Automating bookkeeping with AI: our complete setup (Qonto to DATEV)

Bookkeeping used to be the process everyone here hated. Now it almost runs by itself. If you want to automate your bookkeeping with AI and a DATEV export, here is our complete setup.

No concept, no pitch. The system runs at Flowhouse and our brands every day. I will walk you through every building block, what our tax advisor says about it, and where it gets stuck.

Why we built it ourselves

After 12 years at PwC I thought I knew bookkeeping processes. Then I collected receipts as a founder myself. Every month the same routine: go through the bank statement, hunt for receipts, send everything to the tax advisor.

With three brands plus the agency that is a lot of transactions. Meta ads, fulfillers, Shopify payouts, software subscriptions, inventory purchasing. Almost all of it repeats. And repetition is the signal for automation, as I described in the true cost of manual processes.

We looked at off-the-shelf tools. They could do a lot, and never exactly our case: several companies, Qonto as the bank, DATEV as the target. So we built it ourselves. Why we often build our own tools is in why we build our own tools.

Step 1: pull transactions automatically

The base is our business account at Qonto. Qonto has a clean API. Our system pulls all new transactions through it on a regular schedule.

Every transaction arrives with everything we need: amount, date, counterparty, payment reference. Often a receipt is already attached, because Qonto can store receipts directly on a transaction.

That sounds trivial. It is not. Most founders export a CSV once a month and work through it by hand. On our side every transaction sits in the system a few hours after it is booked.

Step 2: rules first, then AI

Automating bookkeeping with AI does not mean handing everything to the AI. The order is what matters.

Fixed rules run first. A payment to our fulfiller is always close to cost of goods. A Meta invoice is always advertising cost. The Hetzner invoice is always servers. For cases like these you need a rule instead of a model. Rules are deterministic, traceable and free.

The AI comes after that. Anything no rule matches goes to a language model. It sees the payment reference, the counterparty and the amount, and proposes a category with a reason.

On our side the rules catch the bulk of transactions, because in ecommerce almost everything repeats. The AI handles the rest: the one-off trade show booth, the new software, the refund.

Important: the AI's proposal is a proposal. Uncertain cases land in a review list and a human looks at them. That is the same logic we use in customer service, see AI in customer service: AI as a filter, never as a replacement.

Step 3: match receipts automatically

A classified transaction without a receipt is only half the job. The tax office wants the receipt.

Our receipt scanner collects receipts from several channels, among them a Google Drive folder where invoices land. The AI reads out every receipt: issuer, amount, date, invoice number.

Then comes the interesting part: the system looks for the matching bank transaction and attaches the receipt to it. Amount and date have to line up, otherwise there is no match. How that works in detail is in capturing and checking receipts automatically.

The result: no more shoebox session at the end of the month. Most transactions already have their receipt by the time we look at them for the first time.

Step 4: export to DATEV

Our tax advisor works with DATEV, the accounting standard German tax advisors use. So our system has to speak DATEV.

At the end of the period the system produces an export in DATEV format: all transactions with category, account and receipt reference, as a CSV the tax advisor can import directly.

Bookkeeping automation and DATEV go together fine. DATEV is a very well documented target format. The work sits in everything before it: clean classification, complete receipts.

What our tax advisor says about it

At the start he was sceptical. AI and bookkeeping sounded like risk.

Today, instead of a folder full of loose receipts, he gets a finished export with receipts already attached. His questions per month have dropped noticeably. He reviews, he does the final booking, he does the closings. That stays his work, and that is how it should be.

One point mattered to him, and it is also my advice to you: the AI does the preparation, never the bookkeeping itself. Responsibility for correctness stays with you and your tax advisor. Build the system so every step is traceable.

The limits

So you do not get the wrong picture:

• Special cases stay manual. Shareholder contributions, partner matters, fixed assets, complicated chains of reversals. A human decides there.
• The AI gets it wrong sometimes. A cryptic payment reference, an unknown counterparty, and the proposal is off. Without a review step that would slip through.
• Receipts without a bank link are harder. Cash receipts or receipts with a deviating amount do not get matched automatically. They land in the review list.
• Building it costs time. This was no weekend project. Rules do not maintain themselves either, and new suppliers need new rules.
• It does not replace a tax advisor. VAT, depreciation, annual accounts: you still need a professional for that.

What you get out of it

For us the effort clearly paid off. The monthly bookkeeping session shrank from half a day to a short control round. More precisely: we review proposals instead of typing in data.

On top of that comes a side effect I had underestimated: we see our numbers continuously, rather than weeks later at the tax advisor. Classified transactions are real-time controlling as a by-product.

Conclusion

Bookkeeping automation with AI works when you build it in the right order: pull transactions automatically, rules before AI, match receipts automatically, a clean DATEV export, a human as the final authority.

That is the same approach we used to automate the rest of our companies, from fulfilment to support. Automating ecommerce processes gives you the overview.

If you want something like this for your business without building it yourself: that is exactly what we do at Flowhouse. Write to us, we are happy to show you our setup in detail.