A regulated non-profit automates expense compliance across multiple funder reporting standards
The situation
A regulated non-profit with multiple institutional funders reviews around 1,500 transactions per reporting cycle by hand. Each funder has its own reporting taxonomy, documentation requirements, and chart-of-accounts rules. The team classifies spend against those taxonomies, verifies documentation sufficiency by spend type, checks dates against contract boundaries, and reconciles arithmetic from receipts up to bank proof-of-payment. The annual audit is conducted by a Big 4 firm and findings carry real consequences with the funder. The process worked but was slow, expensive in staff time, and inconsistent across reviewers.
What we delivered
We designed and built a three-phase QC pipeline. A pure-RPA consolidation layer pulls source documents from accounts payable and accounting systems and merges them with proofs of payment. A structured-extraction layer handles document classification using a tightly scoped LLM step. A QC agent then applies a versioned, config-driven rule library covering documentation sufficiency, nature-over-purpose categorization (a flight is Travel even when the purpose is Training), out-of-contract date checks, competitive-bid threshold requirements, FX reconciliation, and arithmetic validation across nested layers up to five deep. Every flag traces to a named rule ID. The agent flags issues for human review; it never modifies records on its own.
The result
The prototype was validated against historical data and reproduced the flags human reviewers had caught, while surfacing additional documentation gaps the manual process had missed. Production scoping is in progress. The architecture delivers two properties that matter for funder-reportable work: the AI layer is sandboxed to the work that actually requires language judgment, and every decision is traceable to a named rule and a source quote that an auditor can verify on demand.
Have a similar workflow eating senior hours? Book a scoping call. A 30-minute conversation tells you whether it is something we should build for you.