Rebuilding a Stalled Financial-Aid Estimator With Claude

The ask

A private career university hired us for an AI strategy. I ran discovery with four or five departments, read their documentation, and did vendor diligence.

One department had already done the hard work. They'd named their problems and picked a solution. I told them their plan was right and gave them a fallback if it didn't work out. The others needed more digging, especially finance, where how financial aid gets calculated was causing real problems.

Discovery turned up eight opportunities. The engagement focused on one: a financial-aid estimator for the admissions team. I wrote the product specs for the builds, and a short client-facing brief so everyone agreed on what we were making.

A side project for the CFO

Their new CFO kept getting requests to buy software, or kill it, with no business case behind them. Nobody had asked what the problem cost today, or what else might solve it.

Over a few calls I built him a business-case assistant that walks anyone through those questions. I tested it, he tested it, and after two rounds of feedback he rolled it out to his team. It wasn't in the contract.

What went wrong

I moved off the account. Months later, the estimator still hadn't shipped. The team building it and the client's own leaders weren't aligned on what it should do. The COO wanted one thing. The admissions team who'd use it every day wanted another.

The rebuild

I'd been developing a method for writing specs an AI agent can build from on its own, with a check the next morning that a real person can actually do the job. I needed a real test case. This was it.

It took about an hour and a half of my time: answering Claude's questions, reviewing, setting things up. Then Claude built the web app from scratch, on its own.

Two design choices mattered most:

• The math is code, not the model. Every dollar figure comes from a fixed calculation. The AI explains. It never computes aid.
• Sensitive data never leaves the page. The school didn't want Social Security numbers reaching any AI. Any SSN a user types is scrubbed in the browser before anything is sent.

What happened

I left Cadre around the time it was handed over. I believe the school used it, but I can't confirm that, so I won't claim it.

What I learned

Delivery rarely fails on the model. It fails when the people who'll use the tool and the people who approve it want different things. A good spec forces that argument early, before anyone writes code.