How to automate invoice entry at an accounting firm

At most accounting firms, the slowest part of an invoice isn't posting it: it's copying it. Opening the PDF, typing in the supplier, tax ID, date, net amounts, VAT rates, withholding and total, checking it and filing it. Multiplied by hundreds or thousands of invoices a month, and concentrated in the weeks before every quarter-end.

This guide explains how to automate it step by step, what you can expect from artificial intelligence and where you need to be careful.

First, work out what it costs you

Three figures are enough to get an idea:

• Invoices per month coming into the firm.
• Minutes per invoice: opening, typing in, checking and filing. Time ten normal invoices and take the average.
• Hourly cost of the team doing it.

For example, 800 invoices a month at 3 minutes each is 40 hours a month. At €18 an hour, that's more than €8,600 a year just typing in invoices.

Run the numbers with your own figures in our calculator for accounting firms.

The four pieces of the automation

1. Intake: invoices that arrive on their own

The first bottleneck is that invoices arrive from everywhere: email, WhatsApp, paper, USB sticks. You need to centralise:

• A mailbox per client (or a general one) where the client forwards their invoices.
• An upload link for those who prefer to drag and drop files.
• A clear rule: if it doesn't come in that way, it doesn't exist. It takes a quarter to get clients used to it, and then saves many more.

2. Extraction: from PDF to data

This is where AI comes in. A language model reads the invoice and returns structured data: supplier, tax ID, number, date, VAT breakdown by rate, withholding and total. It works with invoices from any software, without per-supplier templates.

Two different cases:

• PDFs generated by invoicing software: the text is already in the file, and extraction is fast and reliable.
• Scans and photos of receipts: they need a text recognition step (OCR) first. It works well with sharp images and less well with crooked or crumpled photos.

3. Validation: never trust the AI blindly

This is what separates a serious automation from an experiment. The AI can misread a number, and in accounting a misread number is a problem. That's why every invoice must pass automatic checks:

• Tax ID check digit: many tax IDs (such as Spanish and Portuguese NIFs and many EU VAT numbers) include a check character that detects reading errors.
• VAT consistency: the VAT for each rate must match its net amount (for example 21%, 10% and 4% in Spain, or 23%, 13% and 6% in Portugal).
• Totals that add up: net + VAT − withholding must equal the total, and that total must appear on the invoice.
• A plausible date: not in the future, within the period.

Anything that fails the checks is flagged for human review. The team stops typing and only reviews the doubtful ones.

A practical rule: the AI should copy, not calculate. Sums and checks are done by the system with fixed rules, not by the model.

4. Integration: getting the data into the accounting software

The validated data has to end up in your accounting software:

• Through an API, if the software offers one.
• With an import file (Excel or CSV) in the format your software accepts. Most accounting software can import journal entries or purchase invoices.

Privacy: where the invoices are processed

Invoices contain personal and financial data about your clients and their suppliers. Before uploading them to any AI tool, check where they are processed and whether they are used to train third-party models. Many popular tools send documents to servers outside the European Union.

An alternative is to run the AI on a private server in the EU, so invoices never leave it. That's the option we use at Onlitions.

What this automation doesn't solve

To avoid false expectations:

• It doesn't replace accounting judgement: the AI extracts data, but allocation and tax decisions remain the firm's.
• It isn't e-invoicing for issuing invoices (such as requirements for certified invoicing software). Here we're talking about the entry of invoices received.
• Very irregular documents (illegible receipts, handwritten invoices) will still need review.

Where to start

1. Measure how much time the team spends typing in invoices (work it out with the calculator).
2. Choose two or three high-volume clients and centralise how their invoices arrive.
3. Test extraction and validation with their real invoices for a month.
4. Roll it out to the rest of your clients once the process is fine-tuned.

How we do it at Onlitions

We have a demo of the extractor: you upload a PDF invoice and within seconds you see the data, with the tax ID validated, the VAT breakdown checked and a warning if something doesn't add up. On a 30-minute video call we try it with real invoices from your firm. More for accounting firms.