Finance, 14 October 2026
Qonto: Two AI Agents That Do a Small Business's Books in Minutes, Not Hours
Qonto — a European business-banking and bookkeeping platform, used by more than 600,000 small businesses, freelancers and new companies across eight European markets, most of whom have no finance staff of their own.
Problem
Qonto’s own research found its customers spend up to eight hours a month on financial admin they never trained for: running payroll one manual transfer per pay slip at a time, re-typing numbers from client invoices by hand, and tracking cash flow in a spreadsheet nobody has time to look at properly. “They are not finance experts and they don’t want to be,” says Sophie Cornay, who runs Qonto’s AI Lab. The cost isn’t the money — it’s the owner’s afternoon.
Result
5X
- Payroll runs about 5X faster — a 15-to-20-slip monthly run becomes one bulk-approved transfer instead of fifteen separate ones.
- Client invoice creation takes roughly 3X less time, and one customer generated over 500 invoices in a single batch.
- Bank transfers are sent about 2X faster, with some customers delegating transfers up to €100,000 to the agent.
- One solo entrepreneur told Qonto the agents gave back about half a day of admin every month.
Solution
In November 2025, Qonto’s 15-person AI Lab started building agents directly into the banking app rather than bolting on a chatbot. By late April 2026 it had shipped two to all customers: an “Operator” agent that routes a request to a specialist sub-agent — bulk payroll transfers, drafting invoices straight from a client’s CSV, scheduling payments — and executes it end to end; and an “Analyst” agent that just answers plain questions about past and current transactions (“what did we spend on suppliers last quarter?”). The complex reasoning runs on Claude Opus 4.5 and Sonnet via Amazon Bedrock; simpler steps go to a smaller, cheaper model. Nothing moves money or sends anything externally without the owner reviewing and confirming it first — the agents prepare, the human still signs off.
Could a small team do this?
Yes, largely — this is one of the most buildable cases in this series, because the hard part isn’t the AI, it’s the trust mechanics, and those are simple to copy. A solo operator doesn’t need a banking licence to build the “Analyst” half: a narrow agent that reads a small business’s exported bank CSV or invoices and answers “what did I spend on X” or drafts a monthly cash-flow summary is a few days of work with Claude’s existing tool-use APIs, no infrastructure beyond a spreadsheet import. The “Operator” half — actually drafting and sending real invoices from a client list, or queuing real payroll transfers — is harder only because it needs a safe human-approval step before anything goes out, which is itself a small, well-understood pattern (generate a draft, show a diff, require a click to confirm) rather than a deep technical barrier. What’s genuinely out of reach solo is Qonto’s own banking rails and the regulatory load of actually moving the money; the realistic version for a solo AI operator is the same agent wired to a client’s existing bank export and invoicing tool instead of owning the bank account — which is exactly the kind of narrow, unglamorous admin tool small Georgian businesses and freelancers would plausibly pay a monthly fee for.