Document automation: process PDFs and invoices without retyping
LAST UPDATED ·
Which documents work well
Not every document is equally suitable. This is the order in which we start with clients:
| Document | Source | Destination | Suitability |
|---|---|---|---|
| Sales order | Client email with PDF or Excel | ERP sales order | High: fixed fields and high volume |
| Purchase invoice | Supplier email with PDF | Accounting | High: standard fields, and the market has plenty of experience |
| CMR / proof of delivery | Photo or scan from a driver app | TMS (trip) and invoicing | Medium: photos vary, and the signature has to be recognized |
| Packing slip | PDF or scan at receipt | WMS or ERP (goods receipt) | Medium: lines must match a purchase order |
| Supplier price list | Excel or PDF, quarterly | ERP (items) | Medium: many lines, low frequency, costly errors |
| Handwritten form | Paper | Any system | Low: handwriting is unreliable |
Start at the top. Orders and purchase invoices are the most common first projects, because the volume is high and the fields are fixed.
OCR versus AI extraction in plain English
OCR (optical character recognition) turns an image into text. It knows the page says “Item no. 4471-B, 24 units,” but not that this is an item number and a quantity. Classic OCR solutions work with templates: for each sender you set where each field sits. With 40 suppliers you have 40 templates, and a new layout breaks it.
AI extraction reads the text and understands its structure. You ask: “Give me the client, the order lines with item number and quantity, and the requested delivery date.” The model finds those fields whether they sit on the left, on the right, or in a table. No templates. New sender, same question.
In practice we combine the two: OCR to turn a photo into text, AI extraction to turn text into fields.
Confidence and the review queue
No reader is 100% certain. That is why we build every document flow with a threshold and a queue.
- The document comes in and is read.
- Each field gets a confidence score.
- The values are matched with the ERP: does the client exist, does the item number exist, does the total match the lines.
- Everything above the threshold with a clean match goes through automatically.
- Everything below it lands on a review screen: the document on the left, the fields on the right, the doubtful fields highlighted. An employee corrects and approves.
- Corrections are logged, so recurring problems become visible.
The ratio differs per document type. In our cases, 75% to 80% of clean PDF orders go through without correction. For CMR photos it is lower: more doubtful fields go to a person. Those percentages are illustrative. The principle is that a person only looks at what needs a person.
Connecting to ERP and accounting
Reading is half the work. The other half is writing the data to the right place.
- Through an API. Exact, AFAS, Twinfield, SnelStart, King, Unit4, and most modern TMS/WMS packages have an API. The order or invoice is created directly, linked to the right client and items.
- Through an import. Older packages without an API often accept a CSV or XML import. The flow prepares the file; the package reads it in.
- Through a bridge. For custom software or an Access database, we build a small app in between that delivers the data in the right format.
In every case, the ERP stays the source of truth. The document flow writes to it but does not decide.
The results
Two examples from our example cases, with illustrative figures:
- A technical wholesaler (45 employees) went from 42 to 9 hours a week on order entry, with order confirmation from 3 hours to 12 minutes and line errors from 2.6% to 0.4%. See the case order entry from email.
- A transport company (60 employees) went from 22 to 6 hours a week on CMR processing, invoices 3 instead of 12 days after delivery, and saw disputed invoices fall from 6% to 1.5%. See the case from CMR to invoice.
The pattern: fewer hours, shorter lead times, fewer errors. With invoicing, a shorter time to payment comes on top, and you see that in your bank account.
Pitfalls
Handwriting. A handwritten quantity on a CMR is often misread. Solution: a low threshold, the queue, and in time a driver app set up so that quantities are typed in.
Changing layouts. AI extraction catches most of it, but a supplier that swaps the “quantity” and “price” columns can confuse the matching. Solution: match with the ERP (the total has to add up) and flag differences.
Duplicate documents. The same invoice emailed twice, or an order followed by a “correction” with the same reference. Solution: check the reference number and sender before writing anything.
Attachments that are not documents. Terms and conditions, logos, signature images. Solution: a classification step that decides what is an order and what is not.
Wanting everything at once. Start with one document type and one client group. Expand when the queue stays small.
The document flow
As a process diagram: Mailbox or folder → Read (OCR + AI extraction) → Match with ERP → Review screen (human) → Write → Confirmation. In the old situation the bottleneck is the block where a person reads and retypes. In the new situation the highlighted block is the review screen, where the person sees only the doubtful cases.
FAQ
FAQ
Does document automation work with scans and photos?
Yes, if the image is readable. Typed text in a sharp photo is reliable. Handwriting and poor photos enter the review queue more often.
What happens when the AI reads a field wrong?
Every value it reads gets a confidence score. Below the threshold, the document goes to an employee who checks it. Everything is also matched with your ERP: an item number that does not exist always falls out.
Which systems can receive extracted data?
Options include Exact, AFAS, Twinfield, SnelStart, King, Unit4, and most TMS and WMS products through an API or import. A file import or small bridge app works when no API exists.
