave.
For every item, give one exact quote from the transcript, copied character for
character, with the speaker's name. If there is no exact quote, write
"no quote". Never paraphrase inside quotation marks.

It works, with limits. The assistant does not remember the last batch, so you merge results by hand. Nothing checks the quotes, so you search for each one yourself. And there is no map: you build the table and the counts in a spreadsheet.

Common mistakes

• Quotes that are really paraphrases.
• A job that names a product: “Use an app to collect receipts.”
• Outcomes with no metric: “easier”, “better”, “smoother”.
• Two different jobs squeezed into one map.
• Reading interview counts as if they were survey results.
• Analyzing once and stopping: the evidence should grow with every batch.

Sources

1. Steven P. Gaskin, Abbie Griffin, John R. Hauser, Gerald M. Katz and Robert L. Klein, “Voice of the Customer”, Wiley International Encyclopedia of Marketing, 2011. stuff.mit.edu
2. Greg Guest, Arwen Bunce and Laura Johnson, “How Many Interviews Are Enough? An Experiment with Data Saturation and Variability”, Field Methods 18(1), 2006. doi.org
3. Lance A. Bettencourt and Anthony W. Ulwick, “The Customer-Centered Innovation Map”, Harvard Business Review, May 2008. hbr.org

Outcome-Driven Innovation® is a registered trademark of Strategyn, LLC. Product Weaver follows the method as published; it is not made or endorsed by Strategyn.