How to Use Claude with Qonto for Finance

Qonto knows exactly what came in and what went out of your business. What it does not do is tell you whether you are on plan, what next quarter looks like, or what happens to your runway if you hire two people. That work still happens somewhere else, usually in a spreadsheet you rebuild from a fresh export every month.

Claude can take a lot of that off your plate. You can hand it your Qonto transactions and get a real read on your cash: categorised spend, burn, the months that broke trend. The basic version of this needs no special tooling and no permission from anyone. This article walks the workflow that actually works, shows where it quietly breaks, and covers how to make the output last longer than the session you built it in.

The Simplest Version: Export and Ask

You do not need an integration to start. Qonto lets you export your transactions as CSV. Claude reads CSV well.

In Qonto, export the transaction history for the period you care about.
Open Claude, attach the file, and ask a real question: "Categorise this spend, show me monthly totals by category, and flag anything that broke trend."
Iterate. "Which vendors grew the most quarter over quarter?" "What was my average monthly burn over the last six months?" "Which month was the outlier and why?"

For a one-off look at your cash, this is genuinely enough. Claude is good at this: it reads the file, groups the transactions, and explains what it sees in plain language. If all you wanted was to understand last quarter, you are done.

The limits show up the moment you want something you can rely on, repeatedly.

Where the Naive Workflow Breaks

Three failures appear as soon as you go past a one-off look.

The categories drift. Ask Claude to categorise the same export twice and you can get two slightly different groupings. It is inferring categories from labels each time, with no fixed chart of accounts. Fine for a glance, not fine for a number you put in front of a board.

It forgets. Next month you export again, open a new chat, and Claude has no memory of how you categorised last time, which vendors you split out, or what your budget was. You re-explain the whole setup every session. The analysis is disposable by design.

There is no model underneath. Claude gives you a description of what happened. It does not give you a structure you can forecast on. The moment you want "actuals versus budget" or "reforecast the rest of the year on these numbers," you are back in a spreadsheet, wiring it up by hand, and the AI is no longer in the loop.

None of this means the workflow is wrong. It means the export-and-ask approach is a reading tool, not a modelling tool. Knowing the difference is what separates people who use AI on their finances from people who trust the result.

Making It Repeatable: Give Claude a Persistent Layer

The fix is not a better prompt. It is giving Claude somewhere to put the numbers that persists between sessions, with a fixed structure it does not re-invent each time.

Concretely, that means three things:

A fixed mapping. Your Qonto categories map to the same P&L and cash lines every time, so the numbers are comparable month over month instead of re-inferred.
A structure that holds. Your transactions land in a model with explicit relationships between variables, not a flat pile of cells. Change one assumption and the linked figures follow instead of silently breaking.
Memory across sessions. The model, its conventions, and its history stay put. Next month, Claude picks up where it left off instead of starting from nothing.

This is the layer that turns "Claude read my Qonto export" into "Claude keeps my cash model current." Tools built for this exist. Layerz is one: it connects to Qonto, pulls your transactions into a structured model, and lets Claude drive the whole thing through MCP, so the categorisation, the actuals branch, and the reforecast all persist. But the principle matters more than the tool. Whatever you use, the job is the same: separate the structure from the data so the structure can be reused.

A Workflow That Survives Next Month

Here is what the durable version looks like, tool-agnostic:

Connect once, not export every time. A live connection to Qonto beats a monthly CSV. Fewer steps, no stale files, and the sync is repeatable.
Fix the mapping once. Decide how Qonto categories map to your lines, and keep it. Comparability is the whole point of a monthly number.
Keep actuals and plan separate. Land the Qonto figures as actuals next to your budget on the same structure, so "are we on plan" is a glance, not a rebuild.
Let the agent do the running. With an MCP connection, Claude pulls the month, updates the actuals, and reforecasts the rest of the year from the chat, keeping full context.
Export clean when you need to share. A board wants Excel, not a chat log. The output should drop to a clean, auditable workbook any time.

The difference between step 1 here and the export-and-ask version earlier is the difference between a reading and a system.

One Honest Limitation to Know

If you rely on Qonto's custom category labels, know that Qonto's API only exposes its default categories, not your custom ones. This is a constraint of the Qonto API itself, not of any tool reading from it. In practice you map at the default-category level and refine from there. Worth knowing before you design your categorisation around labels that will not come through a live connection.

The Takeaway

Using Claude with Qonto is real and useful today, and you can start with nothing more than a CSV export and a good question. That gets you a reading of your cash.

Getting a number you trust every month is a different job. It needs a persistent layer under the AI: a fixed mapping, a structure that holds, and memory across sessions. Add that, and Claude stops being a tool you re-explain your finances to and becomes one that keeps your cash model current.

If you want the connected version, here is how to connect Qonto to Layerz. If you would rather stay on exports for now, the workflow above still works. The point is to know which one you are doing.