AI for Business

How to Write Case Studies From Customer Interviews Using AI

Sagnik Bhattacharya By Sagnik Bhattacharya 28 Sep 2026 11 min read
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Write Case Studies From Customer Interviews Using AI.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Write Case Studies From Customer Interviews Using AI.

Build the case study from a structured brief, not straight from the transcript. Interview the customer for 30 to 40 minutes on the situation before, their decision, what happened and what changed; have AI extract exact quotes and every number into a brief; draft from that in your fixed format; then get the customer's approval. Expect three to five hours.

The step that needs the most discipline is the results section. Given a transcript with soft outcomes, a model will cheerfully write "improved engagement by 40%" because that's what case studies usually contain. Your results must come from the customer's words and records, which is why the brief step below lists every number with a timestamp before any drafting starts.

Follow me on Instagram@sagnikteaches

What a case study has that a testimonial doesn't

A testimonial says "they were great". A case study shows a buyer like them what the problem was, why they chose you, what actually happened and what changed, with evidence. It answers the questions a cautious buyer or funder would ask on the phone. A reliable structure:

Connect on LinkedInSagnik Bhattacharya
  1. Headline that states the result in plain words.
  2. Snapshot box: who the customer is, what they bought, the timescale, two or three headline numbers.
  3. The challenge, in the customer's words where possible.
  4. Why they chose you, including what else they considered.
  5. What happened, step by step, including anything that went wrong.
  6. Results, measured and observed.
  7. One strong quote, attributed by name and role.
  8. What's next: renewal, expansion, advice to others.

The snapshot box is the part most readers see, so it's worth drafting it first as a target. A filled-in version for the example below:

Subscribe on YouTube@codingliquids 
SNAPSHOT
Customer:   40-bed residential care home (named, with permission)
Service:    weekly music and memory sessions, 45 minutes each
Timescale:  24 sessions over six months, renewed for another 24
Results:    attendance up from 6 to 14 residents by week eight
            (home's activity log); two residents who rarely leave
            their rooms now attend every session
Quote:      "It's the one hour of the week nobody has to be persuaded
            to come to." Activities coordinator

The example running through this tutorial is illustrative: a small charity that runs music and memory sessions for people living with dementia, paid for by care homes. It wants a case study from a 40-bed care home that booked 24 weekly sessions, to use on its website, in funding bids and in proposals to other homes.

Pick the customer and the people to interview

Choose a customer whose situation matches the buyers you want more of, and who has something measurable to show. Then interview two people if you can: the one who felt the problem day to day (here, the activities coordinator) and the one who signed off the spending (the home manager). The first gives you the texture; the second gives you the reasons and the numbers.

A request that tends to get a yes:

Subject: Could we tell the story of your sessions with us?
Hi [name], the sessions at [home] have been some of our best, and
other homes keep asking what they're like in practice. Would you
and [manager] give us 30 minutes on a video call? We'll write it
up, you'll see and approve every word before it's used anywhere,
and you can choose whether the home is named. No residents will
be named or shown.

The last sentence matters for this sector, and it leads to the next step.

Consent and confidentiality when people's care is involved

Case studies in health, care, education or anything touching vulnerable people need extra care, because the most vivid details are often personal data about someone who isn't in the interview. For the charity, the rules it set before recording were:

  • No resident is named, described in identifying detail, quoted or photographed, unless the home has confirmed consent through its own proper process. The charity's default is simply to leave residents out.
  • The care home confirms in writing which uses it agrees to (website, bids, proposals, social media) and whether it's named.
  • Recordings and transcripts are stored in one place and deleted once the case study is approved.
  • Before any transcript goes into an AI tool, names of residents and staff who aren't being quoted are replaced with roles ("Resident A"). Use a business plan or an account with model training switched off.

Health information is a sensitive category under data-protection law such as the GDPR, so if you're unsure, check with your data-protection adviser before you record rather than after.

The interview: twelve questions in four parts

Before

  1. What was going on before you started working with us? What did a typical week look like?
  2. What had you already tried?
  3. What made you decide something had to change, and when?

The decision

  1. How did you find us, and who else did you look at?
  2. What nearly made you say no?
  3. Who had to agree, and what did they need to see?

What happened

  1. What did the first few weeks look like?
  2. Did anything go wrong or need adjusting?
  3. Is there one session or moment that sums it up for you?

Results

  1. What's different now? What can you count or point to?
  2. What do staff, families or others say about it?
  3. What would you tell another manager thinking about it?

When you hear a number, ask where it comes from ("Is that from your activity records?"). That one follow-up saves an awkward correction at the approval stage. When you hear a vague answer ("it's made a real difference"), ask for an example: "What did that look like last Tuesday?"

Interview habits that ruin the transcript

Most weak case studies are decided in the interview, not the drafting. Four habits show up again and again when you read the transcript back:

  • Leading questions. "The sessions have been brilliant, haven't they?" gets "Yes, brilliant." You can't quote a yes. Ask "What did you notice after the first month?" instead.
  • Talking too much. If your lines take up more of the transcript than theirs, you've been pitching. Count the speaker turns; the customer should have roughly three-quarters of the words.
  • Only interviewing the fan. The coordinator loved the sessions but couldn't say why the home paid for them. Without the manager, the "why they chose you" section would have been guesswork.
  • Recording somewhere noisy. A call taken in a busy lounge produced a transcript where every third sentence was "[inaudible]". Ask for a quiet room, or a headset.

From recording to a clean transcript

Record with permission, stated at the start of the recording. Your video-call tool's built-in transcript is usually good enough; dedicated apps such as Otter.ai start at about $8.33 per user a month on annual billing if you do this often. Accuracy matters more than speed here, so spend 15 minutes on a correction pass: names, job titles, product terms and any numbers. AI versus human transcription covers when paying for a person is worth it, which is mainly when the recording quality is poor or the stakes are high.

Then do the redaction described above. It takes five minutes with find-and-replace, and it means the version the AI sees contains nothing you'd regret sharing. An illustrative line before and after:

Before: "Margaret in room 12, she's got quite advanced dementia, she hadn't come down for anything since her husband died, and she sang the whole of the second song."

After: "Resident A, she hadn't come down for anything in months, and she sang the whole of the second song."

The redacted line keeps the moment that matters and drops the name, room number, diagnosis and bereavement. Even so, a detail this specific shouldn't appear in the published case study without the home's say-so, because staff and families could still recognise who it is.

Prompt one: turn the transcript into a structured brief

Below is a corrected interview transcript. Do not write a case study yet.
Produce a brief with these headings:
  CUSTOMER SNAPSHOT: who they are, what they bought, timescale
  CHALLENGE: 3-5 bullet points, each with a supporting EXACT quote + timestamp
  DECISION: why they chose us, alternatives considered, objections
  WHAT HAPPENED: in order, including problems and fixes
  RESULTS: list EVERY number mentioned, with its exact wording,
    timestamp and what it measures. Separate observed results from numbers.
  BEST QUOTES: the 5 strongest, verbatim, with speaker and timestamp
  GAPS: questions we still need answered to make this credible
Copy quotes exactly. If something isn't in the transcript, don't include it.
[paste transcript]

An illustrative slice of the brief for the care home:

RESULTS
- "We started with six coming down, by week eight it was fourteen." (coordinator, 18:40)
  measures: residents attending each session
- "Twenty-four sessions, and we've just booked another twenty-four." (manager, 31:02)
  measures: renewal
- Observed: "Two residents who never leave their rooms come down for it." (coordinator, 21:15)
- Observed: families now ask which day the sessions are on (manager, 27:30)
GAPS
- Is the 6 to 14 figure from the activity log? (coordinator said "roughly")
- Did anything change for staff time? Not discussed.
- Permission to name the home: not yet confirmed.

The GAPS section is where the brief earns its keep. Here it caught that the headline number was "roughly" and needed checking against the home's own log. A two-line email to the coordinator confirmed it (the log showed 6 in week one and 14 in week eight), and the case study could state it plainly.

Prompt two: draft in your house format

Once the brief is checked, draft from the brief, not the transcript. The model is less likely to invent when it's working from a short, verified document.

Using ONLY the brief below, write a case study of 700-900 words in
this structure: [paste your 8-part structure]. Tone: [from your voice
guide]. Use the customer's quotes verbatim. Every number must appear
in the RESULTS section of the brief. Where the brief has a gap, write
[GAP] rather than filling it. Headline: state the main result plainly.

Even with those instructions, check the results paragraph line by line. Here is the difference between a typical unguided draft and the approved version:

Unguided draft: The programme delivered life-changing outcomes, with resident engagement increasing by over 130% and significant improvements in mood and wellbeing across the home.

Approved version: Attendance more than doubled over the first eight weeks, from six residents to fourteen, according to the home's activity log. Two residents who rarely leave their rooms now come down for every session, and families have started asking which day it runs. After 24 sessions, the home booked another 24.

The first version's percentage is technically arithmetic on the real figures, but "mood and wellbeing" was never measured, and "life-changing" is the kind of phrase a funder skims past. The second is less dramatic and far more believable. For a quick way to catch this across any document, see fact-checking AI marketing copy before it goes live.

Headlines: ask for options, then pick the plainest

The headline decides whether anyone reads the rest, and models default to slogans. Ask for five, each built on a different fact from the brief:

Write 5 headlines for this case study, each under 12 words and each
built on a different fact from the RESULTS section. No puns, no
"transform", no "journey". Say which fact each one uses.

An illustrative set, with the verdicts:

  1. "How one care home more than doubled attendance at its weekly sessions" (attendance figure). Chosen: specific and checkable.
  2. "The hour nobody has to be persuaded to attend" (quote). Good as a subheading.
  3. "24 sessions, then 24 more: why a care home renewed" (renewal). Strong for the proposal version.
  4. "Bringing residents out of their rooms with music" (observed result). Warm but vague.
  5. "A symphony of connection" (nothing). Deleted.

The approval round

Send the draft to the customer with specific questions rather than "any thoughts?", which invites a rewrite or silence:

Here's the draft. Could you check three things by [date, 10 days away]?
1. Are the figures in the snapshot box right?
2. Are you happy with how you're quoted? (quotes are highlighted)
3. Are you still happy for the home to be named?
Anything else you'd change, just mark it up.

Expect one of three responses: approval with a tweak or two, a request to soften something that felt too personal, or a manager who wants their own sign-off to go through a head office. Allow up to two weeks, chase once politely, and never publish on "no reply". Keep the approval email with the final text; if anyone ever questions a quote, that's your record, and the rules on testimonials and AI show why that record matters.

One awkward case deserves a plan. Sometimes the customer's head office or communications team "improves" the draft, replacing the coordinator's plain words with corporate phrases ("delivering person-centred enrichment"). Don't argue line by line. Thank them, accept any factual corrections, and ask whether the original quotes can stay as spoken because they're what makes the story believable to other homes. Most comms teams agree once they see that the quotes are attributed and approved by the person who said them.

Three versions from one approved case study

Once approved, one case study gives you several working assets without new interviews:

  • Website page: the full version, with the snapshot box at the top and a link to your enquiry form.
  • Funding-bid paragraph: 150 words focused on need, method and evidence, which is what a funder scores. The charity dropped the renewal detail and led with the attendance figures.
  • Proposal proof block: three lines and one quote, pasted into every proposal to a similar home. Writing proposals faster with AI shows where it fits, and it's equally useful when responding to tenders that ask for evidence of past work.

If you want social posts and emails from the same interview, the moments approach in turning one customer story into a month of marketing reuses the transcript you already have. Check the consent covers each new use first.

The same method for a commercial customer

The charity example is sensitive by nature, but the method is identical for an ordinary commercial case study. Take a boutique hotel writing up a software company that has held four staff offsites there in a year. The interview is with the company's operations manager, who books the offsites, and its founder, who approves the spend.

The brief's RESULTS section would look quite different: 18 staff per offsite, four bookings in twelve months, the switch from a city conference venue, and the manager's comment that planning time dropped "from about a week of back-and-forth to one call and a spreadsheet". The GAPS section might flag that the founder called the hotel "cheaper overall" without a figure, so the hotel either asks for the comparison or drops the claim. The headline could be "Four offsites in a year: why one software team stopped using conference venues", and the proof block goes straight into the hotel's corporate enquiry replies. No personal data concerns beyond the two people quoted, so the consent step is a single line in the approval email.

Checking whether the case study is doing its job

A case study is working if it shows up in decisions. Three months after publishing, the charity checked:

  • Proposals: of the six proposals sent to care homes with the proof block, how many led to a call, compared with the six before it.
  • Enquiry form answers: whether anyone mentioned the case study or the home by name.
  • Funders: whether bid feedback referred to the evidence section.
  • Page analytics: time on the case study page and clicks from it to the enquiry form.

Small numbers won't prove anything statistically, but they tell you whether to write the next one in the same format. Also set a reminder to refresh each case study after a year, because a result from last year that has since improved ("now booked for a third block") is a stronger proof point than the original.

As a time budget for the full case study: 30 minutes arranging, 40 recording, 30 correcting and redacting, 15 for the brief, an hour drafting and editing, and 20 minutes on the approval email and final changes. Around three and a half hours of your time, most of it the parts no AI can do for you.

Case study questions that come up after the interview

Can a case study work if the customer wants to stay anonymous?

Yes, though it carries less weight. Describe the customer precisely ('a 40-bed residential care home') and keep their quotes, attributed by role. Many customers who refuse a public case study will agree to one you share privately with prospects or funders, so offer that option. Never add a fake name or stock photo to make an anonymous story look like a named one.

How long should a case study be?

Write the full version at 600 to 1,000 words, because that length forces you to include the detail that makes it credible. Then cut a 150-word version for proposals and bids and a three-line version for your website listing. Buyers who are close to a decision read the long one; everyone else reads the headline and the numbers.

Should I offer the customer something for taking part?

A thank-you is fine, such as a small discount on the next booking or a donation to a cause they choose, but agree it before the interview and never make it depend on what they say. If you give anything of value, say so if asked, and be careful that the reward doesn't turn their words into a paid endorsement you'd have to disclose.

What if the results are real but hard to measure?

Use what can be observed and counted: attendance, repeat bookings, time saved, staff comments, the customer's own decision to renew. Qualitative results are honest if they're specific ('two residents who rarely left their rooms now come down for every session'). What you must avoid is converting soft results into percentages the customer never gave you.

Further reads

Sources: Otter.ai pricing page (Pro plan, annual billing); general interviewing and editorial practice.

Want a case study routine your team can run?

On a 1:1 call we'll turn your interview questions and house format into prompts in your AI tool, and agree how transcripts and approvals are handled safely.

Book a 1:1 call with me