What decision does this actually settle?
Lakebridge Home & Hospitality is a fictional $410M distributor closing the books across 23 branches, with 4,680 invoices a month competing for eight AP reviewers. Every invoice starts as an untrusted document until someone reconstructs the purchase — matches it to a PO, a receipt, and the right GL line. The model doesn't decide whether an invoice gets paid. It decides which invoices already agree with themselves and can post without a person, so the eight reviewers spend their time on the ones that don't.
From an open batch to a closed exception
Read the case brief first
Day −5 through day +4 in the story above the workbench is the same timeline the invoice stream below replays.
Run the next batch
Each batch pulls in new invoices and auto-matches the ones with a clean invoice / PO / receipt / GL agreement.
Filter by status
All / Auto-posted / Review / Blocked. Auto-posted needs nothing from you — it's the 82% baseline the case brief describes.
Open an exception, e.g. Peakline Equipment INV-84932
The detail panel shows exactly which field mismatched — quantity, price, duplicate number, or late receipt — with the source documents side by side.
Resolve it or hold the payment
Request a receipt correction, or keep the payment on hold. Either way, a person authorizes the release — the model never pays anything.
What each part of the workbench is for
| Area | What it controls | Where the decision happens |
|---|---|---|
| Stream panel | The live invoice queue, filterable by match status. | Pick which exception to open — auto-posted items don't need you here. |
| Detail panel | The specific mismatch: which field, which document, which branch. | Request a correction or hold the payment. |
| Download exception packet | Exports the open exceptions with their evidence. | Take this into the close meeting instead of a raw count. |
Three numbers, in order of usefulness
| Metric | What it means | What to do about it |
|---|---|---|
| Days to close | 11.0 days baseline across 23 branches and 4,680 invoices a month. | Track which branches or vendors are still blocking accruals at day 9, not just the final count. |
| Processing cost | $14.80 per invoice, blended across auto-posted and manually reviewed items. | Compare it against the cost of the exceptions alone — that's the number automation actually moves. |
| Match status | Auto-posted, review, or blocked — driven by the 3-way match against PO, receipt, and GL policy. | Blocked items are the ones with a real dollar or timing risk; review items are usually just missing a field. |
Three questions this actually settles
Run a batch and split by status — the auto-posted share is the honest answer, not an estimate.
Filter to Blocked and read the detail panel — each exception carries the exact mismatch, not just a category.
The exception packet export carries owner and branch on every line, which a raw invoice count never did.
Where this breaks if you push it too far
What goes in, what comes out
| Detail | |
|---|---|
| Input | A pre-loaded illustrative invoice stream across 23 branches — no upload. |
| Agent output | Capture, coding, 3-way match, and an exception packet with evidence attached. |
| Stays with a person | Payment release, vendor-master changes, close certification. |