How Do You Write a Case Study for an AI Agency?
Start with the client's situation before you arrived, in their language and with enough specificity that a similar buyer recognises themselves in the first two sentences. Then give three numbers: what the metric was before, what it is now, and how long the work took. Explain the approach in plain terms without turning it into a technical tour. Close with one direct quote from the client and one clear next step. That is the whole document, and it usually fits on a single page.
The reason most agency case studies fail is not writing quality. It is that they are written from the builder's point of view, which makes them a description of what was made rather than an account of what changed. A buyer reading your case study is not evaluating your craft. They are checking whether you have already solved a problem shaped like theirs, and whether the outcome was worth what it cost.
The Three-Number Case Study
Every case study we consider finished carries three numbers, and the framework is named after them because they are the part people leave out. The before number, the after number, and the elapsed time. Without all three, a reader has no way to judge whether the result was impressive, and no way to imagine the same thing happening to them.
- The before number. The state of the thing when you arrived: how many hours the process took, how many enquiries went unanswered, how long onboarding ran. This is the number clients never think to record, so ask for it during scoping, not at the end.
- The after number. The same measurement taken after the build, from the same source. Same source matters. A before figure from a client's memory and an after figure from a dashboard is not a comparison.
- The elapsed time. How long from kickoff to live. This is the number that does the most persuading in an AI build, because a short timeline is genuinely surprising and it is the part a buyer can verify against their own experience of software projects.
If you cannot produce all three, the honest move is to write the case study qualitatively and drop the numbers entirely rather than inventing a plausible one. A vague claim reads as modest. A specific claim that turns out to be decorative reads as a lie, and the client whose logo is on the page will notice first.
Capture the Numbers Before the Project Starts
The reason agencies have no case studies is almost never that the work went badly. It is that nobody wrote down what things looked like beforehand, and by the time the project is delivered the old state has been forgotten by everyone. The fix is administrative and takes ten minutes: make the baseline part of scoping.
- Ask the measurement question in discovery. How would you know this worked? Whatever they answer is your before number, and it is also your definition of done.
- Screenshot the current state. The old spreadsheet, the manual process, the inbox. These are the images that make a case study concrete and they stop existing the day you go live.
- Put the case study in the agreement. One line saying you may write up an anonymised summary of the work, with the client having approval over anything named. Asking at the start is routine. Asking afterwards is a favour.
- Diarise the follow-up. The real result is often visible thirty or sixty days after launch, not on delivery day. Book that check-in while you still have their attention.
This is one of the practical arguments for a properly structured scoping phase, since the same conversation that gives you an accurate quote gives you a baseline to measure against. If you are already running [a paid discovery phase](/blog/paid-discovery-phase-ai-projects), the measurement question belongs in the document you deliver.
The Six-Part Structure That Works
Use the same six parts every time and write them in this order. The consistency is the point: a reader comparing three of your case studies should be able to find the result in the same place in each, and you should be able to produce a new one in an hour rather than starting from a blank page.
- Situation. Who the client is, in a sentence, and what their world looked like before. Industry and size, not a paragraph of praise.
- Stakes. Why it mattered that this got fixed. What it was costing them in time, money, or missed work. This is the paragraph that makes the reader keep going.
- Approach. What you built and the two or three decisions that made the difference. Plain language, no architecture diagrams, no listing every technology.
- Result. The three numbers, stated flatly, plus anything that changed which cannot be counted.
- Quote. One direct sentence from the client. Their words, not yours, and specific rather than complimentary.
- Next step. One line inviting the reader to a conversation, with a link. A case study without a next step is a brochure.
Keep the approach section shorter than you want to. The build is the part you find most interesting and the part the buyer cares least about, and every extra technical paragraph makes the work sound more complicated than the price you are about to quote. Two decisions explained well beats a full account of the stack.
Getting a Quote Worth Printing
Ask specific questions and you get printable answers. Ask for a testimonial and you get one generic sentence about how great it was to work together, which persuades nobody because every agency page has the same one. The difference is entirely in how you ask, and the ask should happen on a call rather than by email.
- What was the most annoying part of the old process? The answer names the pain in the client's own words, which is exactly the language a similar buyer will recognise.
- What did you expect to go wrong that did not? This produces the objection-handling quote, and it is the single most useful sentence you can print.
- What would you tell someone considering this? A recommendation phrased as advice sounds like advice. Phrased as a testimonial, it sounds like marketing.
- What surprised you about the timeline? On an AI build the answer is usually about speed, and it lands better from them than from you.
Record the call with permission, pull the sentences verbatim, and send them the tidied version for approval. Never write the quote yourself and ask them to sign it off, even though they will usually agree. The specific, slightly awkward phrasing of a real client is what makes it read as real, and smoothing it out removes the only evidence that a person said it. The full process for collecting proof is in [how to get testimonials for your AI agency](/blog/how-to-get-testimonials-for-your-ai-agency).
What to Do When the Client Will Not Be Named
Write it anonymised and keep every specific except the identity. A named logo is nice. Specific numbers and a recognisable situation are what actually persuade, and you can keep all of those while replacing the client with a description. Confidentiality is a formatting problem, not a reason to have no case studies.
- Replace the name with a category. A twelve-person recruitment firm in Manchester tells the reader more than a logo they do not know.
- Keep the numbers, blur the scale if you must. A range is fine. Round numbers are fine. Removing them is not.
- Offer a reference call instead of a public quote. Many clients who refuse a written endorsement will happily speak to a prospect privately, and that converts better anyway.
- Get written permission for the anonymised version too. One email, kept on file, and it prevents the conversation you do not want to have later.
There is a second version of this problem worth naming. Some clients do not want it known that AI was involved in the work at all, which is a live consideration on this kind of build and one to settle before you write anything. The reasoning is covered in [should you tell clients you used AI](/blog/should-you-tell-clients-you-used-ai), and the safe default is that whatever you agreed privately governs what you publish.
Where the Case Study Actually Goes
A case study that only lives on a website page is doing a fraction of its job. The document is a sales asset first and a marketing page second, and its highest-value use is being sent to one specific person who has just raised the exact objection it answers.
- Attach it to proposals. One relevant case study attached to a quote does more than three paragraphs arguing you are qualified.
- Send it after an objection. When a prospect worries the timeline is unrealistic, the case study showing that exact timeline is the whole reply.
- Put it on the site with the result in the heading. Not our work with Acme. The heading is the outcome, because that is what makes someone click.
- Break it into posts. The situation is one post, the surprising decision is another, the result is a third. One case study is a month of material.
- Use it in cold outreach. A short specific result sent to someone with the same problem outperforms any pitch, because it is evidence rather than a claim.
Two or three strong case studies will out-earn a portfolio of twenty descriptions of work, so treat them as assets to be maintained rather than pages to be filled. This is also the strongest material your website can carry, which is the argument in [what to put on your AI agency website](/blog/what-to-put-on-your-ai-agency-website), and it is what turns a first project into a second one.
Short, practical drops on skills, MCP, agents, prompts, and more. No spam, unsubscribe anytime.
Frequently asked questions
How long should an agency case study be?
Under a thousand words in most cases, and one page is a good target. Buyers scan case studies looking for whether you have solved a problem like theirs and what the outcome was, so length works against you. If the detail is genuinely valuable, put the short version on the page and offer the longer write-up to anyone who asks.
What should a case study include?
Six parts in this order: the client's situation before the work, why it mattered, what you built and the key decisions, the result stated as three numbers, one direct client quote, and a clear next step with a link. The three numbers are the before state, the after state, and how long the project took from kickoff to live.
Can I write a case study without the client's permission?
Not with them named. You can write an anonymised version describing them by category and keeping the specifics, and you should still get written permission for that. The cleanest approach is to include a single line in your agreement at the start of the project giving you the right to publish an anonymised summary, with the client approving anything identifiable.
What if I do not have any results to show yet?
Write up your own builds instead. A project you built for yourself, with a real before and after, is a legitimate case study as long as you say clearly that it is your own. It demonstrates the same thing a client project does: that you can define a problem, ship something, and measure whether it worked.
How do I get a good quote from a client?
Ask specific questions on a call rather than requesting a testimonial by email. What was most annoying about the old process, what did you expect to go wrong that did not, and what would you tell someone considering this. Record with permission, pull the sentences verbatim, and send the tidied version for approval rather than writing it for them.
Should the case study mention that AI was used?
That depends on what you agreed with the client, and it is a conversation to have before you write anything rather than after. Some clients are happy for it to be a feature of the story, particularly where the timeline is the impressive part. Others prefer it left out entirely. Whatever was agreed privately governs what you publish.
Last reviewed by Duncan Rogoff on August 22, 2026
