Invoices and Receipts: The Classic Small-Business Automation

If a business implements exactly one automation, invoice and receipt processing is usually the right one. It meets every criterion that makes a process automatable: high frequency, clear rules, structured target systems — and it reliably annoys everybody involved.

Many projects still fail at the same point: the assumption that "automated" also means "unsupervised".

What actually happens

The process in question is unspectacular: an invoice arrives by email, PDF attached, somebody reads out supplier, date, amount, tax rate and invoice number, assigns an account and a cost centre, files the document and passes it to accounting.

Automated, it looks like this: the inbox is watched, attachments are detected, fields are extracted, checked against master data and handed to the target system. Anything that cannot be assigned with confidence lands in a review queue.

That last sentence is the decisive one. A system without a review queue is not an automated process; it is an error source with throughput.

What is realistically recognised

For clean PDF invoices from recurring suppliers, extraction is very reliable today. It gets harder with:

• Photographed receipts — fuel, hospitality, thermal paper
• First-time suppliers — no master record to anchor the assignment
• Consolidated invoices — several cost centres in one document
• Credit notes and partial invoices — sign and reference are the most common silent error source

So the realistic expectation is not "everything goes through" but: the large majority goes through and the rest is made visible instead of overlooked. That is already the entire gain — the review queue is shorter than the original stack.

The part you do not automate away

Substantive approval stays. No system can judge whether a service was actually delivered. Automation replaces the typing, not the sign-off.

Retention obligations stay. Requirements on immutability, auditability and retention apply unchanged, even when the route there is automated. How that applies in a specific case belongs with your tax advisor — it is not a point to decide in passing.

The exceptions stay. And they stand out more than before, because they are now a visible list instead of disappearing into a stack. That feels like more work at first. It is the same work, sorted.

Frequently asked questions

Can't I do this with my accounting software's own feature?

Often yes — most systems now include document recognition. So the honest first step is to properly test the existing feature before building something custom. A dedicated workflow mainly pays off where several sources converge or bespoke rules are needed.

What is the recognition rate?

That depends too heavily on your document mix for a number to mean anything. The answer only becomes defensible through a test run on your own documents from the last three months — including the awkward ones.

What happens when it gets something wrong?

That is what the review queue is for. What matters is that uncertain cases are held rather than slipping through with a default value. A system that guesses when in doubt is worse than none.

Where does the data sit?

With invoices this is not a side question: they contain business and sometimes personal data. Processing location, processing agreement and deletion periods belong settled before the build.

What you can do today

1. Collect a typical month of documents and sort them by type — PDF invoice, photo, consolidated invoice, credit note.
2. Count what share falls into the simplest type. That is your realistic starting volume.
3. Measure the time per document for pure data entry, excluding substantive approval.
4. Run the capacity value through the ROI calculator — including build and maintenance effort; see the business case.
5. Test the feature your accounting software already ships with first.

If it turns out several systems have to be joined up, it is an integration project. How we build those is described under AI & automation — or in a first call.