Over the past two years I have written somewhere north of forty customer case studies, first in-house at a B2B software company and now as a freelancer. For the last fourteen months, an AI assistant has been open in a second tab for every single one of them. I have used ChatGPT, Claude, and Gemini on this work, sometimes all three on the same project, and I have made most of the mistakes it is possible to make. I once shipped a draft where the model quietly turned a customer's 34 percent into “over a third.” I caught a completely invented quote about twenty minutes before it went to a client for sign-off. I have also cut my average drafting time from around six hours per study to under three.
So nothing below is theory. It is the workflow, template, and prompts I actually use now, plus the guardrails that stop AI from ruining the two things a case study lives or dies on: real numbers and a real customer voice. Follow the steps in order and you will end up with a draft that reads like a person wrote it, because a person did. The AI just handled the heavy lifting in the middle.
THE ONE RULE AI drafts. Humans verify. Every number, every name, and every quote gets checked against the interview transcript before anything ships. Break this rule and nothing else in this guide will save you. |
Where AI actually helps, and where it will hurt you
Fourteen months in, this is my honest scorecard. The pattern is simple: AI is excellent at processing your source material and terrible at replacing it.
| TASK | MY VERDICT | WHY |
|---|---|---|
| Writing interview questions | Excellent | It knows the shape of a good case study interview and adapts questions to any industry in seconds. |
| Mining transcripts for numbers, quotes, and timeline | Excellent | The single biggest time saver in the workflow. What used to take an hour of re-listening now takes five minutes. |
| Building the outline | Very good | Give it a fact sheet and a fixed structure and it maps facts to sections reliably. |
| Drafting individual sections | Good, supervised | Quality holds up when you draft one section at a time with tight rules. One-shot full drafts come out mushy. |
| Generating title options | Good | Ten options in seconds. Two or three are usually keepers, and the rest tell you what to avoid. |
| Trimming and tightening a draft | Good, with checks | It cuts filler well, but you must explicitly forbid it from touching numbers and quoted text. |
| Writing customer quotes | Never | It fabricates plausible quotes without hesitation. Quotes come from the transcript only, no exceptions. |
| The final voice edit | Do it yourself | Left alone, every model sands the personality out of a draft. The last pass is human work. |
The six-step workflow
The order matters. Each step feeds the next, and the whole system runs on what you capture in step one.
Step 1. Record a real interview
Book 25 to 40 minutes with the customer and record it with anything that produces a transcript: Zoom, Google Meet, Otter, Grain, or your phone in a pinch. There is no AI substitute for this. A case study written without an interview is a product page wearing a costume, and readers can tell.
Step 2. Feed the raw transcript, not a summary
Paste the whole transcript into the model, typos and filler words included. The first time I ran this workflow I pasted my own tidy summary instead, and the draft came out generic enough to belong to any company in the niche. The specifics you need live in the messy verbatim text.
Step 3. Extract before you outline
Run the transcript miner prompt below to pull out every metric, the strongest verbatim quotes, the timeline, and the pre-purchase problems. Save the output as your fact sheet. From this point on, the fact sheet is the single source of truth. If something is not on it, it does not go in the study.
Step 4. Outline, then draft one section at a time
Build the outline first, then draft the challenge, solution, and results sections in separate passes with tight length and style rules. Sectioned drafting is slower by about ten minutes and better by a mile, because you keep control at every stage.
Step 5. Quotes are copy and paste only
Every quote in the finished study must exist word for word in the transcript. Trimming for length is fine if the customer approves the trimmed version. Generating is never fine. This is the guardrail I am strictest about, because it is the one that nearly burned me.
Step 6. Verify, then get sign-off
Read the draft aloud once, check every figure against the fact sheet, then send it to the customer for written approval of their quotes and numbers. Approval is not bureaucracy. It is what lets you publish the story with their logo on it.
The template
Every study I write follows this structure. Aim for 800 to 1,200 words total. I have tested longer and it has never performed better for me, because people skim case studies for proof, not prose.
| SECTION | WHAT GOES IN IT | TARGET LENGTH |
|---|---|---|
| 1. Title | Customer name plus the strongest specific result, with a timeframe if you have one. Example: “How Meridian Logistics Cut Invoice Processing Time 62% in One Quarter.” | Under 70 chars |
| 2. At-a-glance box | Industry, company size, product used, and the three headline metrics. Most readers only ever read this box, so it has to stand on its own. | 40 to 60 words |
| 3. About the customer | Who they are and what they do, written for someone who has never heard of them. | 40 to 70 words |
| 4. The challenge | What was broken, what it cost them in time or money, and what they tried before you. This is the most under-written section in most case studies, and it is where the reader decides whether the story applies to them. | 150 to 250 words |
| 5. The solution | Why they picked you over the alternatives, how the rollout went, and one or two concrete details about what changed in their day-to-day work. | 200 to 300 words |
| 6. The results | Numbers with context: the baseline, the timeframe, and how they measured it. A bare “3x improvement” convinces nobody. | 150 to 250 words |
| 7. Closing quote | One strong verbatim quote that looks forward, placed right after the results. | 1 to 2 sentences |
| 8. Call to action | One short paragraph, one link, one action. Resist the urge to add three. | 50 to 80 words |
The six prompts, in the order you use them
Copy these as they are, swap in your details, and adjust the tone rules to taste. I wrote them while working in ChatGPT and Claude, and they behave the same in Gemini.
Prompt 01. The interview question generator
Run this the day before your customer call, then cut it down to the ten questions you actually want.
| You are helping me prepare for a customer case study interview. The customer is [COMPANY], a [SIZE] [INDUSTRY] company that used [PRODUCT] to [ROUGH OUTCOME]. Write 12 interview questions in a natural conversational order. Cover: the situation before, what they tried first, why they chose us, the rollout, what changed in their day-to-day work, measurable results, and what surprised them. Make every question open-ended and specific enough to draw out numbers and stories, not yes or no answers. |
Prompt 02. The transcript miner
The workhorse of the whole system. Its output becomes your fact sheet, which every later prompt depends on.
| Here is the raw transcript of a customer interview. Do not summarize it. Instead, extract: 1) every metric, number, or timeframe mentioned, quoted exactly as spoken, 2) the 8 to 10 strongest verbatim quotes, using the speaker's exact words, 3) the timeline of events, 4) the problems they described before using the product, 5) anything they said that sounded hesitant or negative. Present each group as a labeled list. [PASTE FULL TRANSCRIPT] |
Note: Keep the “hesitant or negative” list. It tells you what the customer will not want published, before it becomes a problem.
Prompt 03. The outline builder
Turns the fact sheet into a section-by-section plan and, usefully, tells you where your material is thin.
| Using the fact sheet below, build an outline for a customer case study with this structure: title, at-a-glance summary, about the customer, challenge, solution, results, closing quote, call to action. For each section, list the specific facts and quotes from the sheet that belong there. Flag any section where the source material is thin, so I know what to follow up on with the customer. [PASTE FACT SHEET] |
Prompt 04. The section drafter
Run once per section. The quote placeholder rule is what keeps fabricated quotes out of your draft.
| Draft only the CHALLENGE section of this case study, using the outline and fact sheet below. Rules: 150 to 250 words, plain conversational business English, short sentences, no marketing adjectives, do not invent any detail that is not in the fact sheet, and mark quote placements as [QUOTE HERE] instead of writing quote text. [PASTE OUTLINE + FACT SHEET] |
Note: Run it again for the solution and results sections, changing the section name and the length rule each time.
Prompt 05. The title generator
Run after the results section exists, because a title without the real metric in it is a wasted title.
| Write 10 title options for this case study. Each must contain the customer name, the single most impressive metric, and a timeframe if we have one. No puns, no colons unless genuinely necessary, under 70 characters where possible. Then mark your top 3 and explain why. [PASTE RESULTS SECTION] |
Prompt 06. The tightening pass
The final AI touch before your own human edit. The change log makes its edits easy to audit.
| Edit the draft below to cut 15 percent of the length without losing any facts, numbers, or quotes. Remove filler phrases, redundant adjectives, and throat-clearing openers. Do not change any number or any quoted text under any circumstances. Return the edited draft, followed by a list of every change you made. [PASTE FULL DRAFT] |
Five mistakes I made so you do not have to
Every one of these cost me real time, and one nearly cost me a client. Consider them field-tested.
I summarized the interview first
My tidy notes stripped out every number and every distinctive phrase, and the resulting draft could have been about any company in the industry. Raw transcript in, always.
I let the model write a “representative” quote
It produced a warm, plausible, entirely fictional quote that read better than anything the customer actually said, which is exactly what made it dangerous. I caught it at review. Quotes are copy and paste only now.
I trusted its numbers
In one draft, 34 percent became “over a third” and an 11-week rollout became “about three months”, which quietly turned a true claim into a false one. Models paraphrase numbers the way they paraphrase everything else.
I asked for the whole study in one go
The one-shot draft read smoothly and said almost nothing. Every section had the same rhythm and the same vague confidence. Section by section is the only way I draft now.
I nearly skipped customer approval
Deadline pressure is not a reason. The customer signs off on every quote and figure in writing before anything goes live. It protects the relationship, and frankly it protects you too.
What the human edit is for
A real pair from a recent project, lightly anonymized. Same fact, two drafts.
| AI FIRST DRAFT | AFTER THE TRANSCRIPT EDIT |
| “Meridian Logistics faced significant challenges with their invoice processing workflow, which was impacting operational efficiency across the organization.” | “Meridian's AP team was keying the same invoices into three systems by hand. At month end, two people lost whole days to it, and late payment fees were running about 4,000 dollars a quarter.” |
The second version is not cleverer writing. It is just the customer's own detail, pulled straight from the transcript and tidied up. Specificity is the entire game.
The pre-publish checklist
I run this list on every study, every time. It takes ten minutes and it has caught something on roughly one draft in three.
✓ Every number matches the fact sheet exactly
✓ Every quote exists verbatim in the transcript, or is an approved trim
✓ The customer has approved quotes and figures in writing
✓ The title carries the strongest metric
✓ Total length sits between 800 and 1,200 words
✓ There is one call to action, not three
✓ You have read it aloud once, start to finish
Final verdict: where I landed after fourteen months
The honest summary is this: AI made me faster, and the interview made me better. Drafting a case study used to eat a full working day, usually spread across a week of quiet avoidance. It now takes me around two and a half hours from transcript to client-ready draft, and because I no longer dread the work, I say yes to more of it. What AI did not do is improve results on its own. The studies that get quoted back to me by sales teams are the ones with the sharpest interviews behind them, where I asked the follow-up question about cost and got a real number, and where the customer's odd, specific phrasing survived every drafting pass. No prompt produces that. A good interview does. On tools: I have run this exact workflow through ChatGPT, Claude, and Gemini, and the differences between them are smaller than the difference between a careful prompt and a lazy one. Use whichever one handles your transcript length comfortably and respects a word limit without arguing. And one warning I mean sincerely. Skip the verification pass and you will eventually publish a number or a quote that is not real, and you will hear about it from the customer. That risk is entirely manageable, but only if you treat it as real. AI drafts. You verify. That order is the whole method. |
