Back office automation with OCR has been promised for twenty years. Most of it failed on the same point: the software could read a document it had seen before, and nothing else. What changed between 2021 and today is not the scanning. It is the model behind it.
What back office automation with OCR actually means today
OCR converts pixels into characters. That part has been solved for a long time and is worth almost nothing on its own. The value sits one layer up: deciding what kind of document this is, which fields matter, whether the numbers are plausible, and whether a human needs to look at it. That layer is what the industry now calls intelligent document processing.
A back office runs on documents it did not design. Supplier invoices in four hundred layouts, medical reports from twelve hospital systems, claims correspondence written by hand. The question is never whether a machine can read a page. It is whether it can read a page it has never seen.
Why template based OCR breaks in the back office
- Every new sender needs a new template, so the maintenance cost grows with the business instead of shrinking.
- Layout changes break extraction silently. The field is still found, the value is wrong.
- Anything handwritten, scanned at an angle, or stapled into a multi document PDF falls out of the process.
- The exception queue grows faster than the automated volume, which is the moment most projects are quietly stopped.
We wrote up the difference in detail in the new OCR and measured it against general purpose models in the document AI benchmark.
Where automation pays off first
Not every document type is worth automating. The ones that are share three traits: high volume, repetitive handling, and a downstream system that wants structured data. In our customer base these are the clusters that reach production fastest.
- Accounts payable: incoming invoices with three way match against order and delivery note.
- Claims handling: first notice of loss, estimates, invoices and correspondence in one case file.
- KYC and onboarding: identity documents, proof of address, company registers.
- Sales order processing: purchase orders arriving as PDF, fax or email body.
- Referrals and physician reports: unstructured clinical text that has to land in the right patient record.
- Digital mailroom: everything that arrives centrally and has to be split, classified and routed.
The four numbers that decide whether it worked
Most automation projects are reported on activity rather than on outcome. Four figures tell you the truth.
- Automation rate. Share of documents that pass through without a human touching them. Below 70 percent the business case is thin, above 90 percent it carries itself.
- Field accuracy on the fields that matter. An average across forty fields hides the fact that the IBAN is wrong.
- Cost per document, fully loaded. Licence, infrastructure and the remaining manual handling together, not the licence alone.
- Time to first production document type. If it takes six months, the scope was wrong.
What regulated industries have to add
In banking, insurance and healthcare the model quality is only half the decision. The other half is where the data goes and what you can prove afterwards.
- Processing location. Documents stay in Switzerland or in the EU, with no transfer to a provider outside that perimeter.
- Audit trail. Every extracted value traceable to the page region it came from, every correction logged.
- DORA and the operational resilience requirements that come with it, including exit scenarios and subcontractor transparency.
- No training on customer documents unless it is contractually agreed.
That is the reason banks, insurers and hospitals run document AI differently from a marketing department automating its expense receipts.
Outsourcing or automation
Sending back office work to an external provider still makes sense for volumes that are genuinely irregular. For everything repetitive the arithmetic has moved. Outsourcing scales by adding people, so unit cost stays flat. Automation scales by adding volume, so unit cost falls. The dependency, the compliance overhead and the quality control effort stay with you either way.
A realistic way to start
Take one document type with real volume, run a few hundred historical documents through a system, and compare the extracted values against what your team actually booked. That comparison takes a day and settles the argument better than any vendor deck.
You can do exactly that on your own documents in the platform, or book a demo and we will run the first batch with you.
