The average invoice takes 9.2 days to process and costs $9.40, and only 32.6% reach payment without a human touching them — figures from Ardent Partners' 2025 AP metrics report, based on 212 AP organisations. Those numbers get quoted in every software pitch, and they are the wrong ones to start from, because none of them tells you where the 9.2 days went.

Cost per invoice is a lagging metric. It is the sum of everything upstream, which means it can only tell you that something is slow, never which thing. Buy software against a lagging metric and you automate whichever stage the vendor happens to be good at, which may not be the stage where your invoices are actually sitting.

So before any tooling decision: find the wait. It takes an afternoon, and about half of what it turns up can be fixed without buying anything.

The measurement to take this week

Export the last 200 paid invoices and capture five timestamps for each:

Received — when it first arrived anywhere in your organisation, not when AP saw it
Entered — when it existed as a record in your system
Coded — when GL account, cost centre, and tax treatment were final
Approved — when the last required approval landed
Paid — when the payment was released

That gives you four intervals: intake, coding, approval, and payment. Compute the median and the 90th percentile for each.

Both numbers matter, and for different reasons. The median tells you what the ordinary invoice experiences. The 90th percentile tells you what generates the phone calls — and where the supplier inquiries come from, which Ardent puts at 21.8% of AP staff time. An approval stage with a two-day median and a fourteen-day p90 is not a two-day stage. It is a stage with a tail, and the tail is the problem.

Two practical notes. Getting a true received timestamp is the hard part, because invoices arriving in personal mailboxes have no measurable arrival time — and if you cannot measure it, you have already found your first finding. If your system cannot export a coded timestamp, use the first approval routing event as a proxy and note the substitution.

The five stages, and what a broken one looks like

Stage	What you measure	Healthy shape	Broken looks like
Intake	Received to entered	Under a day, tight distribution	Invoices "discovered" during close; no single arrival point
Coding	Entered to coded	Hours, and mostly automatic	AP guesses, the budget holder corrects it after close
Approval	Coded to approved	p90 within 3× the median	A long tail, with the same names in it every month
Exception	Time from flag to resolution	Owned outside AP, with ageing	One queue, one owner, no clock
Payment	Approved to paid	Governed by due dates and discount windows	Governed by the calendar and nothing else

Exception handling is deliberately not sequential with the others. It runs alongside, it applies to roughly one invoice in seven — Ardent puts the exception rate at 14% — and it is where the p90 in every other stage comes from. Measure it as its own interval or it will hide inside the four above.

Fixes that cost nothing

Work these before any purchase, because each one removes a delay that software would otherwise automate rather than eliminate.

Give invoices one front door. A single address or channel that every supplier is told to use, with personal mailboxes forwarding into it. This is usually the largest single reduction in intake time available to a mid-market team, and it costs a change to your remittance advice and one email to your top 50 suppliers. It also makes the intake interval measurable for the first time.

Write down the default coding per vendor. Most vendors code the same way every single time. Nobody has written that down, so AP guesses and someone corrects it later, which shows up as journal corrections after close rather than as coding time. Take your top 50 vendors by invoice count, agree the default with the budget holder once, and record it against the vendor. Coding stops being a decision and becomes an application of a rule. GL coding automation is worth buying afterwards, when the rules exist to automate.

Audit your approvers by rejection rate. For each approver in the last 200 invoices, count how many they rejected or sent back. An approver with a zero rejection rate across a hundred invoices is not a control — they are a delay wearing a control's badge. Either raise the threshold above which their approval is required, or remove them from that path and route it to someone who does exercise judgement. This is usually the single biggest cut to approval time and it costs one uncomfortable conversation.

Name a backup for every approver, and enforce delegation. Most approval tails are one person's annual leave. A named delegate with an automatic handover after 48 hours removes the tail without changing anything else.

Route exceptions to whoever can close them. A price dispute belongs to the buyer who signed the contract, a receipt discrepancy to the receiving location, a coding question to the budget holder. AP owns the queue and its ageing; AP does not own the answers. Routing every exception to AP converts a distributed problem into one team's backlog — and it is the mechanism behind most of that 21.8% of staff time spent on supplier inquiries.

Move the cutoff before you move the cadence. Teams argue about weekly versus twice-weekly payment runs when the actual problem is that the cutoff is an hour before the run, so anything approved on the day misses it and waits a full cycle. A 24-hour cutoff, published and enforced, fixes more than a cadence change. The full sequence is in payment run scheduling.

Publish a status a supplier can read. Even an auto-reply that states the expected payment date given the current queue absorbs a large share of "where is my invoice" traffic. It is not a portal and it does not need to be.

Fixes that do need tooling

These are worth buying — after the corresponding free fix is in place, because otherwise you are paying to automate a problem you could have deleted.

Buy this	Fixes which stage	Only after you have
Capture and extraction	Intake	Given invoices one front door
Rule-based coding	Coding	Written down the per-vendor defaults
Threshold routing with escalation	Approval	Audited approvers by rejection rate
Duplicate hold on vendor, amount, and date	Exception	Confirmed duplicates are actually occurring
Exception ageing with a named owner per type	Exception	Decided who owns each exception type
Supplier self-service status	Intake and inquiries	Measured how much time inquiries consume

The precondition column is the part that gets skipped. A routing engine configured against approval thresholds nobody has questioned will route invoices to the same rubber-stamp approver, faster. The stage that most often genuinely needs software is capture, because ingesting arbitrary invoice formats is a real technical problem rather than a process one — which is also why it is where ERP-embedded AP modules tend to stop, as covered in AP automation versus your ERP.

Ranking what to fix first

Rank each stage by median days in stage × share of invoices that pass through it, and start with the top one. Exception handling usually ranks higher than teams expect: it touches only 14% of invoices, but the time in stage is often ten times any other, so the product is large.

Then three rules that keep the improvement measurable:

One change per stage at a time. Two simultaneous changes to approval mean you learn nothing about either.
Re-measure after 30 days with the same five timestamps. Not a new set of metrics; the same ones, so the comparison holds.
Watch the p90, not just the median. Most process changes move the median first and the tail last, and the tail is what your suppliers experience.

Expect the first pass to be intake or approval, and expect the second pass to be exceptions once the easy time has come out. Cost per invoice will move too, but it moves last and it moves slowly, which is exactly why it is a poor thing to steer by.

What to add to your monthly reporting

Four numbers, alongside whatever you already report:

Median and p90 for each of the four intervals — the shape of the process
Rejection rate per approver — whether your approval chain is a control or a queue
Exception rate by type — which exceptions to attack next, rather than an aggregate that hides the mix
Supplier inquiries per 100 invoices — the outside view, and the one that correlates best with whether suppliers think you are easy to work with

Report those for two quarters and the tooling conversation becomes straightforward, because you will be able to say which stage you are buying for and what you expect the number to become. That is a materially stronger position than a business case built on an industry average — and it is the same discipline described in finance automation ROI metrics.

FAQ

How do you find the bottleneck in an accounts payable process?

Export the last 200 paid invoices with five timestamps — received, entered, coded, approved, paid — and compute the median and 90th percentile for each of the four resulting intervals. Rank the stages by median days in stage multiplied by the share of invoices passing through them. That product, not the raw duration, tells you where the recoverable time is. Measure exception handling separately rather than inside the other stages, because it applies to roughly one invoice in seven and usually carries far longer durations, so it disappears into an average.

Should you fix the AP process before buying automation software?

Fix everything that does not need software first, then buy. Giving invoices a single arrival point, writing down per-vendor coding defaults, removing approvers who never reject anything, naming delegates, and routing exceptions to the person who can close them cost nothing and remove delay outright. Automating those same delays instead makes them faster without making them smaller, and it hides them, so the next process review cannot find them. The exception is capture: ingesting arbitrary invoice formats is a genuine technical problem, and process changes will not solve it.

What are the standard accounts payable metrics to track?

Median and 90th-percentile duration for intake, coding, approval, and payment; exception rate broken down by type; rejection rate per approver; and supplier inquiries per hundred invoices. Cost per invoice and total cycle time are worth reporting but poor for steering, because they are sums of everything upstream and move last. Ardent Partners' 2025 benchmarks — 9.2 days to process, $9.40 per invoice, 32.6% straight-through, 14% exception rate — are useful for calibration but will not tell you which stage to change.

How long does accounts payable process improvement take to show results?

The no-cost changes show up within one measurement cycle, usually 30 days: a single intake channel and an approver audit both move the numbers quickly because they remove waiting rather than speed up work. Tooling changes take longer, typically a quarter, because configuration has to be tuned against real exceptions before the rules earn trust. Re-measure with the same five timestamps rather than introducing new metrics, and change one thing per stage at a time so the attribution is unambiguous.

Who should own exception resolution in accounts payable?

The person who can close the exception, which is rarely accounts payable. Price and rate disputes belong to the buyer who signed the contract, receipt discrepancies to the receiving location, coding questions to the budget holder, and missing purchase orders to whoever made the purchase. AP owns the queue, the ageing clock, and the escalation. When every exception routes to AP by default, the department spends its time chasing answers it cannot give, which is the most common reason cycle time stays high after extraction accuracy improves.

Let Ken run the queue

Ken gives invoices one front door in Slack, applies per-vendor coding defaults, routes approvals by threshold with automatic escalation when an approver goes quiet, and holds duplicates before they reach a payment run — with an owner and a clock on every exception. Start a free trial and measure the difference against the same five timestamps in 30 days.

Accounts Payable Process Improvement: Fix the Queue Before the Tooling