Blog/Automation in practice

Accounts payable automation best practices: the five that decide whether it pays

Accounts payable automation best practices that decide the return: clean the vendor master first, measure cost per invoice and the exception rate before and after, design for the exceptions, keep approval controls, and calculate ROI on the whole process.

Accounts payable automation best practices come down to a short list of decisions made before the software goes in, and those decisions set the return more than the software does. The benchmark that frames them is Ardent Partners' State of ePayables 2025, which put the average all-inclusive cost of processing a supplier invoice at $9.84 and the average exception rate at 18.4 percent. The full report shows the spread behind those averages: best-in-class teams spent $2.65 per invoice against $12.42 for everyone else, and carried an exception rate of 11.1 percent against 20.9 percent. Ardent's own reading is that exceptions are typically the biggest single reason the costs aren't lower. The five practices below are the ones that move a team from the second group toward the first.

Accounts payable automation best practices start with the vendor master

The vendor master is the list of suppliers the company pays, with each supplier's legal name, tax ID, remit-to address, bank details and payment terms. Every automated step reads from it. The match looks up the supplier to find its orders, the payment run takes its bank details from it, and the fraud check compares an invoice's bank details against it. A vendor master with duplicate records, inactive suppliers still set to active, and bank details nobody has confirmed in years will make an automated process fail in the same places a manual one does, only faster.

The cleanup is unglamorous work. Merge the duplicates, deactivate suppliers with no activity in a set period, confirm bank details through a call-back to a known contact, and record who may change a supplier's details and how the change is verified. Ardent's report recommends that AP teams invest in establishing a baseline of data quality before scaling AI, because supplier data can be incomplete, inconsistent or stored in different systems. It also found that 67 percent of best-in-class AP teams hold some responsibility for the vendor master, which puts the people who see the payment consequences of bad supplier data in a position to fix it.

The vendor master is also a common route for payment fraud. The Association for Financial Professionals found in its 2026 Payments Fraud and Control Survey that 76 percent of organizations experienced attempted or actual payments fraud in 2025. A fraudster who gets a supplier's bank details changed doesn't need to beat the match, because the invoice is genuine and only the destination of the money is wrong. Automation that pays faster against an unverified vendor master pays the fraudster faster too.

Measure the baseline, then design for the exceptions

Measure before you buy. The numbers that matter are the all-inclusive cost per invoice (staff time, overhead and systems, divided by invoice volume), the cycle time from receipt to approval for payment, the exception rate, the straight-through rate, and a count of exceptions by cause. Ardent's 2025 averages give a reference point for each: $9.84, 8.2 days, 18.4 percent of invoices as exceptions, and 35.4 percent processed straight through with no person touching them, against 51.0 percent for best-in-class teams and 29.0 percent for everyone else. A team that skips the baseline can't show a return later, and it can't tell which part of the process the automation improved.

The count by cause matters most, because it shows where the project should aim. In a typical AP queue the exceptions come from a small number of causes: a price on the invoice that differs from the order, a quantity that differs from the receipt, a receipt that was never entered, an invoice with no purchase order, a supplier the system can't identify. Each cause has a different fix, and some of the fixes sit upstream in purchasing or receiving rather than in AP. We walked through how the match produces these exceptions, and how a redesign moves the work upstream, in three-way match, rebuilt.

Design for the exceptions, because the clean invoices were never the expensive part. A clean invoice that agrees with its order and receipt costs very little to process by any method. The cost sits in the fifth of invoices that fail, each of which needs someone to open three screens, reconstruct what happened and decide what to do. Ardent's respondents ranked a high percentage of exceptions as their second-largest challenge (48 percent), just behind approvals that take too long (49 percent). An automation that clears the clean invoices and drops the rest into the same queue as before saves less than the vendor's demo suggests. The arithmetic of that queue is the subject of the unit economics of the exception tail: the cost of the whole process follows from how many invoices need a person and how long each one takes. So the useful design question for every exception cause is what evidence a person needs to resolve it, and whether software can gather that evidence before the person opens the case.

Keep approval controls and segregation of duties

Segregation of duties means no single person can complete a payment on their own: the person who sets up a supplier doesn't approve its invoices, and the person who approves invoices doesn't release the payment run. Automation projects tend to weaken this without anyone deciding to. A workflow tool gets one administrator who can edit suppliers, change approval routes and release payments. An approval threshold gets raised to reduce the queue. A bot runs under a shared login with broad rights, so the audit trail shows the bot and not the person behind the change.

The practice is to carry the existing controls into the new process and write them down as part of the design. Approval limits stay tied to the company's delegation of authority, and changes to them go through the controller. The automation gets its own identity in each system, with only the permissions its steps need, so every entry it makes is attributable. Changes to supplier bank details stay a separate step with a separate approver, outside the invoice flow. The approval step itself can still get faster: routing an invoice to the right approver with the order, the receipt and the variance attached removes most of the delay Ardent's respondents complained about, without removing the approval.

Robotic process automation in accounts payable, and where bots break

Robotic process automation in accounts payable means software bots that operate the same screens a clerk uses: they open an invoice, copy fields into the ERP, download a statement from a supplier portal, or move a file between folders. Deloitte's 2017 global RPA survey, The robots are ready. Are you?, defines RPA as the automation of rules-based processes with software that uses the user interface. That definition explains both its appeal in AP and its limit. Accounts payable automation using RPA needs no change to the ERP and no API, which suits a mid-market company running an older system. And because the bot works at the keystroke level, it depends on each case looking like the last one.

Deloitte's respondents named the problem directly. Process standardization was the top challenge at every stage of adoption, ahead of IT support and integration, and the report notes that processes which look standard in documentation often differ significantly in practice across business units. In AP, the clean invoice is standard. The exception is the case that differs, by definition, and that is where a bot stops and hands the invoice to a person. A bot also stops when a screen changes after an ERP update or a supplier redesigns its portal, which adds a maintenance cost the original business case rarely included. Bots in AP do well on high-volume, stable steps on the clean path, and they leave the exception queue roughly where it was. We discuss why model-based agents fail in different ways in why agents are a different mechanism.

How to calculate accounts payable automation ROI honestly

An honest accounts payable automation ROI starts from the baseline and splits it in two. Take a company processing 4,000 invoices a month at the Ardent average of $9.84, which is about $472,000 a year. Suppose 18 percent are exceptions and each exception costs four times as much to process as a clean invoice; the exceptions then carry close to half the total cost. A project that automates the clean invoices fully and leaves the exceptions alone removes about half the cost at best. A project that also cuts the exception rate and the handling time per exception reaches the other half. Use your own measured numbers in place of these, since the ratio between clean and exception cost varies widely between companies.

The cost side needs the same honesty. Count licenses, implementation, the integration with the ERP, the vendor master cleanup, the time AP staff spend testing, and yearly maintenance, including the cost of fixing bots or rules after system changes. Then compare the payback you are promised with what earlier adopters got. In Deloitte's survey, organizations piloting RPA expected payback in 9.3 months, while those that had implemented and scaled it reported 11.5 months, and 63 percent said their expectations of time to implement were not met. Deloitte reads the gap as an underestimate of time and cost. Budget for the longer number, and measure the same five baseline figures again six months after go-live.

A reasonable counter, answered

A reasonable counter is that these practices belong to the software vendor. Buy a good AP tool, the argument goes, and the vendor master checks, the exception routing and the approval controls come with it, so the finance team's job is to pick the right product. There is truth in this. The features to look for in accounts payable automation tools follow from the practices above: duplicate and bank-detail checks against the vendor master, matching tolerances the finance team sets, exception handling that attaches the order, the receipt and the supplier correspondence to each case, role-based permissions with an audit log, and reliable posting back to the ERP. B2B accounts payable automation with those features is easier to run well than a tool without them. But no product cleans the vendor master, sets the baseline, or decides who may approve what. Ardent found best-in-class teams process invoices at a cost 79 percent lower than their peers, and part of that gap is technology the others haven't adopted yet. The rest comes from how the teams run the process, and that part is the finance team's work whichever product it buys.

Common questions

What are the best practices for accounts payable automation?
Clean the vendor master before anything else, because every match, payment and fraud check reads from it. Measure the all-inclusive cost per invoice, the cycle time, the exception rate and the straight-through rate before the project starts, so the result can be measured against something. Design the automation around the invoices that fail the match, since they carry most of the cost. Keep approval limits and segregation of duties in place, and give the automation its own identity with its own limits.
How do you calculate accounts payable automation ROI?
Start with the all-inclusive cost per invoice today, multiplied by annual invoice volume, and split it between clean invoices and exceptions. Estimate the new cost of each group after automation, including the people who still work the exception queue, and subtract it from the baseline. Then set that saving against the full cost of the project: licenses, implementation, integration with the ERP, the vendor master clean-up and yearly maintenance. In Deloitte's 2017 survey, organizations piloting RPA expected payback in 9.3 months, while those that had implemented and scaled it reported 11.5, so budget for the longer number.
Is robotic process automation a good fit for accounts payable?
Robotic process automation works for the steps in accounts payable that are stable and rule-based, such as keying a clean invoice into an ERP screen or downloading a statement from a supplier portal. A bot follows the screen at the keystroke level, so it depends on the process being standardized and the screens staying the same. It does poorly on the exceptions, where an invoice disagrees with its order and someone has to work out why. RPA belongs on the clean path, with the exception queue left to people or to software that can read and reason about documents.

Book a demo

Start with one department

We map how one team’s work runs, show you where the hours go, and automate from there.