Accounts Payable Invoice Automation

Most of invoice automation is not AI

People hear “invoice automation” and picture a model doing everything. In practice the model does two jobs: it reads documents, and it suggests coding for spend it has not seen before. Everything else is plain rules, which are cheaper to run and much easier to audit.

Step	Done by	Why
Pull attachments from the AP mailbox	Rule	It is just email
Tell an invoice from a statement or a reminder	Model	Layouts vary too much for rules
Read supplier, dates, lines, tax, total	Model	Same reason
Match the supplier to your vendor list	Rule	Tax ID and bank details, then name
Check the arithmetic and look for duplicates	Rule	Exact checks, no judgment needed
Match to the PO and goods receipt	Rule	Tolerances you set
Code a supplier you have paid before	Rule	Ledger history
Code a new supplier or new kind of spend	Model, then a person	No history to go on
Route for approval	Rule	Your approval matrix
Create the draft bill	Rule	Accounting system API

If someone quotes you a project where a model makes every one of these decisions, ask how they plan to audit it.

The companion guide on AI for accounts payable shows exactly what the model returns from an invoice, and when an AP suite is a better buy than a custom workflow.

Best practices before you automate anything

These cost nothing, and they decide whether the automation works. Do them first, even if you end up buying software instead of building.

1. Move every supplier to one AP address. Invoices sent to personal inboxes are the main reason bills get paid late or twice. Email your suppliers once, then forward stragglers for a quarter.
2. Clean the vendor master. Merge duplicate suppliers, fill in tax IDs, and lock who can change bank details. The automation matches against this list, so its mistakes become your mistakes.
3. Write the approval matrix as a table. Amount bands, cost centers, approvers and holiday cover. If you cannot write it down, nothing can follow it.
4. Set your match tolerances. Decide how big a price or quantity difference is acceptable before an invoice counts as an exception. A small fixed amount or a percentage, whichever is larger, works for most teams.
5. Measure a baseline for two weeks. Count invoices, minutes per invoice, days from receipt to approval, and how many needed a follow-up. Without this, nobody can tell you later whether the project paid off.

An example approval matrix for a 60 person company:

Invoice	Approver
Under $1,000, matches a PO	AP lead approves the draft
$1,000 to $10,000	Budget owner of the cost center
Above $10,000, or no PO	Finance director
New supplier or changed bank details	Blocked until verified by phone

How coding from history works

Most coding is repetitive. The same supplier gets coded the same way month after month, and your ledger already knows that. The workflow looks up the last twelve months for each supplier and proposes the dominant coding. It only asks the model when there is no history.

Here is how three lines on one invoice get handled (account numbers are illustrative):

Line	What the ledger says	Suggestion
Corrugated boxes	14 of the last 15 bills coded 5010 Packaging	5010 Packaging, Warehouse. Ready to approve
Pallet wrap	Coded 5010 twice and 6200 Consumables once	5010 Packaging. Flagged because history is split
Annual software licence	Never seen from this supplier	6420 Software, suggested by the model. Needs review

The clerk confirms the flagged lines once. From then on the history does the work.

What to measure after go live

Track these monthly from the tracker table. They are the numbers that tell you whether to widen the automation or fix something first.

• Days from receipt to approval, as a median and for the slowest 10%
• Share of invoices that needed an exception, and the top three reasons
• Share of drafts approved without any edit
• Duplicates and bank detail changes caught before posting
• Minutes of human time per invoice, compared with the baseline

We do not quote industry benchmarks here. Compare month three with the baseline you measured before go live, because that is the comparison you can trust.

When Xero or QuickBooks capture is enough

Both have their own bill capture. Xero includes Hubdoc, and QuickBooks Online can pull bills and receipts in from email. For a small volume of clean PDFs from a few regular suppliers, try those first. They cost nothing extra.

They run out of road when you need an approval matrix with escalation, coding from history, PO and goods receipt matching, statement checks, several entities, or documents that arrive as scans and photos in different languages. That is where a custom pipeline earns its fee.

What breaks most often

• PDFs that hold several invoices, or one invoice split across several emails. The pipeline splits and joins them, but you should watch the first month closely.
• Payment reminders that look like new invoices. Sorting documents before reading them stops these from becoming duplicates.
• Coding drift. When your chart of accounts changes, the history points at old accounts. Re-map them once, at the same time as the chart change.
• Approvers who never open the queue. The reminders help. A weekly list of overdue approvals sent to the finance director helps more.

How we build it

We build the pipeline as a fixed-scope AI automation project, live in your mailbox and accounting system in two to three weeks, with monitoring, error alerts and a runbook. We start from your real invoices and your written approval matrix. If you are still deciding whether to build at all, start with a readiness audit.

For finance teams whose documents cannot go to a public model, the reading step runs privately. See the finance industry page, or browse every process on the processes overview.