AI in AP without losing control
A controller’s guide to adopting AI in accounts payable while keeping decisions, validation and accountability clear.
Adopt the technology, keep the signature. The controls framework is inside.
Complete the form to get your copy.
Every AP vendor now claims artificial intelligence. Almost none of them tells controllers what they most need to know: how the technology fails, and what controls keep those failures out of the ledger.
This paper takes a controls-based middle position. “AI is too risky for financial data” ignores real evidence — top models now exceed 96% zero-shot extraction accuracy on open invoice datasets. “The vendor said it’s 99% accurate, so we’re fine” ignores equally real evidence — the same models can fabricate plausible-looking numbers, and a field returned at 0.95 confidence can still be wrong. The position throughout: the controller, not the algorithm, signs.
In this whitepaper
- The failure modes controllers should plan for, and why a model’s confidence score is not an accuracy guarantee
- How to design human oversight into the workflow: ground, validate, gate, record
- What the fraud evidence says about where humans must remain — vendor-master changes, banking details, payment release
- The centerpiece: a crosswalk mapping the five COSO internal-control components to concrete AP-AI controls, drawing on COSO’s 2026 generative-AI guidance, the NIST AI Risk Management Framework and ISO/IEC 42001
- A fourteen-question vendor-diligence checklist and an implementation approach that keeps controls intact
- 96%+zero-shot field-extraction accuracy of top multimodal models on open invoice datasets (academic benchmarks)
- 5COSO internal-control components mapped to concrete AP-AI controls
- 14vendor-diligence questions to put to any AI-AP platform
