How to set up custom actions in neobank support chat?

Custom actions let your chatbot perform specific tasks triggered by customer requests, such as checking a balance, freezing a card, or updating personal details. Instead of linking to a help article, the chat actually executes steps on the customer's behalf. For a neobank, this turns support into a self-service channel that resolves issues in seconds.

In Chatref, custom actions are built using natural language instructions paired with your existing support tools. The ai-agent interprets what the customer needs, calls the right action, and replies with the outcome, all grounded in your knowledge-base to ensure accurate information.

• A Chatref account with custom-actions enabled (available on all plans, including the $50 free credit starter).
• Your knowledge-base uploaded with neobank policies, FAQs, and step-by-step resolution procedures.
• Access to the shared inbox for monitoring escalations.
• The APIs or internal tools you want to trigger (e.g., core banking endpoints, card management, identity verification) with proper authentication.

1. Define the action scenario
   In your Chatref dashboard, navigate to the custom-actions tab. Create a new action and give it a clear name, for example “freeze_debit_card”. Describe when it should be triggered: “When a customer asks to freeze or block their debit card immediately.”

2. Write the instruction set
   Provide the AI agent with step-by-step instructions using plain language. Include the exact API call or system command to execute, what information to collect from the user (e.g., card last 4 digits, reason for freeze), and how to present the result. Example:

   Collect last four digits of card number and a brief reason.
   Call FreezeCard API with those parameters.
   On success, reply: “Your card ending in {last4} has been frozen. Reason: {reason}. Ref: {tx_id}”
   On failure, hand over to shared-inbox with full context.
   

3. Connect to your bank’s systems
   Under the action’s configuration, enter the endpoint URL and authentication credentials. Chatref will pass collected parameters securely. Test the connection using the built-in test runner to ensure responses come back correctly.

4. Ground answers with knowledge-base
   Any static policy information (e.g., “What are the fees for card replacement?”) should be answered by your knowledge-base documents, not by custom actions. The ai-agent automatically mixes both sources; when the question requires an action, it triggers it, otherwise it pulls from docs.

5. Set up escalation rules
   Not every task should be automated. For actions like disputing transactions or closing accounts, configure the action to require human confirmation. In those cases, the ai-agent collects the necessary details and then passes the conversation to your support team via the shared-inbox, with all context attached.

Before going live:

• Simulate customer requests in the playground: “I lost my card, block it now,” “What’s my balance,” “Enable international transactions.”
• Verify that the ai-agent correctly identifies the action, collects the right data, executes, and delivers the result.
• Check escalation paths: attempt an action that should require a human and confirm it appears in the shared-inbox with the full chat history and action details.
• Review your knowledge-base entries to make sure standard questions don’t accidentally trigger an action.

• Start with high-volume, low-risk tasks such as balance enquiries, card on/off toggles, and transaction history lookups. This builds trust and offloads repetitive work immediately.
• Always fall back to human support for anything involving financial liability or account changes, using shared-inbox to hand over with context.
• Keep action instructions precise; ambiguous prompts can cause the ai-agent to misinterpret. Test all variations of customer language - neobank users might say “lock card,” “block card,” “stop card,” etc.
• Monitor via shared-inbox to spot where actions fail or need refinement. Chatref’s real-time inbox shows every conversation, allowing you to jump in and take over if the ai-agent struggles.
• Combine actions with knowledge-base articles so users get procedural help and can self-trigger actions. For example, “I want to change my PIN” could first show instructions, then offer a button to start the action.