Process Audit: Accounts-Payable Invoice Handling
This is a complete sample of the audit document every OneTPM engagement produces, prepared for a fictional mid-market logistics company. We publish it so you can judge the work before you pay for it. Names and figures are illustrative.
§1Executive summary
Meridian's AP team of four processes roughly 1,900 invoices per month, arriving by email as PDFs from around 340 vendors. Each invoice is manually keyed into the ERP, matched against purchase orders, and routed for approval. We observed the process end to end, timed a sample of 60 invoices, and interviewed all four processors.
Recommendation: automate extraction, PO matching, and routing, with human review gates at two points: any extraction below the confidence threshold, and every payment approval. We estimate the build at six weeks. We advise against automating vendor-exception handling in the first phase (see §6) — the volume doesn't justify the complexity yet.
§2The process as observed
The documented procedure has five steps. The actual process, observed at the desks of the people running it, has nine — including three workarounds that exist because of a 2022 ERP migration and one step that exists only in the head of the most senior processor.
| Step | As documented | As observed |
|---|---|---|
| Intake | Invoices arrive in the AP inbox | Three inboxes; one processor forwards from a legacy address vendors still use |
| Entry | Keyed into ERP | Keyed into a spreadsheet first, then the ERP, because the spreadsheet is the de facto tracking system |
| PO match | Automatic in ERP | Automatic match fails on ~30% of lines; failures resolved by memory or by asking Priya |
| Approval | Routed by amount threshold | As documented, but approvers sit on batches until Friday |
The spreadsheet is the finding that matters. Any automation that ignores it will be worked around, exactly as the ERP is today.
§3Volumes and effort
| Measure | Observed | Note |
|---|---|---|
| Invoices / month | ~1,900 | Seasonal peak near quarter end |
| Median handling time | 6.5 min | Timed sample, n=60 |
| Exception rate | ~30% | Dominated by PO line mismatches |
| Exception handling time | 18 min median | High variance; depends on vendor |
| Estimated team effort | ≈ 380 hrs / month | Roughly 55% of the AP team's capacity |
§4Failure points
- Single point of knowledge. PO mismatch resolution depends on one person's memory. Her calendar is the process bottleneck.
- Shadow system. The tracking spreadsheet is unbacked-up, unaudited, and load-bearing.
- No intake control. Three inboxes mean no single count of what arrived, so "did we get that invoice?" takes twenty minutes to answer.
- Approval batching. Weekly approval batches add a median of four days of latency the automation alone will not remove — this is a policy decision for Meridian, flagged in §6.
§5Recommended architecture
A single intake pipeline replacing all three inboxes; LLM extraction into the ERP schema validated against PO data; automatic filing above a confidence threshold tuned during the pilot; a review queue for everything else, designed around the current spreadsheet's columns so the team's working habits transfer instead of breaking.
- Extraction and matching logged per field, with corrections feeding the validation rules
- Mismatch-resolution knowledge captured as rules during the build, with Priya as the reviewing expert — converting the single point of knowledge into a maintained asset
- Payment approval remains fully human, unchanged by design
§6What we advise against
- Don't automate vendor-exception handling yet. Twelve vendors cause most exceptions. A quarterly vendor cleanup is cheaper than the automation.
- Don't buy an "AP automation suite" for this volume. At 1,900 invoices a month the per-document pricing of the major suites exceeds the cost of the build within two years, and the suite won't absorb the spreadsheet workflow.
- Decide on approval batching separately. It is culture, not tooling.
§7Proposed build
| Phase | Weeks | Deliverable |
|---|---|---|
| Intake consolidation + extraction pilot | 1–2 | Single pipeline, live on a 3-vendor subset |
| Validation + review queue | 3–4 | Full vendor set, confidence thresholds tuned with the team |
| Enablement + handover | 5–6 | Runbooks, admin transfer, team operating independently |
Success criteria, agreed before build: the AP team operates the system without OneTPM present; intake is answerable from one place; exception handling no longer depends on a single person.
End of sample. A real audit also includes the interview notes, the timing data, and an appendix mapping every field in the extraction schema — omitted here for length.