Merchant cash advance runs on documents. Bank statements, applications, stips, offers, contracts, and a broker chasing eleven funders at once. The industry standard is a generic CRM bent until it fits, plus a shared inbox, plus somebody's memory.
FUNDesk is the workspace built for the actual shape of the work: intake that reads the statements, a pipeline that knows what a submission is, and a paper trail underneath all of it.
The invariants
- Row-level security on 131 tables, enable plus force, with a dashboard that reports drift instead of assuming coverage.
- Money in integer minor units, inside a transaction, always. No floats anywhere near a balance.
- Whoever creates a payable cannot approve it.
- Sensitive attributes are filtered before anything reaches a model, from one canonical list.
- Every read of regulated data writes an append-only log entry.
Where the AI actually sits
Not in a chat bubble in the corner. It sits on intake, where a hundred-page PDF becomes structured fields, and on the parts of underwriting where a human is reading for the same six things every time. Every extraction is validated against a schema before it is allowed to become a record, and the source page is one click away. So far that path has extracted 10,662 offers and produced 618 underwriting reports.
What running it looks like
Twenty scheduled jobs, four of them tagged as money paths. The payment reconciler runs nightly and reports what it found rather than fixing it quietly: on the last run, 214 of 222 rows clean, zero healed, ten flagged for a human. Zero healed is the number that matters. A job that silently corrects a ledger is a job that hides the bug.
The nightly QA sweep has come back clean eight runs in a row against production. Real-user web vitals over 42,196 samples land good on all five: 2.12s LCP, 0.068 CLS, 136ms INP, 902ms FCP, 78ms TTFB.