case study · 02 · published 2026-07-04
A CRM of record, generated document packs, e-signature split by legal weight, and one narrowly-scoped Claude agent that reads compliance documents and flags gaps - but never fills, corrects, or clears anything. Built in a one-week triage sprint and a four-week fixed-fee build.
Published anonymized at the client's request. All details verified; the client reviewed this text before publication.
Every document pack was assembled by hand: copy the last similar contract in Word, overwrite the fields, hope. The predictable failures followed - a price or buyer name carried over from the previous deal; the hoja de visita signed on perhaps six viewings in ten, leaving commission disputable on the rest; an expired certificado energético or a nota simple that didn't match the seller surfacing at the notary's office instead of at intake.
With roughly a third of transactions involving international buyers, the document and compliance load multiplied - NIE, source-of-funds, bilingual packs - while deal state lived in inboxes, WhatsApp and a shared drive. Assembling and chasing one complete pack cost around two hours. AML checks, which the agency is legally obligated to perform, ran inconsistently and left no audit trail. The core issue wasn't volume. It was that nothing was verified and nothing was traceable.
A one-week paid workflow audit mapped the document flow; the build followed at a fixed fee. Trigger to done:
Stack: Attio, n8n, Claude (Anthropic API) as the single agent, PandaDoc and DocuSign, the Idealista feed, Holded. Claude runs under a no-train default with a full audit trail; buyer PII and AML data stay on EU infrastructure. The gap-checker was checked against labelled past packs with known gaps before go-live.
“It made us faster without ever guessing. Every field on a contract traces back to a document, and anything the system isn't sure about lands on someone's desk instead of in the paperwork.” Director · independent residential agency · Comunidad Valenciana
What was deliberately not built:
Underneath all of it, one hard rule: no un-sourced value reaches a contract - every merged field traces to a record, every flag to a document.
Reported by the client after go-live, approximate figures, not an independent measurement: assembling and checking a document pack went from roughly two hours of manual work to about twenty minutes of review.
Most consequentially, compliance gaps now surface at intake instead of at the notary; the client reports no signing has stalled on an unchecked document so far. The hoja de visita signs at every viewing, and contract fields merge from verified records rather than copied contracts.
The agency didn't buy artificial intelligence; it bought verified paperwork and traceable decisions. The one place a language model earned its seat - reading documents against a checklist and citing what it found - is narrow, measured against an eval set, and stripped of authority to act. Everything legally weighty kept a human signature over it, and everything routine stopped consuming human hours.
It is the same pattern as the first case study, in a different industry - which is rather the point. Fixed-fee, source delivered at handover, owned by the client team, with a 30-day open line after go-live.