Which part of accounts payable to automate first? Start where the invoice breaks.

Pull one invoice from a mid-market firm's accounts payable inbox and follow it. A Chicago distributor's freight bill lands by email on a Tuesday. It sits two days before anyone opens it. The purchase order number is printed on a different line than the system expects, so the match fails. It routes to a manager who is traveling. It clears the next Monday, six days after it arrived, and nobody in the building could tell you what those six days cost.

That drift is the norm, not the exception. Ardent Partners puts the average cost of processing a single invoice at $12.88 for teams without automation, and the average time to process one at 17.4 days, in a read of its State of ePayables research. The fastest teams cut more than $10 and roughly 14 days off those numbers. The gap is not a mystery. It is the difference between an invoice that clears itself and one a person has to chase.

Get the break points in your own AP run mapped before you buy any software

What accounts payable automation actually buys you

Strip the vendor language and the work is five steps. Capture reads the invoice, pulling vendor, amount, dates, and line items off a PDF or an email instead of a clerk keying them. Coding assigns the general ledger account and the cost center. Matching lines the invoice up against the purchase order and the receipt, the three-way match that proves you are paying for something you ordered and got. Approval routing sends it to whoever owns that budget. Payment schedules the run and captures any early-pay discount.

The money is not spread evenly across those steps. APQC benchmarking, drawn from 1,485 organizations, shows the best performers processing an invoice for $2.07 or less while the bottom quartile spends $10 or more, in its cost-per-invoice benchmark. A five-fold spread on the same task points straight past the typing. The cost lives in the exceptions: the invoices that do not match, the approvers who sit on them, the vendor calls to sort out a wrong number.

What to automate first in accounts payable

Here is the order we would run it, and it is not the order most vendors sell.

1. Capture and coding. Kill the keying first. Reading vendor, amount, and line items off the document is the highest-volume, lowest-judgment task in the room, and it feeds every step after it. Get this clean before you touch anything downstream, because a bad capture just automates a wrong number faster.
2. The three-way match. Auto-clear every invoice that ties cleanly to a purchase order and a receipt. For most firms that is the majority of the pile, and each one that self-clears is a person's afternoon back. This is where the $12.88 becomes $2.
3. Exception routing. The invoices that fail the match are the whole game. Automate the routing, not the judgment: get the broken invoice to the one person who can fix it, with the reason it broke attached, instead of into a shared inbox where it ages.
4. Approval workflow. Rules by amount and cost center, with reminders and an escalation when an approver goes quiet. The traveling manager is a process gap, not a personality flaw, and a second approver on a timer closes it.
5. Payment controls. Schedule to capture early-pay discounts, and put a hard gate on the two events that move money somewhere new. Do this last, because a fast payment engine on a shaky match is how you pay a fraudster on time.

Is accounts payable automation worth it for a mid-market team?

For a team processing a few thousand invoices a month by hand, yes, and the math is not close. Take the Ardent figures: at $12.88 an invoice, 3,000 invoices a month is roughly $39,000 in monthly processing cost, on top of a 17.4-day cycle that quietly forfeits early-pay discounts worth one to two percent of spend. Move even half of those invoices to a touchless path at a top-quartile cost and you free both the cash and the people. The payback rarely comes from the headcount you cut. It comes from cycle time you recover, discounts you stop leaving on the table, and the late-payment penalties and duplicate payments that vanish once a machine, not a person's memory, holds the record. The one caveat: automation only pays if the exceptions have somewhere to go. A tool that speeds up the clean invoices and still dumps the broken ones on one overloaded clerk moves the bottleneck, it does not remove it.

The touchless number that should scare you

Every AP tool sells a touchless rate, the share of invoices that post and pay with no human touch. Chase it too hard and you build a fast lane for fraud. In 2024, 79 percent of organizations were hit by attempted or actual payments fraud, checks were the most-targeted method at 63 percent, and business email compromise tied checks as the top attack avenue, also at 63 percent, according to the Association for Financial Professionals. A forged vendor bank-change email is exactly the message a touchless pipeline waves through. This is the same failure mode we traced in the fraud losses that arrive with valid credentials: the system is not fooled, the process is. The fix is not slower automation. Put a hard human gate on the two events that move money to a new place, a new vendor and a changed bank account, and let everything else fly.

How we would run it in one week

When we map an accounts payable week, we do not start with a tool. We start at the invoices that broke last month and ask why. Pull 60 days of paid invoices, sort them into cleared-itself and needed-a-human, and read the second pile until the two or three exception types eating the month have names: a vendor whose PO format never matches, a cost center nobody codes the same way twice, a plant that emails photos of invoices. That pile is your automation spec, and it is the same place a bank's back office hides its real cost. It is also the automate-first logic we would run in a claims shop: clear the clean ones by rule, spend the people on the exceptions. What an operator walks away with is a one-page map of which invoices the team should never touch again and which exception types are worth a build, plus a first target: get the clean-match invoices onto a touchless path and put every hour of saved attention on the exceptions and the fraud gates.

Monday, the first move is not a demo. It is to pull last month's paid invoices, split the ones that cleared themselves from the ones that needed a person, and read the second pile until the three exception types eating the week are named. Automate those, gate the vendor and bank-account changes, and let the clean invoices run. The stack of paper on the corner of someone's desk was never the job. Finding out which of it a machine should have cleared a week ago is.