AI document processing is software that reads a business document in any layout, works out what each value means, checks it against the data in your ERP and writes the result back as a record. Template OCR copies text from fixed positions on a page. AI document processing reads the page for meaning, so a new supplier layout needs no new template.
For a back office, that difference decides who does the typing. A customer order lands as a PDF, an invoice as a scan, a confirmation as three lines in an email. Today someone reads each one and keys it into Business Central, SAP or Exact. At Lleverage we think the reading is now the simple part. What counts is what the system does with the record afterwards: which checks it runs, what it writes into the ERP, and which documents it hands to a person before anything is final.
This post is for operations and finance leads who own that work. It covers the documents a back office handles, what our AI agents write back for each of them, how to measure accuracy and where a person still approves. For the category itself, our guide to intelligent document processing covers the vendors and the five stages. If orders are your bottleneck, see our order intake agents, and you can book a demo to see them on your own documents.
What is AI document processing, and how does it differ from template OCR?
AI document processing reads a document the way a clerk does. It recognises the document type, finds the customer, the lines, the quantities and the dates wherever they sit, and checks each value against what your systems already hold. Template OCR turns pixels into characters and copies them from zones someone drew for each layout.
The practical difference shows up on the day a supplier redesigns its invoice. A template breaks, because its fields have moved. Character recognition still produces text, but it does not know which number is the VAT and which is the line total. A model that reads for meaning finds the total wherever it sits, in Dutch or in English.
| Approach | What it reads | Setup per layout | What comes out | Where it breaks |
|---|---|---|---|---|
| Template OCR | Fixed zones on a known layout | One template per sender | Text copied into mapped fields | A new or changed layout |
| Rule-based automation | Data that is already structured | Rules per format | Keystrokes into a screen | Free text, photos, a changed screen |
| AI document processing | The whole document, read for meaning | None per layout | Fields checked against your ERP data | Unreadable or ambiguous documents |
| AI agent on top | The document plus your rules and history | Rules per customer or supplier | A record in the ERP, or a question to a person | Cases your rules do not cover yet |
Character recognition still has a job: it turns a scanned image into text a model can read. The step that saves hours comes after it. We compare the two on invoices in template-free AI versus OCR for invoice data capture.
Which documents does a back office run through AI document processing?
Most back offices handle 5 document flows that repeat every day: customer orders, supplier invoices, order confirmations, delivery notes and certificates. Each arrives from many senders in many layouts, and each ends in a record someone keys by hand. The more senders a flow has, the more a template approach struggles and the more reading for meaning pays.
| Document | Who sends it | What gets checked | Where it ends up |
|---|---|---|---|
| Customer order | Customers, by email, PDF, Excel or portal export | Customer, articles, prices, delivery date | A sales order, or a draft for review |
| Supplier invoice | Suppliers | Purchase order, goods receipt, tolerances | An invoice ready to post, or an exception |
| Order confirmation | Suppliers, against your purchase order | Prices, quantities and dates against the order | Updated purchase order lines, or a flag to the buyer |
| Delivery note | Carriers and suppliers | Delivered quantities against the order | The goods receipt, or a flagged difference |
| Certificate | Suppliers and labs | Batch, specification values, expiry | A record linked to the batch, or a hold |
Most of our published customer stories start with customer orders. SIG Benelux, a building materials distributor, receives orders where around 3% carry an article number and roughly 47% have only a description. Order history does the translation, and 88% of lines match the right article first time. At Microtechniek, roughly half of what lands in the shared inbox is not an order at all, so the first step is classification: timesheets and quote requests go one way, purchase orders the other.
Supplier invoices and confirmations sit on the buying side. Our purchase order confirmation automation post covers confirmations in detail. Certificates have their own rules per product, which we cover in certificate of analysis software.
What does the AI agent write back into the ERP?
The agent writes the same fields a person would fill, through the ERP's own interfaces, so the ERP validates the record the way it validates a person's work. For a customer order that means a sales order with the customer, delivery address, delivery week and every order line. For a supplier invoice it means an invoice matched to its purchase order and receipt.
Topa Bathroom Products shows what that looks like for orders. Its agent reads every order in the inbox, plain email, PDF, CSV or Excel, and creates the sales order directly in Business Central. Over 90% of incoming orders are processed into Business Central without any manual input, and some customers receive their confirmation within 30 seconds. The team members who used to key orders, about 3.8 full-time equivalents, moved to after-sales support and service planning.
Write-back also works where the ERP offers no standard route in. Microtechniek runs Ridder iQ on-premise, and between 45 and 71 purchase orders a week are written into it without anyone typing them, around 60. Each one took about 10 minutes by hand. The same approach works with the ERPs we connect to, from Business Central to SAP.
What the agent writes stays inside the limits you set. A sales order is created and confirmed. An invoice is posted only when it falls within your tolerances. On the accounts payable side the agent never releases a payment: that stays with your finance team.
Where does a person still approve before anything is posted?
A person approves wherever the agent is unsure, wherever a rule says so, and wherever a mistake would cost more than the check. The agent scores how sure it is of each document. Below the line a person decides, above it the work goes through on its own, and only your team moves that line.
Some customers keep a person on every document by design. Xpol, a fresh flower wholesaler, chose an agent that prepares a draft webshop order in Business Central and never posts a final sales order on its own. A person signs off each draft with a single click, and ambiguous items are left blank for that person to fill in. Across 150 orders a week, that saves about 20 minutes per large order. SIG Benelux works the same way: 73% of order emails arrive in Dynamics 365 as a draft ready to review, and the salesperson still approves each one.
Other flows move further. Topa's agent creates the order itself and still stops on the cases it should not decide. When a commission number is missing, it drafts the email to the customer. When an article code is unknown, it flags the order for review.
When a person does step in, the evidence is already in front of them. In Lleverage the request opens beside the source PDF, and hovering over a field highlights the matching value in the document, so the reviewer checks each value at a glance. Our post on supervised agents and background automation explains how to choose between the two set-ups.
"It's a matter of building trust in the organisation with these kinds of initiatives. You can't just throw something like this over the fence." Cees Maaskant, General Manager, Xpol
How is AI document processing accuracy measured?
Accuracy in AI document processing is measured per document flow with 2 numbers. The touchless rate counts documents that reach the ERP without a person. The correct rate counts documents whose final ERP record needed no later fix. A good set-up keeps the correct rate at or above your manual baseline before the touchless rate grows.
Field accuracy on a test set is a common headline figure and a weak one to buy on. A document with 19 correct fields and 1 wrong delivery address is still a wrong order. Count whole documents, on your own inbox, over a few weeks.
| Metric | What it counts | Why it matters |
|---|---|---|
| Touchless rate | Documents that reach the ERP with no human step | Shows how much work left the desk |
| Correct rate | Documents whose final record needed no later fix | Shows whether that work was right |
| Flag rate | Documents sent to a person, with a reason | Shows where rules or data need work |
| Silent error rate | Wrong documents that went through untouched | The number to keep near zero |
Two things move these numbers more than the model does. The first is context: which customers, articles and prices exist, and which customer rules apply. Xpol's agent applies 25 customer rulesets for unit conversions, weekday-specific label text and multi-depot splits. The second is the confidence line. A higher line sends more documents to a person and lets fewer errors through. Where to set it is a business decision about the cost of a wrong record. We go deeper in touchless rate vs correct rate.
How common is AI document processing in European companies?
Adoption is still early. According to Eurostat's survey on ICT usage in enterprises, 19.95% of EU enterprises with 10 or more employees used AI technologies in 2025. Only 11.75% used AI that analyses written language, the kind of AI that reads documents.
The gap by size is wide. In the same survey, 17% of small enterprises used AI in 2025, against 55.03% of large ones. Mid-sized manufacturers and wholesalers sit in between, often with document volumes that keep several people busy.
At Lleverage, our forward deployed engineers build the first process with your team, in your ERP, and it typically runs supervised within a few weeks.
How should a back-office team roll out AI document processing?
Start with 1 flow: the highest-volume document that ends in someone retyping. Run the agent next to the current process, let a person approve every record for a few weeks, and measure the correct rate on your own documents. When the numbers hold, let the routine documents through and keep people on the exceptions.
Three tests pick a good first flow. It should be high volume, so saved minutes add up. It should have a clear definition of correct, so the agent's work can be checked. And it should end in your ERP, because a document read and left in a folder saves nobody any time.
Xpol followed that path. Its largest retailer went live first, then the supplier set covering roughly 60 to 70 percent of weekly volume, with long-tail customers added as the rule library grew. SIG Benelux started with one location, one process and the central order mailbox, and ran the automation alongside the team before any customer saw it.
Once the first flow runs, the next one reuses the same ERP connection and the same review set-up. Orders often come first, then invoices through AP automation, then supplier confirmations through procurement automation.
Frequently asked questions
What is the difference between AI document processing and OCR?
OCR turns an image into characters and, with templates, copies them from fixed positions on a page. AI document processing reads the whole document for meaning, finds each value wherever it sits and checks it against your ERP data. It handles a new layout without setup, where a template-based system needs someone to build a new template first.
Which documents can AI document processing handle?
Any recurring business document with a stable meaning and a changing layout. In a back office that usually means customer orders, supplier invoices, order confirmations, delivery notes and certificates, sent as PDF, Excel, CSV or plain email text. The more senders a flow has, the more an agent saves compared with one template per sender.
Does AI document processing post records into our ERP?
It writes into the ERP within the limits you set. Sales orders are created and confirmed, invoices are posted when they fall inside your tolerances, and anything the agent is unsure of waits for a person. Some teams have every record prepared as a draft that a person approves with one click.
How accurate is AI document processing compared with a person?
Accuracy is measured per document flow on your own inbox. Count the correct rate, the share of records that needed no later fix, alongside the touchless rate. Start with high-volume repeat documents, where a few weeks give enough examples to judge. Rare documents from new senders need more supervision until the rules cover them.
How long does it take to get AI document processing running?
With Lleverage, the first process typically runs supervised within a few weeks, built by our forward deployed engineers with your team. Records are approved by a person at first. Once the correct rate holds on your own documents, routine documents go through on their own and people keep the exceptions.
Start with the flow that costs you the most hours
Pick the document your team retypes most often, collect a week of real examples, including the awkward ones, and bring them to a call. We will show what the agent reads, what it would write into your ERP and which documents it would hand to a person. Book a demo and we will plan the first process with you.
