We built their AI. Then taught them to build.

The situation

Every company has one. The person who's been there forever and knows everything — the one everyone messages when a client asks something hard. For this cross-border real-estate firm, that person had held the knowledge for two decades. It worked, right up until it didn't scale: answers were slow, they varied across languages, and they all funnelled through a single human.

The challenge

Get that knowledge out of one head and into the hands of the whole team — so anyone could answer a client email accurately, in the right language, fast — and do it without leaking personal data along the way.

What we built

• A knowledge assistant, built overnight. It reads the incoming client email, strips personal data before anything else happens, searches a knowledge base of 261 real answers, and drafts a grounded, cited reply in the client's language — in the company's voice, with a review step for anything sensitive. We hardened it with the team in a workshop the next day.
• A training program, so they don't need us. Rather than hand over a black box, we taught seven of their people to build the way we do — with Claude Code and spec-driven development, the write-the-spec-first method that produced the assistant in the first place. They left able to extend it themselves.

What changed

• Twenty years of expertise became a searchable, multilingual base the whole team draws on — grounded in real answers, not guesses.
• Seven people can now build, not just use. Each saves an estimated 7–11 hours a week, and the firm can make the next change without waiting on us.
• It proved a pattern: an overnight prototype plus a hands-on workshop is a fast, low-risk way to show a team what AI does — and leave them able to keep going.