Skip to main content
RoutMe

Buyer guide

AI chatbot for banking and fintech: what to allow

By the RoutMe team. Last updated .

In banking and fintech an AI chatbot should be treated as an information and routing layer, not an advisor or an actor. It can explain products and fees from approved documents, look up non-sensitive status after verification, and send anything consequential to a person. It should not give financial advice, move money or change account details, and your compliance team should approve its scope before launch.

What is reasonable to automate

Financial services teams answer many repetitive questions that carry little risk when answered from approved text: opening hours, how to reset an app login, which documents are needed to open an account, what a published fee schedule says, how a card replacement works, where to find a statement. These suit an assistant that quotes your own documents. The banking and finance industry page frames what RoutMe covers and where it stops.

What to keep out of scope

Decide in writing, before building, that the assistant will not:

  • Recommend a product, a plan or an investment.
  • Comment on markets, returns or tax treatment.
  • Interpret regulation or a customer's legal position.
  • Initiate or approve payments, transfers, limit changes or closures.
  • Collect full card numbers, passwords or one-time codes in chat.

Each item should appear in the assistant's instructions and in your test list. The structure for instructions is in what to put in an AI agent's instructions.

Verification before account talk

General product questions need no identity check. Anything about a specific customer does. Use whatever verification your institution already approves, through a channel and method that does not require the customer to type secrets into an open chat. Often the safest pattern is for the assistant to say it can explain how to find something and then point to the authenticated app or portal, rather than reading the data itself. If a limited status lookup is warranted, such as whether a card replacement has shipped, return only that field.

Read-only as a design rule

A read-only agent can only read. That sharply limits what a manipulative message can achieve, since there is no tool that moves money. The reasoning is in what read-only AI means and prompt injection and read-only tools. In a regulated setting this is more than convenience: it is easier to explain to a risk committee, an auditor and a regulator than an agent with broad write access.

Auditability

Keep records of conversations and tool calls for as long as your own policies require. When a customer or a supervisor asks what the assistant said, you should be able to answer from logs, not from recollection. Review a sample regularly, and review every conversation the assistant flagged as uncertain. How to measure AI support quality suggests measures that fit this kind of review.

Escalation triggers

Write rules that hand off immediately, with context, when a customer:

  1. Mentions fraud, a scam, an unrecognised transaction or a compromised account.
  2. Disputes a charge or files a complaint.
  3. Shows signs of distress, vulnerability or confusion about a major decision.
  4. Asks for advice or a recommendation.
  5. Asks for something the assistant cannot verify.

Fraud conversations should also tell the customer how to contact your official security line, and the assistant should never ask for sensitive credentials. Handoff mechanics are in human handoff best practices.

Channel and data considerations

WhatsApp is widely used for customer service in many markets, and it has opt-in, window and template rules described on the WhatsApp compliance page. A bank may decide that certain topics never go over a consumer messaging app, and the assistant should redirect those to a secure channel. Check data residency, retention and processor terms with your own counsel, and read the security page.

How RoutMe approaches this

RoutMe agents are read-only by default and connect to systems through MCP or APIs, so a chat session has no path to change an account unless you build one. Guardrails keep the assistant inside the topics you approve, handoff passes the conversation to staff. RoutMe does not make compliance claims on your behalf: whether a deployment satisfies your regulator is for you and your advisers to assess.

Sources

Explore the product: RoutMe for Shopify, solutions and the documentation.

Frequently asked questions

Can the chatbot give financial advice?

It should not. Configure it to decline advice and recommendations and route those requests to a qualified person.

Can it check a balance?

That is a design decision for your security and compliance teams. A common safer pattern is to direct customers to the authenticated app, and allow limited status lookups only where verification is strong.

Is it compliant with financial regulation?

No tool can promise that. Compliance depends on how you scope, configure, monitor and document it, and on your jurisdiction. Involve your compliance team and legal counsel.

What about prompt injection?

Assume customers and documents may try to steer the model. Read-only tools, narrow scope and logged calls limit the damage, as explained in the prompt injection guide linked above.

Put these ideas to work