Use case · Accounts payable
Invoices that match themselves. Exceptions that find the right person.
An AI agent reads every supplier invoice, checks it against the purchase order and goods receipt, and posts the clean ones to your ERP. The ones that don't match arrive on the right desk with the evidence already attached.
For finance directors and AP managers at mid-market and enterprise companies processing thousands of supplier invoices a month.
The situation
It is the third working day of the month.
The AP inbox holds four hundred unread messages. Some are invoices, some are statements, some are a supplier asking why they haven't been paid. A PDF from a logistics provider has twelve pages, and the invoice number is on page nine.
Each invoice follows the same path: someone opens it, keys the header into the ERP, finds the purchase order, checks the quantities against what the warehouse received, and chases a budget holder for approval when the numbers are off by more than tolerance. Most invoices are routine. The team still touches every one.
Meanwhile the controller wants accruals by Friday, a supplier has changed bank details by email, and nobody is sure whether invoice 88412 was already paid last month under a slightly different number.
How it works
Follow one invoice through.
Three stages. The work that used to happen by hand, done before anyone opens the queue.
Conceptual flow, not a live system or measured result.
Read everything that arrives
The agent watches the AP mailbox, supplier portal and scanned mail. It separates invoices from statements, credit notes and correspondence, then extracts supplier, dates, totals, tax and line items — including from multi-page and scanned documents. Every extracted field keeps a pointer back to where it came from on the page, so a reviewer can check a value in one click.
Match before anyone looks
Each invoice is checked against the purchase order and goods receipt: two-way or three-way, at the line level, inside the tolerances your finance policy already defines. The agent also looks for duplicates by content, not just invoice number, and flags supplier master-data changes such as a new bank account. Invoices that pass are prepared for posting; everything else becomes an exception with a reason.
Route the exception, not the invoice
Price variances go to the buyer. Missing receipts go to the warehouse. Unknown cost centres go to the budget holder. Each person sees only the question they need to answer, with the invoice, PO and receipt side by side. Their decision is written back to the ERP with an audit trail, and repeated exception patterns are reported so the root cause — a stale price list, a supplier who never quotes the PO — can be fixed upstream.
Where people stay in control
Automation for the routine. Judgement for the rest.
Payment runs stay human
The agent prepares invoices for payment; it does not release funds. Payment proposals are approved in your ERP or banking platform by the same people who approve them today.
Bank-detail changes are always escalated
Any change to a supplier's remittance details is held and routed for independent verification, whatever channel it arrived through. This is the most common invoice-fraud pattern, so it never takes the automated path.
Tolerances come from your policy
Match tolerances, approval limits and segregation of duties are configured from your existing finance policy and delegation of authority — not inferred by a model.
Low-confidence reads go to a person
When the agent is unsure of a value, it says so and asks. Confidence thresholds are set conservatively at launch and only relaxed when the review data supports it.
Worked example
Model it with your own numbers
Adjust the inputs to your AP operation. The defaults are illustrative starting points, not results from a ColdAI engagement — replace them with figures from your own team.
Hours today = invoices × minutes ÷ 60. Hours after = invoices × (1 − touchless share) × exception minutes ÷ 60. Monthly capacity released = the difference × cost per hour. This excludes platform and implementation cost, discount capture and fraud avoided.
Illustrative model, not a quote or a measured result.
Judging the pilot
Five numbers that tell you whether it is working.
Agree these before the pilot starts and measure them on the same invoice population before and after, so the result is a comparison rather than an impression.
- 01First-pass match rate
- Share of invoices that match PO and receipt within tolerance without anyone touching them.
- The headline measure of automation — but only meaningful alongside the error rate below.
- 02Posting error rate
- Invoices posted automatically that are later corrected, reversed or credited.
- A high match rate with rising corrections means tolerances are too loose, not that the agent is good.
- 03Exception reason mix
- Exceptions grouped by cause: price variance, missing receipt, unknown PO, master-data mismatch.
- Shows which upstream fixes would remove the most work — often a supplier or a buyer, not the AP team.
- 04Duplicates caught before payment
- Potential duplicates flagged and confirmed before the payment run, compared with those found afterwards.
- Moves duplicate detection from audit findings to prevention, which is where the money is recovered.
- 05Receipt-to-post cycle time
- Elapsed time from an invoice arriving to it being posted or routed as an exception.
- Determines accrual accuracy at month end and whether early-payment terms can actually be taken.
An honest boundary
When this is the wrong answer.
We would rather tell you now than six weeks into a pilot.
- You process a few hundred invoices a month. A well-configured ERP capture module is likely to be cheaper than an agent.
- Most spend has no purchase order. Without a PO to match against, the agent can read and code invoices but cannot validate them — fix PO compliance first.
- Supplier master data is unreliable. Duplicate or outdated supplier records will generate false exceptions; a data clean-up should come before automation.
- You want a system that approves payments on its own. We will not build that: payment release stays with authorised people.
Works alongside
ERP and inbox integrations
- SAP S/4HANA
- Oracle Fusion Cloud ERP
- Microsoft Dynamics 365 Finance
- NetSuite
- Sage Intacct
- Coupa and SAP Ariba supplier portals
- Microsoft 365 and Google Workspace mailboxes
Questions finance teams ask
Before you start.
How is this different from the OCR our ERP already has?
OCR turns a document into text. The agent does the work that follows: classifying the document, finding the matching PO and receipt, applying your tolerances, spotting duplicates by content, deciding who owns an exception, and writing the outcome back with evidence. Where your ERP's capture is good, we use it and build on top.
Does the agent approve or pay invoices?
No. It prepares clean, matched invoices for posting according to rules you set, and routes everything else. Approval limits follow your delegation of authority, and payment runs are released by authorised people in your ERP or bank.
What happens with handwritten or poor-quality scans?
They are read, but low-confidence fields are flagged rather than guessed. A reviewer sees the image region next to the extracted value and confirms or corrects it; corrections are logged and used to evaluate the extraction step.
How long does a pilot take?
A bounded pilot usually covers one entity, one mailbox and a defined supplier segment. The length depends on ERP access and how clean the PO and receipt data is; we agree scope and acceptance criteria before starting rather than quoting a fixed number of weeks.
Where is invoice data processed and stored?
Deployment is designed around your data-residency and security requirements, including hosting in your own cloud tenancy where needed. Invoice images and extracted data follow your existing retention policy.
How do you prevent duplicate payments?
Duplicates are checked by content — supplier, amount, date, line items and document similarity — not just invoice number, so re-sent invoices with a changed reference are still caught before posting.
Go deeper
The pieces behind this workflow.
What would you like to move forward?
Start with the outcome that feels closest. We'll help give the next step a useful shape.
One conversation. A clearer direction.
Tell us what you want to improve. A few sentences are enough to start.
shayan@coldai.orgLast reviewed 25 September 2026.