Why AI agents break in enterprise finance: 25% of invoices lack a PO

vasuman · x · 2026-11-05

Vasuman (founder of Varick Agents) catalogs real-world failure points where AI agents hit walls in large-enterprise finance workflows, based on firsthand audits:

Missing POs: 25%+ of invoices arrive without a purchase order; coders just check "no PO" in Coupa and it gets paid unchecked — sometimes for the same vendor six months running.
Free-form tickets: Partner billings start as free text; one ticket misstated a $360K credit as $3M — and the customer was paid that amount.
Missed deadlines: AR needs close data by business day 2 but gets it on day 4, forcing revenue accruals.
Manual asset retirements: 500-700/month, each requiring manual typing of three data columns in the ERP; one wrong digit retires the wrong asset.
Dirty master data: Every new invoice group spawns a new customer account; the CRM integration can't be turned on, and nobody knows the B2B customer count.
Rigid approvals: 11 chart-of-accounts segments each routing to different reviewers, thresholds that add approvers but never remove them, 2 days to 2+ weeks per request.
Fragmented systems: Intake/approval and ERP coexist; users don't know which is the source system.
Legal redlines strip metadata that flags non-standard discounts, caught only when the system recalculates.
Accruals via email: PMs are asked monthly about unbilled work over $5K, then reconciled in Excel.

The takeaway: enterprise agent failures stem less from model capability than from broken processes — missing POs, manual clicking, approval stacking, and messy master data that must be fixed before automation can work.

Original post →

Why AI agents break in enterprise finance: 25% of invoices lack a PO

More from coding & agent