N
B
Supplier Invoice Workflow With AI

Supplier Invoice Workflow With AI

· by Panayot Dimkov

Automating the supplier invoice workflow is not "adding an OCR". It is seven stages — intake, capture, validation, matching against order and delivery note, posting, approval and filing — from the moment the invoice arrives somewhere until it is posted, approved and filed. Many of the projects that stall only automated the second one: reading the data.

This article walks through the workflow stage by stage, with what has to be decided at each one and where it usually breaks. The idea that holds it together: AI reads the invoice, rules validate it and match it against the order and delivery note, whatever matches is posted automatically and people only see the exceptions. If you are after figures, the cost and the break-even point are in automating supplier invoices: cost and when it pays off.

1Intake: normalising how they arrive

Before extracting anything there is something duller to solve: invoices arrive through four or five different channels and each one has its own chaos.

ChannelWhat has to be solved
Email A single receiving mailbox. It is the most profitable change and the one that meets most resistance, because today everyone receives their own
Supplier portals Someone logs in to download them manually. Either you connect through an API or you accept that this stage stays manual
Paper A scan with a fixed destination. If the scanner drops the PDF on somebody's desktop, no automation is possible
Structured formats The best of the lot: the data arrives labelled (Facturae, UBL, EDI) and needs no interpreting. They go through a different, cheaper path, and will gain ground once mandatory B2B e-invoicing applies in Spain
The single mailbox is the decision that most determines the project's outcome. Without it, the flow only sees part of the invoices and two processes coexist: the automated one and the usual one. And the usual one tends to win.

2Capture: extracting the data

The real difference from the tools of a few years ago is that an AI model does not need a template per supplier: it understands the document, and a new supplier needs no prior setup. For the workflow, two things have to be decided at this stage.

01
Which fields to extract
Header — supplier, tax ID, number, date, net, taxes, total — and line items. Lines are the expensive part: they multiply the difficulty and are not always needed. Deciding early avoids over-engineering.
02
Confidence level per field
A good system does not only say what it read, it says how sure it is. That figure is what lets you route only the doubtful cases to human review instead of reviewing everything.

The most common design mistake is demanding 100 % accuracy. It does not exist, with AI or with people. What you can design is a system that knows when it is not sure.

3Validation against your master data

Extracting well is not enough: you have to check that what was extracted matches the reality in your system.

  • The supplier exists and is active. If not, the invoice is not posted automatically: it is routed to supplier creation, which is another process.
  • The tax ID matches the one on file. This is the check that catches invoices from a different legal entity than expected.
  • The invoice number is not a duplicate for that supplier. A duplicate payment is one of the most expensive errors and one of the easiest to prevent automatically.
  • The bank account is the registered one. If it changes, it is not accepted without verification through another channel. This is the barrier against payment-diversion fraud, which still works because often nobody checks.
  • The arithmetic adds up. Net plus taxes equals total. An invoice that does not add up internally means either a reading error or a real problem, and both need looking at.
The fourth check can justify the project on its own once a company handles any real volume. A manual process does it when someone remembers; an automated one does it on a hundred per cent of invoices, every time.

4Three-way matching: invoice, order and delivery note

If you buy against purchase orders, this stage decides whether the invoice approves itself or goes through a person. This is three-way matching: the invoice is compared with the order (what was ordered and at what price) and with the delivery note or goods receipt (what actually arrived).

SituationWhat the flow should do
Everything matches within tolerance Post and approve with no intervention
Price or quantity difference within the margin Approve and log the variance for later review
Difference above the margin Hold and route to whoever can decide, with the comparison already done
No purchase order Approval workflow by amount and cost centre

Setting the tolerances is a business conversation, not a technical one, and it should happen before anything is built. Too narrow a margin sends everything to review and the project saves nothing; too wide and it lets errors through.

5Posting into the ERP

The entry needs an account, a cost centre and tax treatment. There is an underlying decision here worth taking consciously.

📖
Explicit rules
"This supplier always goes to this account." Predictable, auditable and easy to explain. It covers most cases with little effort and it is where to start.
🧠
Suggestion from history
The system proposes the allocation based on how similar invoices were posted. Useful for the remaining cases, but as a reviewable proposal, not a closed decision.
🔌
How it enters the ERP
Through an API if there is one, file import if not, and interface automation only as a last resort. That last one works, but it is fragile: a screen change in an ERP update can break it.
↩️
What happens when it fails
The stage has to be retryable without duplicating the entry. It is the technical detail that causes the most trouble afterwards and gets the least thought beforehand.

6Approval and exceptions

Invoices that do not approve themselves need a workflow of their own. And that is where the time the project meant to save gets lost, so it deserves proper design.

What works: notifying the specific person — not a shared mailbox — giving them everything needed to decide inside the notification itself, allowing approval without logging into the ERP, and escalating only when a deadline passes. What does not work: a daily email with twenty pending items and a link to an inbox where you have to go looking.

A well-built approval flow tends to cut cycle time more than the data extraction does. An invoice that takes ten days usually is not late because nobody read it, but because it spent five of them waiting on the wrong desk.

7Filing, traceability and payment

The last stage gets forgotten and is then missed. The original document must be stored and linked to the accounting entry, with a record of who did what and when. That is what turns an audit into a five-minute query.

On payment itself, automate the proposal — what gets paid this week, based on due dates and early-payment discounts — and keep execution under human control with dual validation. This is money going out: not the place to save yourself a check.

One more piece of data is worth having ready: once mandatory B2B e-invoicing in Spain applies, Royal Decree 238/2026 requires the receiving company to report whether each invoice is accepted or rejected and the date it is paid in full. A workflow that already records those statuses at approval and payment has that part largely solved.

8How to roll out the supplier invoice workflow

01
Start with the usual twenty
In most companies a handful of suppliers account for a large share of invoices. Automating those first delivers results sooner and produces the learning needed for the rest.
02
Coexist with the manual process
For a while, both paths run in parallel. Attempting a big-bang switch is the fastest way for the team to lose confidence in the system.
03
Measure before you start
Invoices per month, minutes per invoice, percentage of exceptions and cycle days. Without that baseline you will not be able to show afterwards that it was worth anything.
04
Design the exception on day one
What happens when the flow does not know what to do. If the answer is "it stops", the project will last until the first odd invoice.
Use caseInvoicingAutomation

Frequently asked questions

Do I need to change ERP to automate invoices?

No, and it is almost never a good idea. The flow connects to the ERP you already have, through an API if it offers one or file imports if not. Changing ERP is a project of another order of magnitude, and mixing it with automation multiplies the risk of both failing.

What is three-way matching of invoice, order and delivery note?

It means checking that the invoice agrees with the purchase order (what was ordered and at what price) and with the delivery note or goods receipt (what actually arrived). If all three match within the agreed tolerance, the invoice can be posted with no intervention; if not, it is held with the difference already flagged. For purchases with no receipt to record, such as many services, the match is two-way, invoice against order, and without an order the invoice goes to the approval workflow.

What about invoices that arrive on paper?

They are scanned and enter the same workflow, provided the scanner leaves the file in a fixed destination rather than in someone's personal folder. Paper does not prevent automation; what prevents it is digitisation depending on a person remembering where to save things.

Can the system post invoices with nobody reviewing them?

Yes, for invoices that pass every validation and match their purchase order and delivery note within tolerance, which in a well-tuned workflow can be a large share of the volume. The rest goes through a person. That split is exactly the point of the design: people spend their time on exceptions rather than typing in normal invoices.

How long until a first working version?

It depends on the number of intake channels, whether there is order matching and how accessible the ERP is. A narrow scope — one channel, the main suppliers, no line items — goes live considerably sooner than the full workflow, and it also teaches you what the rest genuinely needs.

Keep reading

In short

The workflow has seven stages and data extraction is only one of them. Projects that stop there tend to end up with a very good OCR feeding a process that is still manual from the second step onwards.

The stages that pay back most tend to be the least glamorous: normalising intake, validating against master data and building an approval workflow that reaches the right person. That is how we approach AI process automation.

How many supplier invoices do you process a month?

With that figure, the channels they arrive through and your ERP, the full workflow can be sketched out and we can say which stage to automate first. We review it with you, no strings attached.

Request a process analysis

Nexum Blog

Related articles

How much does a SCADA system cost?

24 Jun 2026

How much does a SCADA system cost?

How much a SCADA system costs in 2026: price ranges by size, cost factors (tags, licences, engineering), cloud vs on-premise, TCO and ROI explained.

Read article
ATEX Electrical Panels: Hazardous Area Guide

22 Jun 2026

ATEX Electrical Panels: Hazardous Area Guide

ATEX electrical panels for hazardous areas: zone classification, Ex marking, protection modes and Directive 2014/34/EU, explained step by step.

Read article
Automating Processes in Line with the GDPR

16 Oct 2026

Automating Processes in Line with the GDPR

How to automate processes in line with the GDPR: the five usual failures, Article 22, the processor agreement and a flow with no legal surprises.

Read article