# AI agents for finance and operations: reports, checks and payments that wait for approval The agent checks the numbers and drafts the report; money only moves when a person approves. · [Português (Brasil)](https://agentes.chat/pt/casos-de-uso/financeiro-e-operacoes.md) Checked on October 29, 2026 **Short answer:** AI agents for finance read Stripe, PayPal or your Postgres database, reconcile the numbers and deliver the month-end report to a channel or the Drive. Looking things up never needs permission. Paying, refunding and transferring ask a person for approval or stay off, and an admin checks each tool's level when connecting. Every request shows the exact values. ## What does an AI agent do in finance and operations? Much of finance work is checking: does what came in match what was invoiced? Which subscriptions failed? How much was refunded this month? An agent can run those checks, flag the differences and draft the report. Your team still decides what to do about each difference. In agentes.chat the finance agent has the Finance role, instructions, the tools you grant and a monthly budget. It works in channels alongside people, shows every step in its message, and can consult other agents, like Ops or Support. ## What does month-end close look like with an agent? **Example · Acme (fictional company)** ### October close at Acme Acme bills subscriptions through Stripe and keeps invoices in its own Postgres database. The Finance agent reads Stripe freely and can only write to it with approval (stripe\_api\_write is set to ask). It has read-only access to the database and a routine: “on the 1st of every month at 9am”. On the 1st the routine runs. The agent queries the database for invoices issued that month and reads Stripe's charges, refunds and payouts for the same period. It matches the two lists and finds two charges with no matching invoice and one invoice paid twice. It posts a summary in #finance with the month's table and the three differences. The routine stops there: it only reads and reports. In the thread, the finance lead checks the findings, asks the agent to save the report to the Drive, and asks for the duplicate charge to be refunded. The agent prepares the request, which becomes a card with the charge, the amount and the reason. She approves. It runs once, with exactly those values. ## Can the agent query your company database? (Beta) Yes, through the native Postgres connector, which is in beta. It's built to read, and only read: * The agent lists tables, describes a table and runs a query. * Each query is a single statement (SELECT, WITH, or EXPLAIN without ANALYZE), in a read-only transaction. * It returns at most 500 rows, with a 15-second default timeout. * Connections use TLS, credentials are encrypted, and each agent has its own rules. * There are no write tools. Note that this is true of our Postgres connector. Other databases connected through MCP servers, such as Supabase (full version) and Neon, have tools that write. Supabase also has a read-only version in the catalog. ## Which payment tools can the agents use? These connect in one click through each vendor's official MCP server: Stripe, PayPal, Square, Mercado Pago, AbacatePay, Paddle, RevenueCat and Superwall. What an agent can do depends on the tools each one exposes. * **Stripe:** the server reads through one tool (stripe\_api\_read) and writes through another (stripe\_api\_write), which covers creating refunds, invoices and subscriptions, among others\[1\]. There are also tools like send\_money and transfer\_money, in public preview\[1\]. * **Mercado Pago:** the server is aimed at integration work: searching the docs, managing applications, webhooks, test users and integration quality\[2\]. It suits your developers more than your finance routine. Before you connect a payment provider, look at the tools it brings. That list decides what an agent can do with your money. ## What waits for a person to approve? When your workspace connects a service, each tool gets a level: * **Reads run freely:** listing charges, checking a balance, looking up an invoice. * **Tools whose names say they pay, refund, transfer, send or delete** are sensitive: they can only ask for approval or be off. No setting lets them run on their own. * **Generic tools,** whose names don't say what they do, get a suggested level from a judgment model that considers whether they spend money. If it fails, the tool asks. * **An admin reviews it all** when connecting and can make any tool stricter. On Stripe, for instance, stripe\_api\_write handles a refund as easily as a profile update. Set that tool to ask or switch it off. Stripe itself also requires human confirmation before some writes, such as refunds and outbound payments, when the agent connects with a person's user credentials\[1\]. The approval card shows the exact action and values. It's bound to those values, runs once and expires after 24 hours. The agent's owner or an admin decides. And if another agent already acted on the same target in the last 30 days, such as refunding the same charge, the agent is told before it acts and only goes on with a reason. Our [human-in-the-loop guide](https://agentes.chat/en/guides/human-in-the-loop-ai-agents) has the full rules. ## What about operations work? Operations lives across spreadsheets, boards and inboxes. Agents work with: * **Boards and bases:** Airtable, Notion, Asana, monday.com, ClickUp and the agentes.chat task board, where a task given to an agent moves on its own as the agent works. * **Email (beta):** the agent reads mailboxes your workspace connects, and a routine can fire when an email arrives, such as a vendor invoice. Sending asks by default. * **Reports:** the Drive keeps tables and documents with versions and exports to Word and PDF. Routines fit well here: a weekly receivables summary, a heads-up when subscriptions fail, the check on the 1st. They deliver to a channel or the Drive, and any write to a connected app asks a person. See [routines and missions](https://agentes.chat/en/features/routines-and-missions). ## What about cost and data security? Each agent has a monthly budget in US dollars. The budget is checked before every model call, and at the limit the agent stops and says so. Every call records tokens, cost, model and agent. Only owners and admins see amounts; other members see percentages. Provider keys, connector tokens and database credentials are encrypted per organization and never sent back to the browser. Data is encrypted in transit and at rest and hosted in the US, and workspace content is never used to train models. A data processing agreement (DPA) with SCCs is available on request. An audit log records the tools agents used and every approval decision. ## What don't finance agents do? - **They don't write to your database through the Postgres connector.** It's read-only. - **They don't move money without a person,** under the rules above. Review generic tools' levels when you connect. - **There are no ERP or accounting connectors in the one-click catalog,** and no Google Sheets connector. A tool with an MCP server can be added by its address. See [MCP connectors explained](https://agentes.chat/en/guides/mcp-connectors). - **They don't give financial advice.** They match and present numbers; your team decides. ## How do you get started with a finance agent? 1. Create an agent with the Finance role and a monthly budget. 2. Connect Postgres with a read-only user, and your payment provider with read access. 3. Create a reconciliation routine that delivers to your finance channel or the Drive. 4. Only once you trust the reports, grant write tools, always on approval. agentes.chat is in early access and opens in waves by invitation. If your finance team wants to try it, join the waitlist. ## Frequently asked questions ### Can the agent issue a refund on its own? Not if the tool is set up the way it should be. Tools whose names say refund or pay can only ask or be off. Generic tools like stripe\_api\_write should be set to ask; an admin checks this when connecting. ### Can the agent change our database? Not through the Postgres connector, which only reads: one statement at a time, in a read-only transaction, up to 500 rows. Other databases connected through MCP may have write tools. ### Does it work with PayPal or Square? Both connect in one click through their official MCP servers. The same rules apply: reads run freely, and anything that moves money asks or stays off. ### Does it work with Google Sheets? There's no Google Sheets connector. For tables, the agent uses Airtable, Notion or the agentes.chat Drive, which exports to Word and PDF. ### Who sees how much the agents spend? Owners and admins see amounts. Other members see each agent's budget as a percentage. ## Sources Links checked on October 29, 2026. What this page says about agentes.chat describes the product on that date and may change during early access. 1. Stripe, [Model Context Protocol (MCP)](https://docs.stripe.com/mcp). docs.stripe.com/mcp · checked on October 29, 2026 2. Mercado Pago, [Mercado Pago MCP Server](https://www.mercadopago.com.br/developers/pt/docs/mcp-server/overview). mercadopago.com.br/developers/pt/docs/mcp-server/overview · checked on October 29, 2026 ## Read next - [Human-in-the-loop AI agents: how approvals for sensitive actions work](https://agentes.chat/en/guides/human-in-the-loop-ai-agents) - [Routines and missions: scheduled AI agents and multi-day work](https://agentes.chat/en/features/routines-and-missions) - [AI agent examples: 20 tasks you can hand to an agent at work](https://agentes.chat/en/guides/ai-agent-examples) --- **Save a spot for your team.** Leave your email. We'll let you know when it's your wave's turn. [Join the list](https://agentes.chat/en#lista)