When a business grows fast, the growth shows up on the revenue chart long before anyone notices what it did to the finance team. For one high-growth equipment and field services enterprise, rapid multi-branch expansion meant a steady rise in entities, vendors, purchase orders, and invoices, all landing on an accounts payable operation that was never designed to absorb that kind of volume. This case study looks at how the company put AP on autopilot, processing roughly 289,000 invoices in six months with AI that reads every field and a three-way PO matching layer that keeps hands-off posting safe, all without scaling the AP team to match.

Prior state and business challenges

The company's AP operation is about as complex as they come. Nearly every invoice, around 97%, is tied to a purchase order, which means the real work isn't just reading the document; it's matching each invoice to the right PO and goods receipt before it can post. On top of that sits a long-tail vendor base of rentals, parts, freight, and fuel, the kind of spend where almost no two invoices look alike and vendors bill in different formats and units.

Growth made operations even more difficult. Every new branch and acquired entity added vendors, approval paths, and volume. Eight legal entities eventually ran under a single finance organization, and invoice volume climbed to a peak of nearly 80,000 in a single month. The classic response to that kind of load is to hire, but headcount scales in a straight line while complexity compounds. Adding people wouldn't have removed the bottleneck; it would have moved it downstream.

When traditional automation stops scaling

Like most enterprises, the company had automation in place: OCR to read documents and a rules engine to route the clean invoices. For a while, that works. Templates catch the invoices that look the same every month and the repetitive work shrinks.

But rules-based automation quietly breaks down as a business grows. When a vendor changes their layout, bills in different units than the PO, or splits one order across several deliveries, each variation becomes an edge case, and a rules engine's only answer to an edge case is another rule. At enterprise scale, the rule library grows faster than anyone can govern it. 

Automation slowly turns into maintenance, and the exceptions still land on a person's desk. The ceiling wasn't a technology limit so much as a math problem: you can't write a rule for every way a growing business gets messy. The company needed a system that reads every field itself and learns the account, rather than waiting to be handed the next rule.

Evolving to an AI-first approach

The company moved to an AI-first AP model built around two engines working together. The first is an AI that predicts every field on an invoice, with no manual data entry, and keeps learning from the corrections people make. The second is a three-way PO matching layer that compares each invoice against its purchase order and goods receipt before it can post, gating risk so that clean invoices can flow straight through to the ERP with no human touch.

Crucially, the rollout was paced rather than flipped on all at once. Autopilot, the models processing an invoice end to end with no manual entry, was extended deliberately, climbing from about 2% of fresh invoices to nearly half as results came in and trust was earned in step with the evidence. Existing controls stayed in place: invoices under $5,000 auto-approve, while anything at or above that threshold always routes to an approver.

Accuracy that compounds

Reading invoices accurately is half the job, and the models did it well. Across the full window, field-prediction accuracy reached 97.7%, and every field was corrected by a person less than 2% of the time. The fields that matter most for payment, the total paid and the vendor, were among the most accurate.

More importantly, accuracy improved as the models learned the account. Since the early ramp, overall accuracy rose from 95.4% to 98.5%, vendor corrections fell from 3.0% to 0.2%, and cost-category corrections dropped from 3.3% to 0.8%. What the finance team noticed first wasn't speed; it was confidence. When you can trust what the system extracted, you stop re-checking everything by reflex, and reporting gets cleaner because the underlying data is clean.

PO matching: the skill that makes hands-off safe

The other half of the job is matching each invoice to the right purchase order, and knowing when not to. Nearly every invoice is three-way matched before it can post, and 87.9% land on a clean match. The rest are handled by design rather than by luck. If the model picks the wrong PO, or none, the invoice simply can't match, so it is automatically held for a person and never auto-posts. The highest-impact error is structurally prevented: a bad match surfaces as review work, never as a bad payment. When a vendor bills in different units than the PO, the system can still match on total amount within the PO balance, releasing invoices that line-level matching alone would stall.

Approvals that scale with the organization

Hands-off posting only works if the human-approval layer behind it is solid. The company runs approvals on an HRIS integration: employees, roles, and reporting lines sync automatically from Workday, and role-based approval flows route every invoice that needs a person to the right approver. More than 9,000 employees and roles and over 1,600 approval flows stay current across all eight entities, with no manual approver maintenance. As people are hired, change roles, or leave, routing updates on its own, so the invoices that need a human always reach a valid, current approver and the auto-approve path stays clean.

The tangible ROI of an AI-driven approach

Autopilot changed both how much the system does and how the AP team spends its time.

• Scale absorbed without added headcount. The operation processed roughly 289,000 invoices in six months, peaking near 80,000 in a single month, without scaling the AP team to match. Autopilot grew from 2% to 47% of fresh invoices, the system now executes nearly all of the posting itself, and about 63% of recent invoices flow fully hands-off, end to end.
• A team focused on judgment, not data entry. With the routine work automated, the AP team shifted from keying invoices and matching line by line to handling the exceptions the system deliberately holds back. Around 530 people across the business act as approvers on exactly the decisions that warrant one.
• Automation without the risk. High automation only counts if the money stays right. On the hands-off cohort, a person changes any field on just 3.54% of invoices, and net cash exposure came to roughly 0.05% of the dollars processed, about $9,000 against $16.9M, with more than 80% of what little remained concentrated in just eight vendors that a few approval rules contain.

Positioned to keep growing

The result is a finance function that scales with the business instead of capping it. Because roles sync automatically and clean invoices post themselves, folding in a newly acquired entity becomes closer to a configuration task than a staffing project, and the next branch or region no longer triggers a hiring scramble in AP. The AI reads every field and improves as it learns; PO matching turns that accuracy into safe, hands-off posting by structurally holding anything that doesn't match; and the HRIS-synced approval layer guarantees exceptions reach the right person. Together, those engines took a long-tail, PO-heavy enterprise AP operation to 87.9% clean matching, roughly 63% hands-off flow, and a net cash exposure of 0.05%, at the scale of nearly 289,000 invoices in six months, and left the company ready for whatever its growth plan brings next.
‍

[Download the full case study]