How to write a case study for an AI project

A one-page structure that hiring managers actually read, with an example of each part.

By MT, AI Solutions Engineer. Updated 2026-10-02.

Key takeaways

  • One page is enough. Lead with the problem and the number.
  • Spend the most space on what broke and how you fixed it.
  • Use real numbers only, and never inflate them.

Why the write-up matters

A project nobody can understand in two minutes might as well not exist. The case study is how a hiring manager, a client or your own boss judges your work, usually in less time than it took you to build one feature. A clear page beats a long one. If you don't have a project yet, start with five portfolio project ideas.

The one-page structure

  1. The problem. One or two sentences: who had it and what it cost them. "A five-person team spent about six hours a week typing invoice details into a spreadsheet."
  2. The scope. What you built, and what you deliberately left out. "Reads PDF invoices into vendor, date and amount. Doesn't handle handwritten invoices yet."
  3. How it works. Three or four steps in plain language, including what happens when the AI isn't sure.
  4. What broke. Your first failure rate on examples you hadn't seen, what caused the failures, what you changed, and the new rate.
  5. The number. Minutes saved, errors caught, or cost to run, with your assumptions written out.
  6. What's next. One or two improvements you'd make with more time.

Spend the most space on what broke

This feels backwards. Your instinct is to hide failures. But the failure section is the one that shows judgment, and judgment is what people are hiring for. Anyone can show a demo that works. Few people can show that they found the cases where it didn't, understood why, and fixed it.

A good example: "On the first run it got 6 of 10 hidden invoices right. Most errors were dates in European format. I added an instruction to return dates as YYYY-MM-DD and to flag ambiguous ones. Second run: 9 of 10 right and 1 correctly flagged for review."

Getting an honest number

The simplest honest formula is minutes saved per item, times items per week, minus anything the tool costs to run. State your assumptions out loud: "Assuming 3 minutes per invoice and about 120 invoices a week." A modest number you can defend beats a big one you can't.

Where to put it

Put every case study in one place you can send as a single link: a simple web page, a Notion page, or a live project dashboard. Then link it from the top of your LinkedIn profile and your resume. When someone asks what you do, send the link instead of describing it. The Command Center Kit is the dashboard I use for exactly this.