Qonto's bank agents on Claude: 2x faster transfers, user approves

Qonto, a business bank serving more than 600,000 companies across eight European markets, built AI agents into its app that prepare transfers, draft invoices and answer questions about transactions, according to a customer story Anthropic published (undated; read on 25 October 2026). The page is vendor-published, so the figures below are Anthropic's account of Qonto's results.

What Qonto reports

• Bank transfers are 2x faster, with no manual beneficiary search, amount entry or reference copying.
• Client invoice creation takes 3x less time, with agents drafting and the customer confirming before sending.
• Payroll runs are 5x quicker: agents prepare every pay slip, and a 15-to-20-slip run becomes one bulk transfer after a single review.
• Some customers delegate transfers as large as 100,000 euros. Qonto's AI Lab lead says the team expected small amounts first.
• One customer generated 500-plus monthly invoices in a single batch.

The baseline is also Qonto's. It cites a Forrester analysis for Qonto's customers putting financial admin at up to 8 hours a month. The page does not say how the 2x, 3x and 5x figures were measured, over how many customers, or against what sample.

How it is built

Qonto's 15-person AI Lab started the project in November 2025 and released the first agent six weeks later to several thousand beta customers, the page says. An operator agent takes high-frequency tasks and hands each request to a specialised agent scoped to that task; a separate analyst agent answers questions on past and present transactions.

Qonto routes each step to a model on latency, performance and cost. According to the page, complex missions run on Claude Opus 4.5 and Sonnet, simpler steps go to smaller models including Haiku, and access runs through Amazon Bedrock, with GDPR and EU AI Act compliance handled at the provider level. Every new release triggers fresh evals so each use case stays on the best model.

The control design

Qonto's stated rule is that the agent can prepare, recommend and give insights, but the user makes the decision. Any action with a financial or customer-relationship impact requires confirmation, and the review screen shows just enough to approve a transfer or judge whether an insight's data is accurate. On the most sensitive use cases, deterministic checks sit alongside the model.

Qonto's AI Lab lead is quoted tying delegation to trust: users who trust the agent hand over more, and that is where the value comes from.

What the page leaves out

• Cost: no spend per task, per customer or on model usage.
• Errors: no rate of wrong drafts, rejected approvals or incidents, and no description of what the deterministic checks block.
• Evidence: no sample size, time window or method for the speed-ups, and no independent check.
• Scope: banking reconciliation is described as next on the roadmap, so it is not part of these results.

What to do

• If an agent in your product can move money or send documents, copy the pattern of prepare, review, confirm, and write down which actions need a person.
• Route by step: reserve your largest model for complex steps and test a smaller one on the rest, as Qonto says it does.
• Re-run your evals on every new model release before switching, and ask any vendor quoting a speed-up for the baseline and sample behind it.