A global corporate-services and fund-administration firm ran outsourced bookkeeping for thousands of end-clients across multiple countries — each invoice keyed by hand and coded to that client's own chart of accounts, across several accounting systems. The work was expensive to scale, and every reviewer's judgment evaporated after the close, so each similar invoice became a fresh review task.
~95%
steady-state auto-approval
~200 hours
reviewer time returned / close
1 week
to 100% user adoption
Prosights installed an AI operating layer over the firm's existing ledgers: it extracts every field with a citation to the exact spot on the source, applies the client's COA memory, and clears the routine majority under stated policy while routing only genuine exceptions to a human. Each correction is saved as a versioned, client-scoped rule and applied to the next batch — so auto-approval compounds close over close.
The Challenge
Outsourced bookkeeping at scale meant manual data entry and per-client COA coding across countries and systems — high cost to serve, and judgment that disappeared after every close, turning each similar transaction into another review.
The Approach
Not another tool in the stack — an operating layer across the existing ledgers. AI absorbs the document volume; a policy gate (thresholds, sum checks, codes restricted to the client's own COA) decides where human judgment is spent; everything released is logged with evidence.
The Build
Extraction with cell-level citations boxed on the source PDF; a client-scoped COA memory that versions every coding decision; adoption inside Excel and a simple web app with no firm-wide UI rollout; and a KPI layer tracking auto-approval rate, reviewer hours returned, exception mix, and decision reuse.
The Outcome
~1,000 documents cleared to ~6 human decisions per batch, ~95% steady-state auto-approval, 100% of intended users live within a week — and a governed, exportable memory the firm owns. Judgment compounds instead of evaporating.
