How to run and analyze interviews about AI useA 30-minute script, an analysis prompt and an Airtable table

Choosing a format

Dots are a score from 1 to 5. For risk, more dots is worse.

Preparing for the meeting

• Mention the note-taking assistantIn the invitation and at the start of the call: the assistant takes notes and a transcript that only the team sees, and answers go into the report without names. With Granola, no audio is kept
• tl;dv or Granolatl;dv is connected to the calendar and joins the Google Meet or Zoom meeting on its own. Granola works without a bot, using your computer audio, and keeps only the transcript and notes
• Airtable row created in advanceName, group and numbers from the table of all people, status "In progress", before the first message
• Guide open on a second screenSeven script blocks and a 30-minute timer
• One host, no note-takingtl;dv or Granola does the transcript, ChatGPT does the analysis

A 30-minute script

Minutes per block

1. 1
   Intro

   "I'm looking at how people in the company actually get their work tasks done. This isn't an evaluation of you. Both successful and unsuccessful cases are equally useful to me. A note-taking assistant is with me: only we see the transcript, and answers go into the report without names. May I turn it on?"
2. 2
   Role and context

   "What do you spend most of your week on?"

   Which tasks repeat · where you look for information and draft things
3. 3
   The last work task

   "Think of the last work task that took you noticeable time. What did you need to get?"

   What means you had · what you chose and why · how you got to the result
4. 4
   Result and effort

   "What did you end up with, and how much effort did it take?"

   What you used as is · what you finished yourself · what constraints there were: data, access, security
5. 5
   A successful case

   "Think of a time a similar task went well. What was different then?"

   Tool · data · preparation
6. 6
   Quitting and switching

   "Have you ever stopped using a tool or switched to another? How did that go?"

   Which tool · what you switched to · why. If this never happened, skip it
7. 7
   Wrap-up

   "What didn't I ask that matters for understanding your work?"

   Thank them · say the answers go into the shared analysis without names

Staying on topic

Logging in Airtable

The table keeps everyone in the sample, their conversations and the report data in one place. Decision: whom else to invite so each group has enough interviews, whom to send a follow-up, and which gaps between self-reported and measured use to discuss with the person.

Voice: the tl;dv or Granola transcript is copied into the Notes field. Text: the chat log is copied into the same field with the date and time of each message.

ChatGPT fills in the orange field from the conversation in Notes. "Per report", "Tokens", "$" and "Activity" come from the ChatGPT Enterprise usage report. The "≠" field is a formula that compares the two frequencies.

Data limits: the report only sees the corporate tool and only for 3 months. Personal accounts and other tools do not show up in it, so a mismatch is a question to clarify, not a conclusion.

Text: wording

In a chat or a questionnaire it is hard to ask again, so a poorly worded question ruins the answer with no chance to fix it. Five common mistakes:

Fill the table with Claude Code

The whole table can be handed to an agent: Claude Code or any other MCP-capable agent connects to Airtable and creates rows, pulls report data and analyzes conversations on its own.

• Create rowsFrom the list of people selected for interviews: name, group, status
• Pull in report dataTokens, $, activity, frequency per report
• Analyze NotesWith the same prompt as in ChatGPT: fields, difficulties, self-reported frequency, quotes
• Update statuses"Partial" (incomplete answer) if the analysis contains "no data"

Questionnaire

The questionnaire is written after the interviews, once you know which tasks and difficulties occur. It goes to the whole company to learn how widespread they are. Questions are short, and the answer options come from the interview analysis.

1. 1
   Which of these tasks do you do at work?choice: tasks from the interviews · other, open fieldtasks
2. 2
   How often do you do these tasks?choice: every day · every week · less often · I don'ttask frequency
3. 3
   Which tools do you use for them?choice: ChatGPT · Claude · Cursor · another AI tool · I don't use AItools
4. 4
   Which difficulties do you run into in these tasks?choice: difficulties from the interviews · I don't run into any · other, open fielddifficulties
5. 5
   How often does this difficulty get in the way?choice: in almost every task · sometimes · rarelydifficulty frequency
6. 6
   How much does it matter for the result?choice: gets in the way a lot · gets in the way · barely mattersimpact
7. 7
   How do you work around this difficulty now?open field · I don'tworkarounds
8. 8
   What else matters that the questions missed?open fieldanything missed

Interview analysis: the analysis prompt and a check against report data

1. 1
   Give the Notes text to ChatGPT

   With the same prompt for every interview
2. 2
   Move ChatGPT's answer into the person's row

   Task, context, chosen way, result, effort, constraints, difficulties, self-reported frequency
3. 3
   Group the table by difficulty

   People with the same difficulty end up next to each other: you see how many there are and from which groups
4. 4
   Check with the questionnaire

   Repeats become answer options in the company-wide questionnaire; the team breakdown is compared with Uniformity

Prompt for interview analysis

Below is a conversation with an employee about their work tasks.
Return strictly by field, only what the person said themselves:
1. Task: the last specific work task
2. Context: for whom and why, what means they had
3. Chosen tool or way, and why
4. Result
5. Effort: time and manual rework
6. Constraints: data, access, security
7. Difficulties: categories from the list {categories from the report's sign table} or "no difficulties"; if none fits, a new category in your own words
8. Self-reported frequency: weekly / regularly / occasionally / doesn't use / unknown
9. 1–2 verbatim quotes
If something is not in the conversation, write "no data" and don't make it up.

{text from Notes}

Further reading