Expense approvals now live in Claude and ChatGPT: Rydoo MCP connector goes global

For an expense platform, that headline alone is easy to shrug off. “Chat with your finance tool” has been a demo since 2023. What makes this one worth reading is the plumbing underneath it: Model Context Protocol (MCP), the open connector standard Anthropic introduced in November 2024 and which OpenAI, Google and the rest of the assistant ecosystem have since adopted. Rydoo is using it to make the AI chat interface a real — and governed — front end for expense workflows, not a toy. The best use case isn't the one in the demo video. Any employee can already log into Rydoo, open a report and click “approve”. The actual pain point is that finance work happens between tools: a trip ends, the card transactions arrive, the policy question lands in Slack, and an approval gets stuck because the manager lives in a chat client all day. Rydoo's bet is simple: put the action where the manager already is, as FF News reports.

What the connector actually does

Based on the announcement and the toolset Rydoo has been iterating on since July, the practical scope is:

• Balance checks — ask “what do we still owe on the Berlin offsite?” and get a live number.
• Spend summaries — reviewed by trip, project or cost centre in natural language, not by building a custom report.
• Approvals and rejections — managers can approve or reject an expense claim inside Claude or ChatGPT, with the decision written back to Rydoo.
• Card operations — the July 2026 update of the Claude connector added virtual card requests and review of unmatched card transactions.

None of this is unique on paper. Expensify and SAP Concur have all chased conversational finance for years. The difference is architectural: instead of forcing every employee into a proprietary assistant inside one vendor's app, MCP makes the capability portable across the assistants people already use. Rydoo's server was added to Anthropic's Claude connector directory back in July, which tells you this wasn't a rushed weekend project. And for a platform used by over one million people in more than 130 countries, that consistency matters. Notice what is deliberately missing: no claim about a “fully autonomous agent” that pays suppliers by itself. The human still makes the final call on an approval. That is not a limitation — in Europe, it is the feature.

The governance layer is the real product

For an EU-based company, the interesting part is not the natural-language parsing. It's the enterprise governance architecture that Rydoo says IT controls: user authentication, system consent and granular access permissions. This is exactly the difference between a shadow-AI experiment and a finance system. In a consumer setup, you log into ChatGPT and grant it access to a tool in your personal context. In a company, that decision must be made centrally: who in this organisation may act on expenses through an AI client, under which identity, with what limits, and whose approval does the action record as having taken? If the answer lives with the IT department and the finance controller rather than individual employees, the connector becomes audit-ready instead of an audit nightmare. European finance teams should still run their own checks before enabling this. The announcement does not specify where data flows sit between the chat client and Rydoo's backend, which model endpoints are used (enterprise API or consumer subscription), or whether EU data residency applies. Those details matter under the EU AI Act and GDPR when you are moving expense receipts — which contain travel patterns, meal purchases and sometimes invoices — into a third-party AI assistant. The July 2026 “Digital Omnibus on AI” simplification package streamlined compliance procedures and risk-tiering for exactly these integrations, but it did not abolish the questions. It just made answering them less bureaucratic. There is also a less obvious legal nuance. If an AI agent auto-rejected an expense claim without any human review, that could count as automated decision-making with legal effect under GDPR. Rydoo's current design keeps the human as the decision-maker inside the chat — which is the only version of this feature a European data protection officer should accept.

What about the actual cost?

Rydoo did not publish pricing for the connector or any quantitative ROI figure. Our own rough math suggests the per-interaction AI cost is not the concern — for context, current Claude-family pricing sits around $10 per million input tokens and $50 per million output tokens. A typical approval conversation with a modest context of 10,000 tokens therefore costs roughly $0.10 in input, well under €0.10 at current exchange rates before output tokens. Even at high volume, the token bill is the cheapest part of this whole exercise. The real cost sits in integration, identity management, audit logging and the finance team's time spent defining what the AI may and may not do. If a company only has 50 employees using the connector, the AI compute is rounding error. If it has 5,000, the governance work is the budget line that matters — which is presumably why Rydoo put the controls at the centre rather than in a settings menu somewhere.

Where the healthy scepticism belongs

What is missing from the announcement is any measured result: no average time saved per approval, no error rate on natural-language queries, no benchmark versus the existing Rydoo web and mobile flows. Rydoo's own AI Policy Assistant and Smart Audit features, which grew custom branch limits and mobile rollout (app version 7.5.0) back in July, already do part of the compliance checking. How the MCP connector combines with those tools in a real policy violation case — say, a claim that should be routed to a second reviewer — is precisely the kind of edge case that demos gloss over. In our own experience running production AI services, connectors like this look flawless on a clean Slack message and get painful the moment permissions, policy rules and multi-step approvals meet. The MCP standard has matured considerably since November 2024, but the enterprise readiness will not be proven by the launch press release. It will be proven by the first European CFO who can demonstrate, in an audit, exactly which AI client approved which expense, under which identity and with which consent. For Rydoo's existing mid-market and enterprise customers, the takeaway is practical: the feature is live, the governance model is sensible, and the natural next step is a pilot with a small finance team before rolling it out to every manager who lives in ChatGPT.