48% reduction in procurement cycle time across 14 categories

Context

A North American mid-market manufacturer was processing upward of 4,000 purchase orders per month across 14 spend categories. Every PO moved through the same manual sequence — invoice matching, exception review, approval routing — and approvers, AP staff, and managers were collectively the bottleneck. On average, 3 to 5 days of every purchase cycle had no value-added work in it.

The approach

The engagement started with the exception taxonomy, before any agent was built. Every exception type the AP team handled — price variances, quantity mismatches, duplicate invoices, missing PO references, vendor disputes — was documented, classified, and assigned a resolution path with an explicit escalation threshold. Below the threshold, the agent resolves autonomously. Above it, the case routes to a human with a pre-packaged summary. That taxonomy became the governance layer the rest of the system was built on.

The invoice processing agent reads incoming invoices in PDF and EDI format, extracts line-item data, and matches against open POs in the ERP using fuzzy matching to handle inconsistent vendor formatting. For clean matches within tolerance, it triggers payment release without human involvement. For mismatches, it classifies the exception type and routes accordingly. No AP staff member touches a clean invoice.

Approval routing was rebuilt as a structured workflow agent rather than an email chain. The agent reads the approval matrix — category, vendor tier, and spend threshold determine the approver — and routes each item with full context attached: PO history, vendor performance data, and prior exceptions on the same vendor. Approvers see what they need to decide, not a raw attachment. Median approval time fell from 2.1 days to 4 hours.

The exception handling agent manages the cases that previously sat in AP team queues. For exception types covered by the taxonomy, it runs the resolution path end-to-end — contacting vendors, requesting documentation, updating the ERP — and escalates only when the path can't proceed without human judgment.

Results

48% reduction in procurement cycle time

3–5 days eliminated from every purchase order

14 spend categories fully automated

78% of exceptions resolved without human involvement

AP headcount: unchanged

What made it work

Most procurement automation projects treat exception handling as the long tail — something to address after the clean-path automation is running. This engagement treated it as the primary design constraint. Building the exception taxonomy first gave the agents their escalation criteria, gave the governance layer its structure, and gave the AP team a clear picture of what the system would and wouldn't handle before it touched live transactions. Without it, the agents would have been operating in a gray zone that produces exactly the kind of unexpected behavior that ends automation deployments.