This is the only named case study I publish. It's here because the professor involved gave explicit permission, and because the lesson inside it has re-applied itself in four subsequent engagements.
Short version: a university professor was about to renew a $50,000/year software contract. We built her a custom AI platform in six weeks that did the same core work, cost a one-time fee under five figures, and belongs to her department forever.
The problem
The department used a specialized piece of software for a repeating analytical workflow — intake a dataset, categorize against a research taxonomy, produce a structured report. The vendor was fine. The software worked. The problem was scale: the department had grown past the vendor's small-team tier, and the renewal quote tripled.
Worse, the department's actual use case had drifted from what the tool was built for. Two-thirds of the feature set was unused, and the one feature they most needed (a custom taxonomy specific to their research) required a consulting engagement on top of the license.
The approach
- 01Weeks 1–2: Mapped the actual workflow. Not the software — the work. Found that 80% of the value came from one stage (classification) that a modern LLM could do with the right prompting and the department's own taxonomy.
- 02Weeks 3–4: Built a Claude Project with the full taxonomy, 40 labeled examples, and a structured output format matching the team's downstream needs. Tested it side-by-side against the existing tool on historical datasets.
- 03Weeks 5–6: Wrapped it in a small web interface the professor's grad students could use without any prompting knowledge, wrote the documentation, trained three people, and handed over every credential.
The outcome
The new workflow matched the old tool's accuracy on their real datasets within the first two weeks. By week six, it was better — because the taxonomy and examples were theirs, not a generic off-the-shelf model.
The department cancelled the renewal. One-time build cost landed under $15,000. Ongoing cost is a Claude subscription plus the grad students' time, which they were paying for anyway. They estimate five-year savings above $200,000, not counting the consulting fees they would have paid the vendor for custom work.
More importantly: the professor and two grad students can maintain and extend the platform themselves. When the taxonomy evolves, they update it. When a new research question comes up, they add examples. I'm not involved, and that's the point.
“The software wasn't bad. It just wasn't ours.”
What this is not
This isn't a case for replacing all your software with AI. Most software is well-priced, well-built, and doing work AI isn't ready for.
It's a case for noticing when the economics and the fit have drifted — and asking the simple question: what would it cost to build just the part we actually use, and keep it?
For some workflows, the answer is "a lot, don't bother." For others, it's surprisingly small. Worth asking.
Want to talk about where this fits for you?
Thirty minutes. No pitch, no slides. Tell me where your team is. I'll tell you honestly whether I can help.
Schedule Your Free Intro Call