How to Write a Client Case Study Using AI in 20 Minutes

Why AI case studies sound like fluff

AI writes fluff about client work for one structural reason: when you ask a single prompt to read, interpret, structure, and polish all at once, it has to guess in the gaps. Guessing is what produces the invented percentage and the word “transformational”. Split the job into stages and there are no gaps left to guess in.

That is the method we teach at PromptPeakAI as the Evidence Relay: four handoffs — Mask, Extract, Assemble, Verify — where each stage may only pass forward facts it actually received from the stage before it. Nothing new is allowed to appear mid-chain.

The difference shows up immediately in what you get back. Here is the same set of client notes run two ways.

If this staged approach feels familiar, it should — it is the same logic behind chaining prompts for multi-step work, applied to a deliverable that has a client’s name on it.

Mask client data before you prompt

Before any client material reaches an AI tool, swap the identifying details for neutral tokens. This takes about ninety seconds with find-and-replace, and it is the step that separates a workflow you can defend to a client from one you cannot. Skip it and you are not writing a case study — you are uploading someone else’s commercial position to a third party.

Most professionals get the risk boundary wrong, and it is worth being precise about. The line is not free versus paid — it is a personal account versus a business or enterprise account. Per OpenAI’s enterprise privacy commitments, updated January 2026, the contractual no-training guarantee covers ChatGPT Business, Enterprise and the API. Anthropic draws the same line: Anthropic’s commercial data policy excludes Claude for Work and API inputs from training by default.

A paid personal subscription does not buy you that guarantee — it buys you a setting you have to manage yourself. Our guide to whether ChatGPT is safe to use at work walks through where those controls actually live.

What to mask, what to keep

Mask identity, keep shape. The AI needs the structure of the story to write it well; it does not need to know who the client is.

• Replace: company name → [CLIENT], individual names → [CONTACT], exact figures → [METRIC 1], [METRIC 2]
• Keep: industry, company size band, the problem, the sequence of what you did, the direction of the result
• Keep a key in a separate file mapping each token to its real value — you will swap them back at the end

If your source is a raw recording rather than notes, clean it up first — the same tidying pass described in turning a meeting transcript into usable minutes makes everything downstream sharper.

You are reviewing a document before it is shared outside my company.

Read the text below and list every detail that could identify the client, its staff, or its commercial position. For each one, tell me the exact text to find and a neutral token to replace it with.

Include: company names, product names, people's names, job titles that are unique enough to identify someone, locations, exact revenue or contract figures, and named third parties.

Do not rewrite the document. Return a two-column list only: FIND / REPLACE WITH.

Text:
[PASTE YOUR NOTES OR TRANSCRIPT]

The four parts of a case study

Every client case study that gets approved does the same four things in the same order: states the problem, explains what you did, shows what changed, and lets the client say it in their own words. When you write a client case study using AI, fixing the length of each part is what stops the model padding.

Note how tight the Results section is. Buyers scan it first and read it last, which is why it earns the same discipline as a proper executive summary — short, specific, and defensible line by line.

How to write a client case study using AI

Here is the full run. To write a client case study using AI properly, each step gets its own prompt, and you read the output before moving on — that reading pass is not optional, it is the part that catches problems while they are still cheap to fix.

Stage two asks for bullets, not prose. You are building the evidence list that every later stage is restricted to.

You are an analyst preparing source material for a case study. Do not write any prose yet.

From the masked notes below, extract four bulleted lists:
1. THE PROBLEM — what was wrong before, with any baseline figures actually stated
2. THE ACTIONS — what was done, in chronological order
3. THE OUTCOMES — what changed, with figures ONLY where the text gives them
4. QUOTE CANDIDATES — up to five sentences the client said that could work as a pull quote, copied word for word

Rules: use only information present in the text. Where a figure is missing, write [NO FIGURE GIVEN] rather than estimating. After each outcome bullet, note which line of the source it came from.

Notes:
[PASTE MASKED NOTES]

Now hand the AI its own extracted list and nothing else. Starting a fresh chat here is a good habit — it removes any chance of the model reaching back into the raw transcript.

You are a B2B copywriter. Write a client case study from the fact list below.

Structure, in this order and within these limits:
- THE CHALLENGE (75–100 words)
- THE APPROACH (100–150 words)
- THE RESULTS (75–100 words)

Make the client the subject of the story and our work the thing that helped. Write in plain professional English.

Do not use these words: transformational, seamless, revolutionary, game-changing, cutting-edge, synergy, robust, best-in-class.
Do not introduce any fact, figure, or claim that is not in the list below.
Where the list says [NO FIGURE GIVEN], describe the outcome in words without a number.

Fact list:
[PASTE THE OUTPUT OF PROMPT 2]

Take the strongest quote candidate — the one the client actually said — and clean it up without rewriting what it means. Spoken language has filler; that is all you are removing.

Below is a sentence a client said out loud in a recorded conversation.

Clean it into a two-sentence testimonial: remove filler words and false starts, fix grammar, and keep the sentence order.

Keep their vocabulary and their level of formality. If they said "sorted it out", do not upgrade it to "resolved the matter". Do not add praise they did not give.

Return three versions so I can choose.

Raw quote:
[PASTE THE VERBATIM QUOTE]

The last stage is an audit, and you run it against the draft rather than the notes. This is where the Evidence Relay earns its name — anything that cannot be traced back to a real handoff gets flagged.

Compare the draft case study against the fact list it was supposed to be built from.

List every claim, figure, date, or outcome in the draft that does NOT appear in the fact list. Quote the exact sentence for each one.

Also flag any sentence that describes a result without evidence behind it.

Do not fix anything. Just produce the flag list.

Fact list:
[PASTE PROMPT 2 OUTPUT]

Draft:
[PASTE PROMPT 3 OUTPUT]

Each of these prompts is built the same way — a role, a goal, the context, and a required format — which is the Logic System doing quiet work in the background. Swap the real names and figures back in, and the draft is ready for a human read.

Getting client sign-off on the quote

The quote is where approvals go to die. Clients reject testimonials not because the content is wrong but because it does not sound like them — and a client who reads a sentence they would never say starts doubting the rest of the page.

Send the client two or three polished versions of their own sentence rather than one finished quote. Choosing feels like authorship; approving someone else’s draft feels like homework. Most account managers find this single change collapses the revision cycle more than any wording tweak.

Whether to mention that AI helped draft it is a separate call, and one worth making deliberately rather than by omission — we cover the trade-offs in telling clients you use AI.

What AI cannot do here

The Evidence Relay narrows the risk; it does not remove it. Three parts of this job stay yours no matter how good the drafting gets, and quietly handing any of them to a model is how professionals end up publishing something they cannot stand behind.

Invented figures are not a sign you prompted badly — they are a known behaviour, and worth understanding rather than just guarding against. Our explainer on why AI invents confident details covers the mechanism, and the broader routine in fact-checking AI output before your name goes on it applies directly to client-facing work.

Fixing the four common failures

Four things go wrong often enough to be predictable. All four have a prompt-level fix, which means you can solve them permanently rather than editing around them every time.

The second row is the one professionals underrate. Telling a model what not to write shapes output more reliably than any amount of positive instruction, which is why negative constraints belong in every prompt in this chain, not just the drafting one.

Save the whole set somewhere you will find it again — a notes file, a pinned document, or a project folder in whichever tool you use. Three things are worth keeping:

• The five prompts, in order, with your banned-words list already filled in
• Your masking conventions, so tokens mean the same thing across every client
• One finished case study as a reference for length and tone

The second time you write a client case study using AI it takes ten minutes, because the only new work is the masking pass and the reading.

And keep the discipline that makes the whole thing hold together: every sentence in a client case study should be traceable to something a person actually said or a file you can open. That is the only standard that survives a client’s legal review.