Automated invoice processing applies software to the accounts payable invoice workflow, so the system handles intake, extraction, validation, matching, coding, routing, and posting without a person keying or moving anything. It rarely breaks down at capture. It stalls at exception handling and at coding invoices that arrive with no purchase order behind them.
Automation projects rarely break down at capture. They stall further along, at the invoice with no purchase order (PO), no clean line data, and a question typed into the body of the email. Software can run the whole accounts payable (AP) invoice workflow, from intake through to posting, without a person keying or moving anything. How many stages it can finish depends less on the software than on your invoice mix.
Ardent Partners published its State of ePayables 2025 benchmarks in January 2026, covering 2025 performance. The average organization spent $9.84 to process a single invoice and took 8.2 days from receipt to approval. Its average invoice exception rate was 18.4 percent, and AP staff spent 21.9 percent of their time on supplier inquiries.
Those four numbers describe one condition. Cost and cycle time are outputs of the exception rate. The exception rate is an output of how invoices arrive and how clean the data behind them is. An AP team with 18.4 percent exceptions runs two operations at once. One is a production line that works. The other is a research desk, staffed by the same people, resolving invoices the line could not finish.
Adoption is no longer the differentiator. Gartner reported in November 2025 that 59 percent of finance leaders used AI in their finance function during 2025, in its finance AI adoption survey. Among adopters, accounts payable process automation was the second most common use case at 37 percent. The same research found that 91 percent of respondents initially experienced low or moderate impact. Buying the technology and getting the outcome are separate events, and the gap between them is where most of the interesting detail lives.
Start with straight-through processing (STP). It is the number every vendor quotes and almost nobody defines the same way.
An invoice is straight-through when the system takes it from intake to posting with no human intervention at any stage. That is the strict reading. In practice, definitions vary in ways that change the number by tens of points. Some counts include invoices where a person cleared an exception in under a minute. Some include only PO-backed volume, which excludes the services and utility invoices that never match. Some start the clock at capture rather than intake, which quietly removes every invoice that arrived by paper or landed in the wrong mailbox. Some count an invoice as touchless if no key was pressed, even when someone read it, decided, and clicked approve.
There is a useful tell here. Ardent Partners does this measurement for a living. It publishes straight-through performance for its top performing group only as a relative, at more than 1.8 times as many invoices processed straight through as the average. It never publishes an absolute percentage. It treats invoice cost and cycle time the same way, at 79 percent lower and 79 percent faster than the average. The most careful benchmark house declines to state an absolute, so a vendor stating one is describing its own definition.
Ask for the denominator. What share of invoices did the vendor count? Which channels were included? Were non-PO-backed invoices in scope? What counts as a touch? Two vendors quoting the same headline can be describing different worlds.
The stages themselves do not change under automation, and a companion piece defines each invoice processing stage. What changes is who does the work and what it costs to be wrong.
Automation consolidates channels into a single point of entry. Email, supplier portals, electronic data interchange (EDI), and structured e-invoices land in one queue with one timestamp. The gain is measurement as much as speed. You get an arrival time for every invoice, which is the first time most teams can see how long invoices sat before anyone touched them.
Extraction changes from keying to reading. Modern capture reads header fields at high accuracy on most documents. It struggles more with line detail, non-standard layouts, and handwriting. Accuracy at the document level and accuracy at the field level are different measurements. A system reading 99 percent of fields correctly still puts a real share of invoices in front of a person when each one has thirty fields.
This is the stage where automation quietly catches duplicates. Every invoice gets checked against the vendor master, the open payables ledger, and the invoices already in flight, on every submission, without fatigue. A human checker compares against memory and a search box. Duplicate detection is one of the few stages where the machine checks more invoices than a person can, because it runs on every invoice rather than on the ones that look suspicious.
Matching automates well when a PO, a receipt, and clean line data exist. Multi-line matching against partial receipts and staged deliveries is harder, and how you configure tolerance determines most of the outcome. A tolerance wide enough to clear the queue lets through variances you would want to see. A tolerance tight enough to catch them creates exceptions your team has to work through.
Automation earns or loses the business case here. Non-PO-backed invoices cover rent, utilities, professional fees, subscriptions, and most services. They arrive with no requisition to inherit coding from. Rules-based coding handles the recurring ones and fails on anything new, because a rule needs a pattern that already exists. AI-based coding proposes a general ledger (GL) account from the vendor, the description, and prior treatment. It works well on volume, with a confidence threshold and a review path for the rest.
Routing automates cleanly and saves less time than expected. The right approver gets the invoice faster, then the decision still waits on a person. Reminders, escalation, and substitution rules shorten that wait more than routing logic does.
Automation reduces how many exceptions occur and changes almost nothing about how they get resolved. The invoice still needs a person with authority over the underlying disagreement, in procurement, receiving, tax, or the budget owner's seat. Most projects route every exception back to AP, which turns the AP team into a switchboard for decisions it cannot make. Designing that routing is the work, and our guide to AP invoice processing exceptions sets out the options.
Posting automates fully once coding, matching, and approval are settled. It is the least interesting stage to automate and the easiest to demonstrate, which is why demonstrations tend to end there.
A program that automates only the first half of the workflow relocates effort. Keying disappears, and chasing answers expands. Your team stops typing invoice numbers and starts fielding questions about why an invoice is blocked, which is the same volume of contact arriving under a different name. Ardent's finding that AP staff spend 21.9 percent of their time on supplier inquiries describes work that no extraction engine removes.
Two more limits are worth stating plainly. Automation inherits your data quality. A vendor master with duplicate records produces more unmatched vendors after automation, because the machine stops making the charitable guesses a person made. And a high straight-through rate achieved by widening tolerances is a control decision, because every variance you stop reviewing is a variance you stop detecting.
Scope tends to follow the demonstration. Capture, matching, and posting look impressive in a room, so they get bought, configured, and measured. The invoice population that generates the cost is outside that scope. It arrives without a PO, needing a coding decision and a conversation.
The second pattern is measuring the wrong thing. Touchless rate rewards volume that was already easy. Exception rate by cause, resolution time by owner, and aging by queue tell you where the money goes, and few programs report any of them at go-live.
Our platform automates the whole run, from intake through posting, and puts the same attention on the work that surrounds the invoice. Our AI reads invoices from every channel, codes non-PO-backed invoices from vendor and historical treatment, matches multi-line POs against receipts, and posts through more than 100 ERP connectors.
The difference shows in exception design. Our AI Agents resolve the recurring reasons invoices stall, answering payment status questions, verifying bank account change requests, and stopping duplicates before they reach a person. Customers reach automation rates of 80 percent or more, and Qualcomm moved from 14 percent to 61 percent autonomous invoice processing on SAP S/4HANA.
Every action an Agent takes leaves an audit trail with the evidence behind the decision, which is what makes higher autonomy defensible to an auditor. Our view of AI in accounts payable sets out the maturity levels, and our AP automation capabilities cover capture, coding, routing, and matching in detail.
Before you compare vendors, measure your own invoice mix. Count what share of your volume is non-PO-backed, and count your exceptions by cause rather than in total. Those two numbers set the ceiling on what any automation can reach, and they tell you which stage to design first.
Touchless invoice processing describes an invoice the system takes from intake to posting with no human intervention at any stage. Vendor definitions vary widely on which invoices enter the count and what qualifies as a touch, so a touchless percentage is only meaningful alongside its denominator.
Ardent Partners reported an average 2025 processing cost of $9.84 in January 2026, with its top performing group running 79 percent below that average. Ardent publishes those figures as relatives, so treat any absolute cost promised by a vendor as an estimate rather than a benchmark.
Exception handling is the hardest stage, together with coding on non-PO-backed invoices. Both need a decision from someone outside AP, so they resist automation that only reads documents. Ardent Partners put the average exception rate at 18.4 percent for 2025.
It improves detection coverage, because duplicate and vendor checks run on every invoice rather than on a sample. It does not remove the risk. A fraudulent invoice that looks legitimate and falls inside your tolerance still gets paid, which is why tolerance settings deserve review alongside your automation design.