Because an order is not one prediction, but a few dozen predictions that all have to be right at the same time. With 95% accuracy on each field, a 10-field order is entirely correct only about 6 times out of 10. To let an order go into the ERP without review, you need much higher accuracy per field, and above all an agent that knows when it isn’t sure.
Automating means letting an order go through without review
At Eledone, we sell automation: orders and quotes that enter our customers’ ERPs with no human intervention. For an order to go into a management system without review, every prediction the model makes has to be correct, and the model has to be certain of it.
An error doesn’t stay in the ERP. It goes to the warehouse, then to the customer. Delivering 1,000 pallets instead of 100 is an original way to surprise a customer, but not the one we’re after.
An order is dozens of fields
Reading an order means filling in a series of fields, each with its own chance of error:
- the customer, and sometimes the entity that pays when it isn’t the one being delivered;
- the delivery address, which may be a building site or another location;
- the reference of each item, often written the customer’s own way;
- the quantity and its unit: piece, carton or pallet;
- the unit price, to be checked against the negotiated price list;
- the requested delivery date, the customer’s order number, any special conditions.
A three-line order already has around fifteen fields. A twenty-line site order has well over a hundred.
The calculation that changes everything
If each field has a 95% chance of being right, the probability that 10 fields are right at the same time is 0.95 to the power of 10, or about 60%. In other words, 4 orders in 10 contain at least one error.
The table below shows the share of entirely correct orders depending on per-field accuracy and the number of fields.
| Accuracy per field | 10 fields | 20 fields | 30 fields |
|---|---|---|---|
| 95% | 60% | 36% | 21% |
| 98% | 82% | 67% | 55% |
| 99% | 90% | 82% | 74% |
| 99.5% | 95% | 90% | 86% |
| 99.9% | 99% | 98% | 97% |
This assumes errors are independent from one field to the next, which isn’t quite true: a badly scanned purchase order often causes several errors at once. But the order of magnitude holds, and it’s unambiguous. A model that looks excellent field by field can still be unusable order by order.
What’s left to certify
At 95% per field, only about 60% of the order and quote flow is left where automating entry is even possible. And even within that 60%, you still need to know which orders are correct.
That is the hardest step, because AI models are probabilistic. A language model can be wrong with great confidence: the wrong answer looks just like the right one. So it isn’t enough to ask it whether it’s sure.
A useful agent isn’t one that is right on average. It’s one that can say: “I’m not certain about this one, I’ll show it to the sales admin team.”
Knowing how to say “I’m not certain”
To certify an order, we don’t rely on a single signal. Each field gets a confidence score, and that score is checked against the company’s real data:
- the price is compared with the customer’s price list in the ERP;
- the quantity is compared with the customer’s habits and the item’s packaging units;
- the item must match exactly one record in the catalogue, without ambiguity;
- the customer and address must exist and be consistent with each other;
- the totals must add up to those on the document.
An order is certified only if everything matches. At the slightest doubt, it is shown to the team with the reason it was flagged: an unusual price, an ambiguous item, a quantity outside the usual pattern. Better to show the team a correct order than to let a wrong one through. We explain how these checks work for items in customer references and for quantities in unit of measure conversion.
Why we chase 1 or 2%
We spend a lot of time gaining 1 or 2% of accuracy per field. From a distance, that can seem pointless: 95% or 97%, what’s the difference?
The table answers that. At 99% per field, 10-field orders are perfect 90% of the time. At 99.9%, that reaches 99%. Every point gained per field is multiplied at the order level, and gives far more room for certification. It’s what takes automation from “the AI prepares, the team reviews everything” to “the team only sees what needs human approval”.
What it changes for your team
For a sales administration team, the right measure isn’t the model’s accuracy, but two very concrete numbers:
- the share of orders that go through without review, because they were certified;
- the number of errors that still get through, which must stay as close to zero as possible.
The first number grows over the weeks, as rules are refined and catalogue data improves. The second must never be sacrificed to push the first one up. The rest of the flow isn’t lost either: it reaches the team already read and prepared, with doubtful fields highlighted, which takes seconds to approve instead of minutes to enter.
Frequently asked questions
Why not simply have the team review everything?
You can, and it’s already a gain over manual entry. But systematic review keeps the team in the middle of the flow: processing time still depends on their availability, and attention drops when you review a hundred correct orders to find one wrong one. Certifying the safe orders focuses review where it matters.
What is a “certified” order?
An order where every field was read with enough confidence and checked against the ERP data (price list, catalogue, customer, ordering habits). The required level is set with each customer. An order that doesn’t meet these conditions is shown to the team.
How do you assess the reliability of an order automation solution?
Ask for accuracy per order, not per field, and test it on your own documents, including the hardest ones. Above all, ask what the tool does when it isn’t sure. You can start by trying Eledone on one of your orders, then read how it works on the order entry page.
Want to see what share of your orders could go through without review?
Book a demo