Finom puts live bank data inside Claude and ChatGPT, but payments stay locked

What the connector actually does

Finom already had an AI assistant. It shipped in July 2026 inside the web app and the iOS and Android clients, according to the same announcement. The new layer moves the conversation out of the bank's interface and into tools the customer probably has open all day anyway. That shift is the product: nobody wants to log into a second application to ask what the balance looked like on Tuesday.

The scope, per Finom's announcement, covers transactions, wallets, invoices, cards, accounting details and a daily briefing. The assistant can read all of it. It cannot move money, issue a card or change anything in the account.

The connector is built on the Model Context Protocol, the open standard Anthropic published and that most assistant vendors now read. For a bank, that choice matters more than it looks. An MCP server is one endpoint, not two custom integrations written twice and maintained twice.

Outside Finom's own release the trail is thinner than a launch post usually leaves it: Zapier and viaSocket list third-party MCP connectors that wrap Finom, not Finom's official endpoint, and an older Finom feature request asking for an official MCP server is the clearest public sign of the demand behind it. The launch therefore rests on Finom's announcement and those third-party listings rather than on an independent review of the server itself.

Read-only, scoped, revocable

Access runs through the customer's own secure Finom login and is limited to the user role already assigned in the account. An employee with view rights on invoices does not get the year's payroll queries by connecting their ChatGPT account. Access can be withdrawn at any time from Finom's side.

Read-only is the sensible first release for a finance product. A write-capable connector would put a payment tool within reach of a language model, and the failure mode there is not a wrong sentence. Invoice text and inbound email are attacker-controlled input in most SME workflows. With no payment function exposed, an injected instruction has nothing to call.

The token bill for asking your bank a question

Finom charges nothing extra: the MCP connector sits inside every subscription tier. The assistant side is where the cost lives, and it is smaller than most people expect.

Take a realistic daily briefing request – 4,500 input tokens plus 400 output tokens for the written summary. The input figure assumes the narrow case: a system prompt, one or two tool schemas and a compact daily-summary payload, not a dump of every transaction on the account. A connector that returned a month of statement lines would push the input far higher, so the number is best read as the floor of a tightly scoped briefing.

Prices for the two models sit at opposite ends of the market, and for GLM-5.3 Flash they also differ between resellers. We use OpenAI's published $2.00 / $10.00 per million tokens for its mid-tier GPT-6.1 Sol. The $0.04 / $0.15 per million figure for glm-5.3-flash is one reseller list price, carried by LLM7; other published rates range from about $0.03 input / $0.085 output (OpenRouter) up to $0.075 / $0.25 (z.ai launch promo). The table uses the LLM7 listing so the arithmetic stays tied to a named source.

That is about 1.3 cents per query on OpenAI's mid-tier model against about a fortieth of a cent on Zhipu's budget model at the LLM7 rate, a gap of roughly 54 times. At the OpenRouter rates the same briefing costs about 0.017 cents and the multiple rises to roughly 77 times; at the z.ai launch promo it costs about 0.044 cents and the multiple falls to roughly 30 times. Push the briefing to 9,000 input tokens and the GPT row climbs to about 2.2 cents a day while the GLM row reaches about 0.042 cents at the LLM7 rate; the multiple barely moves, at roughly 52 times. In euros, at any recent exchange rate, a month of daily briefings on the expensive option stays under half a euro.

Token price is therefore not the constraint. The constraint is the two minutes a human spends checking whether the summary is right, and that cost is identical on both rows. For readers who want to compare what the same workload costs locally, our AI Arena benchmarks track tokens per second and time to first token on both cloud and self-hosted models.

Cheap tokens come with a policy question attached. GLM-5.3 Flash is served from China. Plenty of European finance teams will not send account context to a Chinese endpoint even at a fiftieth of the price, which pushes them toward EU-hosted options or open-weight models they run themselves.

Where the data ends up

This is the part of the announcement that needs a follow-up question. Finom's release documents the permission model carefully and says nothing about where the assistant side processes the query content, or which data processing agreement covers it. The customer's bank keeps the account data; the prompt, the returned context and the answer all leave Finom's boundary and land with Anthropic or OpenAI.

Under GDPR, the SME remains the controller of what it asks about. Whether Finom or the assistant vendor acts as processor for the query content is exactly the detail a finance team should demand in writing before switching the connector on. The release does not answer it.

The AI Act side is quieter. Regulation (EU) 2024/1689 put Article 50 transparency obligations into full application on 2 August 2026, with enforcement by the EU AI Office and national authorities. Those duties fall on whoever operates the assistant, not on the bank exposing a read tool. Finom's exposure is the GDPR one.

What stays manual

Read-only means no write-back. A user can ask which invoices are overdue and get an accurate list, because the numbers come from a tool call rather than from the model's imagination. What the assistant cannot do is mark the invoice paid, tag a transaction to the right cost centre, or file the VAT code. Reconciliation, the task most SMEs actually resent, remains a human job with a keyboard.

Arithmetic is the other soft spot. When the model summarises figures returned by the MCP server, it is still generating prose, and prose about numbers can drift. Cross-checking a total against the statement takes seconds and is worth the habit.

Availability

The rollout is pan-EU and rolling now, tied to Finom's existing European customer base. No non-EU markets were mentioned. A Claude or ChatGPT account is required on the assistant side; Finom has not published a list of other MCP-compatible clients it supports, even though the protocol itself is client-agnostic.