Accounts payable is absorbing roughly a full-time equivalent in keying alone. This report prices the fix, names the one thing that must be fixed first, and ships with a working agent already scored against the firm's own ledger.
The audit reviewed 60,241 invoices across 8 entities and 340 vendors, then replayed an AP intake and coding agent in shadow mode over the most recent 90 days — 8,214 invoices, 25,726 coding decisions, every one scored against the GL code a human actually keyed. Because the window is historical, this is a backtest, not a demonstration.
The straight-through rate rises from 39% in the opening week to 60% in the closing one. Most of that gain lands in the first month, as the model absorbs reviewer corrections on the vendors it was least sure about; it then flattens through April before grinding upward again. Ninety days is a trajectory, not a trend line. The playbook assumes 70% in year one — this window is consistent with reaching it and does not on its own prove it.
Scored 1–5 across five dimensions. A low score does not stop the engagement — it changes what gets built and in what order. Ownership scores 2 here, and that single score is the reason one of the five opportunities in §4 is a recommendation against building.
Header and line fields are present on effectively every posted invoice across all eight entities. The gap is purchase-order linkage: a material share of invoices from PO-required vendors arrive with no PO reference at all, which is what forces manual matching.
Fix: Enforce PO reference capture at intake for the vendor set that requires it.
GL coding is highly repeatable within a vendor, which is why a model can learn it. The vendor master itself carries near-duplicate records, inconsistent legal-name formatting and stale terms, so the same supplier can appear more than once.
Fix: De-duplicate the vendor master and adopt a single naming standard before go-live.
QuickBooks Online exposes the required API surface, the shared AP mailbox is accessible via Microsoft Graph, and all four supplier portals support CSV export. No system replacement or middleware purchase is required.
Portal and email invoices are available same-day. Mail-received invoices sit in the scanning queue for several business days before they are visible, which is a direct cause of missed early-pay windows.
Fix: Move scanning to daily and OCR at the point of scan.
Vendor records are created by whoever onboards the supplier, across both the firm's own books and client accounts, with no review step and no accountable owner. Banking detail completeness is inconsistent. This is the one dimension that materially constrains what should be automated.
Fix: Assign a named vendor-master owner and a two-person change control on banking detail. This is a prerequisite, not a nice-to-have.
Every claim carries its evidence reference in the margin — I‑n for an interview, A‑n for an artifact, S‑n for a system walkthrough. Hover any reference for its source. Measured values are queries against the client's own data, not figures typed into prose; the query is printed beneath each reference.
A time study over 60 observed invoices puts average manual touch at 3.2 minutes. Against measured volume of 2,738 invoices per month across 8 entities, that is 148 hours of AP labour every month — about eight tenths of a full-time role spent moving data from PDFs into QuickBooks.
6.2% of posted invoices required a post-posting correction. 47% of those corrections trace to a wrong GL code rather than a wrong amount. That matters for build design: coding is a learnable, high-signal problem — the same vendor codes the same way most of the time — which is exactly where a model earns its keep.
25% of invoices arrive as scanned PDFs with no OCR layer. They take 48 seconds longer to key than clean digital documents and carry 1.1x the rework rate. They also sit in the scanning queue for several business days, which is a direct cause of the missed discount windows in F-06.
Reconciling six vendor statements against posted invoices identified genuine duplicate submissions, of which 46 were paid twice over the 24-month history, $85,283 gross. Some is recoverable through vendor credits; the report counts only the portion that historically is not. Detection today depends on a clerk recognising the invoice.
340 active vendor records are maintained across the firm's own books and seven client accounts with no review step, no naming standard and no change control on banking detail. This scores 2/5 on the data-quality rubric. It is the reason this report recommends against automated payment release (OPP-5) regardless of how well the coding model performs.
Of the early-pay discount dollars available, 78% were taken. The misses are concentrated in small-dollar office-supply and freight invoices where nobody is watching the clock — roughly $7,899 a year left on the table. This is a processing-speed problem, not a policy problem.
2.2% of invoices were paid after terms, drawing roughly $3,123 a year in fees. The firm is not systematically late; the exposure sits with invoices that stall waiting for an approval that is being chased over email.
Controller-tier approvals take a mean of 35 hours to come back, all of it conducted over email threads with no state. The time does not show up in the touch-time study because nobody is touching the invoice — it is simply waiting.
13% of invoices from vendors the firm designates as PO-required arrive with no PO reference. Each one becomes a manual hunt. This is a capture-at-intake fix, and it is cheap.
Completeness, accessibility and freshness all score 4/5. Every system the build needs is reachable today with no replacement and no middleware purchase: QuickBooks Online exposes the API surface, the mailbox is reachable via Graph, and all four supplier portals export CSV. On this evidence we recommend proceeding — conditional on the vendor-master ownership fix in OPP-2, which is scoped and priced separately at two weeks.
Prioritised with prices and effort. The fifth entry is a recommendation against building. That is policy, not personality: where the data-quality rubric says the controls are not in place, the report says so and prices the fix first.
(a) process automation
Capture invoices from the shared mailbox and supplier portals, extract header and line data, suggest GL coding from the firm's own history, match to PO, and route exceptions to a human queue. Nothing posts unreviewed above the confidence threshold.
Why this rank: Highest measured value, lowest integration risk, and the data quality supports it today. This is the build the ROI model prices.
(a) process automation
De-duplicate vendor records across all eight entities, standardise legal naming, backfill terms, and stand up two-person change control on banking detail.
Why this rank: Prerequisite. The coding model is only as good as the vendor it resolves to, and the ownership score of 2/5 is a live fraud exposure independent of any automation.
(b) custom AI agents / copilots
A read-only portal for client-account contacts showing invoice status, expected pay date and exception state, with a natural-language query surface over their own AP.
Why this rank: Real value and a genuine differentiator for the firm's outsourced-AP offering, but it depends on OPP-1 being live first. Sequence it second, not first.
(a) process automation
Extend the extraction pipeline to employee expense reports and receipt images, with policy checks and mileage validation.
Why this rank: Sound idea, wrong order. Volume is roughly a twentieth of AP, so the payback is far longer. Revisit once the AP pipeline is stable and the extraction models are tuned.
(b) custom AI agents / copilots
Straight-through payment execution: approved invoices released to the bank without a human in the payment path.
Why not: Do not build this. Vendor-master ownership scores 2/5 and banking detail has no change control, so automated release would remove the last human check standing between a compromised vendor record and money leaving the building. Fix the ownership finding first; even then, we would advise keeping a human release step. We would decline this scope.
Every line is a measured basis multiplied by a stated rate. No dollar figure below is hardcoded, and each row names the population it was measured over and the query that produced it. Intervals are nonparametric bootstraps resampled at the invoice level — rare-event drivers resample the full 24-month population rather than annualising a 90-day window, because the sampling error on a handful of events would otherwise swamp the estimate.
| Line | Basis | Measured | Rate | Annual | 95% CI | Playbook | Measured over |
|---|---|---|---|---|---|---|---|
| D-1 AP touch time released ap.v_savings_components -> saved_seconds | Net AP hours released per month (gross saved less 10 h/mo supervision) | 87.5hours/month | $38.00 | $39,912 | $39,375 – $40,463 | $41,000-3% | 90-day replay windown = 8,214 |
| D-2 Rework avoided ap.v_savings_components -> rework_prevented | Correction hours avoided per month (10 min per corrected invoice) | 21.6hours/month | $38.00 | $9,856 | $8,876 – $10,812 | $8,500+16% | 90-day replay windown = 8,214 |
| D-3 Early-pay discounts captured + late fees avoided ap.v_savings_components -> discount_*/late_fee | Net dollars per month over 24 months at 93% discount capture and 80% fee avoidance | 656.3dollars/month | — | $7,876 | $7,152 – $8,591 | $7,000+13% | 24-month historyn = 60,241 |
| D-4 Duplicate payments prevented ap.v_savings_components -> duplicate_caught | 24 months of double-paid invoices at 88% measured agent recall, 55% never clawed back | 1,710.1dollars/month | — | $20,521 | $12,988 – $29,930 | new | 24-month history x measured recalln = 46 |
| Playbook-comparable total (D-1 … D-3) | $57,644 | $56,500 | |||||
| Including the new finding (D-4) | $78,166 | payback 10.4 months on $68,000 net | |||||
These are the numbers we chose rather than measured. They are listed so the client can argue with them — which is the point. The posture is deliberately mid-range: supervision overhead is deducted before a dollar is claimed, a reviewed invoice is charged half its original touch time, and a blocked exception is charged more than the manual baseline.
| Task | Mean | SD | n | Ref |
|---|---|---|---|---|
| Open, read and identify vendor | 41s | ±14s | 60 | I-2 |
| Key header fields into QuickBooks | 52s | ±19s | 60 | I-2 |
| Key and code line items | 63s | ±31s | 60 | I-3 |
| Match to PO / chase missing PO | 19s | ±28s | 60 | I-3 |
| Route for approval and follow up | 17s | ±22s | 45 | I-4 |
| Total manual touch | 192s | 3.2 minutes per invoice | ||
| Assumption | Value | Basis |
|---|---|---|
| AP loaded rate | $38.00/hr | $38/hr fully-loaded AP cost — payroll + benefits + overhead |
| Supervision overhead | 10 h/mo | Daily exception digest and monthly accuracy review, deducted from gross savings before any dollar is claimed |
| Auto-post confidence threshold | 0.88 | Lines below this route to a human; nothing posts unreviewed |
| Reviewed-invoice cost | 50% | A reviewed invoice still costs half the original touch time |
| Blocked-invoice cost | 115% | An exception costs more than the manual baseline — counted against us |
| Discount capture with agent | 93% | Portfolio rate, against the measured human rate |
| Late-fee avoidance | 80% | Not 100% — some late payment is a cash-timing choice, not an error |
| Duplicate recovery haircut | 55% | Only the share never clawed back by vendor credit is counted |
| Rework correction time | 10 min | Locate, reverse, recode, and call the vendor where needed |
| Bootstrap | 2,000 × 95% | Nonparametric, resampled at the invoice level; rare-event drivers resample the 24-month population |
| Ref | Source | Captured |
|---|---|---|
| A-1 | Chart of accounts export33 active expense accounts across 9 categories, exported from QuickBooks. | 2026-06-16 |
| A-2 | 24-month AP transaction exportFull invoice, line, payment and approval history for all eight entities. | 2026-06-17 |
| A-3 | Vendor master exportActive vendor records with terms, default coding and banking-detail completeness. | 2026-06-17 |
| A-4 | Correction / journal-entry logPost-posting corrections with reason codes, used to derive the rework rate. | 2026-06-18 |
| A-5 | Vendor statement reconciliationsSix vendors' statements reconciled against posted invoices to test for double payment. | 2026-06-19 |
| A-6 | Payment register & terms schedulePayment dates against terms, used to measure discount capture and late-fee exposure. | 2026-06-19 |
| I-1 | Managing PartnerScope, priorities, approval authority and the firm's appetite for automation in client-facing work.L. Marchetti | 2026-06-15 |
| I-2 | ControllerEnd-to-end AP walkthrough; observed 60 invoices keyed live for the time study.S. Adeyemi | 2026-06-16 |
| I-3 | AP ManagerCoding conventions, exception handling, month-end close pressure and the correction workflow.J. Brantley | 2026-06-16 |
| I-4 | AP Clerks (2, joint)Day-in-the-life; approval chasing, supplier portals, and the scanned-mail queue.R. Ellison, T. Nakamura | 2026-06-17 |
| I-5 | Client Accounting LeadHow client-account AP differs from the firm's own books; vendor onboarding.D. Okafor | 2026-06-18 |
| I-6 | IT / Systems AdministratorQuickBooks access model, mailbox routing, retention, and existing integrations.M. Villanueva | 2026-06-18 |
| S-1 | QuickBooks OnlineRead-only walkthrough of AP entry, approval and posting; API availability confirmed. | 2026-06-16 |
| S-2 | Shared AP mailbox (Microsoft 365)Invoice intake volumes by sender and attachment format. | 2026-06-17 |
| S-3 | Supplier portals (4)Portal-sourced invoices and CSV export capability reviewed. | 2026-06-17 |
| S-4 | Scanning workflow / network folderMail-received invoices scanned to a shared drive; no OCR layer in place today. | 2026-06-18 |
| S-5 | Approval email threadsSampled approval chains to measure turnaround by threshold tier. | 2026-06-18 |
Generated 2026-08-05T17:30:52+00:00 from snapshot findings-snapshot-latest.json ·
60,241 invoices · 181,758 lines · 8,214 replayed ·
25,726 coding decisions scored
Every figure in this report is a query against PostgreSQL 16; no value is typed into prose.
Synthetic throughout — a modeled archetype, not a client.