AI order and invoice processing is as accurate as the way you measure it. The touchless rate counts documents that reach the ERP without a person. The correct rate counts documents that arrive right. A good system keeps the correct rate at or above your manual baseline and flags what it is unsure of, so the touchless rate only rises as trust is earned.
The question comes up on almost every call we have with an order desk or finance team: how accurate is it compared with our own people, and what happens if it invents an article number. At Lleverage we think the touchless rate on its own is the wrong headline. A system that guesses can push its touchless rate up by sending wrong orders straight into the ERP, and nobody sees the error until the customer does. We measure the correct rate first and let the touchless rate grow only as fast as the correct rate allows.
That is how our AI agents for order processing and accounts payable are set up: every document is checked against your customers, items, prices and purchase orders, and anything the agent is not sure of waits for a person with the reason attached. This article defines the numbers that matter, sets out what a manual baseline looks like, and explains how to test accuracy on your own documents before anything goes live. If you would rather see it on your own orders and invoices, book a demo.
What is the touchless rate in AI order and invoice processing?
The touchless rate is the share of documents that go from the inbox to the ERP without anyone opening, correcting or approving them. For orders it means the sales order is created and confirmed on its own. For supplier invoices it means the invoice is matched, inside tolerance and booked with no human step.
It is the number most vendors lead with, because it maps directly to hours saved. Ardent Partners' State of ePayables 2025 report uses the same idea under the name straight-through processing, and found that the best-performing accounts payable teams process more than 1.8 times as many invoices straight through as their peers.
The touchless rate says nothing about whether the documents that went through were right. A wrong delivery address, a unit of measure read as boxes instead of pieces, or an invoice booked against the wrong purchase order all count as touchless. That is why it cannot stand alone.
What is the correct rate, and how does it differ from the touchless rate?
The correct rate is the share of documents whose final record in the ERP matches what the customer or supplier meant, every field included. It counts a document as correct only when nothing about it needed fixing later: no amended order, no credit note, no call from the warehouse asking what was meant.
The difference matters most in one corner: documents that went through untouched and were wrong. Those are silent errors, and they are the ones that cost money, because nobody looked. A document that was flagged and corrected by a person is slower, but it is not an error that reaches a customer.
| Metric | What it counts | What it hides |
|---|---|---|
| Touchless rate | Documents that reached the ERP with no human step | Whether those documents were right |
| Correct rate | Documents whose final ERP record needed no later fix | How much human effort it took to get there |
| Field accuracy | Individual fields extracted correctly | Whether whole orders or invoices were right |
| Flag rate | Documents sent to a person, with a reason | Whether the flags were needed |
| Silent error rate | Touchless documents that turned out wrong | Nothing: this is the number to drive towards 0 |
| Manual baseline | Your own team's correct rate on the same documents | It is rarely measured, so it is often assumed to be 100% |
Read together, the numbers tell you what you need. A high touchless rate with a low silent error rate means the agent is saving time safely. A high touchless rate with a rising silent error rate means it is guessing. A low touchless rate with a very low silent error rate means it is careful, and the threshold can move once you have seen enough clean runs.
How accurate is manual order and invoice entry today?
Manual entry is far from error free, and most companies have never measured their own rate. Published research on hand-keyed data puts field error rates at a few percent, with free text fields the worst. On the invoice side, nearly 1 in 5 invoices needs a person before AI enters the picture at all.
The clearest study we found is a 2013 paper in BMJ Open by Hong and colleagues, which compared hand-keyed pathology records with electronic imports of the same reports. Across all fields the error rate was 2.8%, with individual fields ranging from 0.5% to 6.4%. Fields in text format were significantly more error prone than numbers and measurements. Only 76% of the records had every field right.
Those were hospital records, not sales orders, but the pattern carries over to an order desk. The fields that go wrong are the ones that need interpretation: a product description with no article number, a delivery instruction in the email body, a customer reference that could mean 2 different sites.
For invoices, Ardent Partners' State of ePayables 2025 report puts the average invoice exception rate for accounts payable departments at 18.4%. An exception there is usually a real disagreement in the data, such as a price that differs from the purchase order or a quantity that was never received. No reader, human or AI, should book those without a person.
That gives you the honest comparison. The fair test is whether the agent is more accurate than your team on the same documents, and whether its errors are flagged rather than hidden.
Why can field accuracy look high while orders still go wrong?
Field accuracy averages over many fields, and most of them rarely go wrong. A purchase order with 20 fields can be 95% right and still be wrong as an order, because the one wrong field was the article or the delivery date. Measure accuracy per document, and weight the fields that cost money.
The BMJ Open study shows the gap clearly: a 2.8% field error rate still left nearly a quarter of records with at least one mistake. An order desk sees the same thing. Customer name, order date and reference number are almost always right. Article matching, units, prices and delivery addresses are where errors live, and they are exactly the fields a dashboard of average field accuracy hides.
When you evaluate any system, ask for 3 numbers per document type: the correct rate per document, the accuracy on the fields that cost money, and the silent error rate. If a vendor can only give you field accuracy, ask how many whole orders were right.
What happens if the AI invents an article number?
In a well-built system it should not get the chance. The agent matches every line against your own item master and order history, so it can only choose an article that exists. When no match is confident enough, the line is left open and flagged for a person. Flag, don't guess, is the design rule.
The published stories show what that looks like in practice. At SIG Benelux, around 3% of orders carry an article number and roughly 47% have only a description. The agent matches lines using years of order history, and it flags low-confidence lines for review instead of guessing. At Microtechniek, every extraction and every match carries a confidence, and anything below the line goes to a person. At Topa Bathroom Products, an unrecognised article code is flagged for human review.
Xpol took the rule furthest. Its agent prepares a draft order in Business Central for a person to sign off, and ambiguous items are left blank rather than guessed. Each blank field also points at something to clean up in the source data, such as a framework contract with overlapping dates. Cees Maaskant, Xpol's general manager, describes the reason for building it that way:
"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
The same rule applies to invoices. The agent matches each line three ways, against the purchase order and the goods receipt, and only the lines that genuinely disagree reach a person. Releasing a payment always stays with a person. If you want the technical side of why a model can be made to behave this predictably, we wrote about getting deterministic outcomes from LLMs.
How do you test AI accuracy before it goes live?
Run the agent in shadow on your real documents first. It processes the same orders or invoices as your team, in parallel, without writing anything to the ERP. You then compare its output, field by field, with what your team booked and what was corrected later. That gives you both correct rates on the same documents.
SIG Benelux worked this way. The second half of its project was building and shadow running: the automation processed order emails alongside the sales desk, so everyone could see where it failed without any customer noticing. Microtechniek set the same bar, and the agent had to prove it could handle real order data, including inconsistent customer abbreviations and multi-location orders, before it touched a live order.
A shadow run is only as good as its comparison. These steps keep it honest:
- Use the last few weeks of real documents, including the awkward ones, not a clean sample picked for the test.
- Compare against what was finally booked, after amendments and credit notes, so your team's own errors show up too.
- Count results per document and per costly field, and record every flag with its reason.
- Look at the silent errors one by one. Each is either a rule to add or a reason to keep that case with a person.
How does trust ramp from review to touchless?
After the shadow run, the agent goes live with a person approving everything, then the threshold moves. In our control layer every process runs at a confidence threshold you choose: above it a person decides, below it the work goes through on its own. It only moves when you move it, never because the model became more confident by itself, and one setting returns a process to full supervision.
Rollout by customer works the same way. Xpol put its largest retailer live first, then the supplier set covering roughly 60 to 70 percent of weekly volume, while the long tail keeps running in parallel until its rules are stable.
What touchless rate should you expect for orders and invoices?
It depends on your documents more than on the AI. Orders with article numbers from repeat customers go touchless early. Orders written as free text, or invoices with real price and quantity differences, should keep a person involved. Expect the touchless rate to start low during review and rise as each customer or supplier proves clean.
The published numbers show the range. Topa Bathroom Products processes over 90% of incoming orders directly into Business Central without any manual input, and some customers receive their confirmation within 30 seconds. SIG Benelux has a much harder mix of descriptions, photos and WhatsApp messages: 73% of order emails reach Dynamics 365 as a draft ready to review, and 88% of lines match the right article first time. Both are good results for their documents.
For invoices, the ceiling is set by your own exception rate. If 18.4% of invoices disagree with the purchase order or the receipt, as Ardent Partners reports for the average department, a touchless rate near 100% would mean the system is booking differences it should flag. The goal is that every invoice without a real difference goes through, and every one with a difference reaches a person with the reason. We compare the ways to get there in our article on invoice matching software.
We see the same pattern across back office automation: the touchless rate is an outcome of clean rules and data, and it improves as the rules that live in people's heads are written into a shared knowledge layer the agent and your team both use.
Frequently asked questions
What is a good touchless rate for AI order processing?
There is no single good number, because it depends on your order mix. Orders with article numbers from repeat customers can run almost fully touchless. Free text and photo orders need more review. Judge the touchless rate together with the silent error rate: a lower touchless rate with no hidden errors beats a high one with errors.
How do I measure my team's manual error rate?
Take a few weeks of booked orders or invoices and compare the original documents with what was finally in the ERP after amendments, credit notes and customer calls. Count every document with at least 1 wrong field, as well as the wrong fields themselves. Most teams have never done this, and the result gives you the baseline the AI must match or beat.
Can AI order processing invent a SKU or customer?
Not when it is built to match against your own data. The agent chooses articles and customers that already exist in your ERP and order history. When no match is confident enough, it leaves the field open and flags it for a person instead of guessing. A wrong guess costs more than a short review.
Is AI invoice processing more accurate than manual entry?
The honest answer comes from a shadow run on your own invoices. An agent checks every line against the purchase order and goods receipt every time, which a busy team under month-end pressure cannot always do. Invoices with real differences still go to a person, and releasing a payment always stays with a person.
How long should a shadow run last before going live?
Long enough to see your normal variety of documents, including month-end peaks and the awkward customers or suppliers. For most order desks that means a few weeks of real traffic. Go live with a person approving everything first, then move the threshold per customer or supplier once their documents run clean.
See the numbers on your own documents
The accuracy question is best answered with your own documents. Bring a handful of real orders or supplier invoices to the demo, including the ones your team dreads, and we will show the agent reading them, matching them against your data and flagging what it is unsure of. We will also walk through how a shadow run on your traffic would measure the correct rate against your team's, and where the threshold would start. Book a demo and bring your order desk or AP lead.