Back-office decision model · Finance 02

The invoice arrived.
The evidence did not.

Lakebridge Home & Hospitality closes 23 branches through one finance team. Every month, 4,680 invoices compete for the same eight AP reviewers—including the 82% that never needed a person.

Open the close workbench Read the case

// fictional company · illustrative data · no upload · runs in your browser

01 · Capture

Extract invoice fields with source-page evidence.

02 · Reconcile

Match invoice, purchase order, receipt, vendor, and GL policy.

03 · Route

Post clean items; send only exceptions to accountable owners.

Use-case story

Eight reviewers were
checking every receipt.

Lakebridge is a fictional $410M multi-location distributor. Its branches buy freight, equipment, labor, packaging, utilities, and local services. The problem is not a lack of invoices. It is that every invoice begins as an untrusted document and stays that way until a person reconstructs the purchase.

01 / DAY −5

PDFs arrive without context.

Branch mailboxes forward 260 invoice formats. Vendor names vary, PO numbers are missing, and one invoice can contain freight, equipment, and tax lines with different GL treatment.

02 / DAY −2

Six hundred exceptions appear at once.

AP discovers quantity differences, price drift, duplicate invoice numbers, and goods received after cutoff. Buyers and warehouse leads answer through email, not a shared evidence trail.

03 / DAY +4

The close waits for AP.

Finance cannot finalize accruals while unmatched receipts and late invoices remain unresolved. The controller gets a count, but not which owners, vendors, or branches are blocking the ledger.

04 / WITH THE AGENT

Normal work becomes invisible.

The agent captures, validates, matches, codes, and posts clean invoices. Exceptions arrive with the exact mismatch, source documents, owner, and recommended action already assembled.

Autonomous invoice stream & close cockpit

Move the policy. Recalculate the close.

Adjust volume and autonomy, filter the invoice stream, inspect evidence, and resolve simulated exceptions. Figures are calculated from the fictional assumptions below.

Month-end close

3.2 daysfrom 11.0 business days

Cost per invoice

$2.10from $14.80 · Base policy

Monthly cost avoided

$59,436gross processing model

Human exceptions

84218% of 4,680 invoices

AP capacity returned

523 hrper month · labor-time model

What the number means

A faster close is a queue equation.

This simulator does not treat AI accuracy as the business result. It models how invoice volume, straight-through rate, review time, and control policy change the work remaining for finance.

Boundary: The agent prepares postings and exception decisions. Final payment release, vendor-master changes, policy overrides, and close certification remain authorized human actions.

Model input	Why it matters	Base value
Monthly invoices	Documents entering AP across 23 operating branches.	4,680
Current processing cost	Illustrative blended labor, rework, and system cost.	$14.80 / invoice
Straight-through rate	Invoices posted without AP review under the selected policy.	82%
AI-assisted processing cost	Illustrative blended cost after capture, matching, and exception routing.	$2.10 / invoice
Manual versus assisted time	Current 8.4 minutes versus 1.7 minutes of residual AP work.	523 hr returned
Software and implementation	ERP integration, security review, change management, and product price are unknown.	Not modeled

The operating change

Finance closes the exception.
The agent clears the queue.

A credible pilot starts with one branch, two high-volume vendors, and historical invoices with known outcomes. Measure field accuracy, false auto-posts, exception precision, AP minutes per invoice, and reconciliation breakage before allowing any live posting or payment release.