What is function calling in AI Agents? A Guide for product managers in banking

Function calling is the mechanism that lets an AI agent ask a system to do something specific, like look up an account, create a payment, or fetch policy details. In practice, it means the model does not just generate text; it can trigger real software functions with structured inputs and use the results in its response.

How It Works

Think of function calling like a bank branch assistant who cannot move money on their own, but can fill out the exact form needed for the payments team.

The AI agent receives a user request, decides whether it needs external data or an action, then selects a function from a predefined list. It sends structured arguments to that function, such as:

• •customer ID
• •account number
• •payment amount
• •date range

The application executes the function and returns the result to the model. The model then turns that result into a user-friendly answer.

For a product manager, the key point is this: the model is not improvising actions. It is choosing from approved capabilities your team exposed through APIs or internal services.

A simple flow looks like this:

1. •Customer asks: “What’s my card balance?”
2. •The agent recognizes it needs live data.
3. •It calls get_card_balance(customer_id).
4. •Your backend returns £2,480.
5. •The agent replies: “Your current card balance is £2,480.”

That’s different from plain chat because the model can now work with real systems instead of only predicting text.

Why It Matters

• •

  It makes AI useful for real banking tasks

  • •Without function calling, an agent can explain policies.
  • •With it, the same agent can check balances, open tickets, or route requests into operational systems.
• •

  It reduces hallucination risk

  • •For regulated workflows, you do not want the model guessing account status or policy terms.
  • •Function calling forces the agent to retrieve facts from approved sources.
• •

  It creates controlled automation

  • •Product teams can expose only safe actions.
  • •Example: allow “check eligibility” but block “submit transfer” unless extra checks pass.
• •

  It improves auditability

  • •Every function call can be logged with inputs, outputs, and timestamps.
  • •That matters for compliance reviews, dispute handling, and incident investigation.
• •

  It helps separate UX from business logic

  • •The AI handles conversation.
  • •Your services handle rules, permissions, and side effects.
  • •That separation is cleaner than stuffing business logic into prompts.

Real Example

Imagine a mortgage servicing assistant in a retail bank.

A customer says: “Can I get my outstanding mortgage balance and next payment date?”

The agent should not invent those numbers. It should call two functions:

{
  "name": "get_mortgage_summary",
  "arguments": {
    "customer_id": "C12345",
    "mortgage_account_id": "M99881"
  }
}

Your backend responds:

{
  "outstanding_balance": "£184,220",
  "next_payment_date": "2026-05-01",
  "next_payment_amount": "£1,245"
}

Then the agent replies:

Your outstanding mortgage balance is £184,220. Your next payment of £1,245 is due on 1 May 2026.

Now add a second step. The customer asks: “Can I change my payment date?”

That should trigger a different function:

{
  "name": "check_payment_date_eligibility",
  "arguments": {
    "customer_id": "C12345",
    "mortgage_account_id": "M99881",
    "requested_date": "2026-05-10"
  }
}

If eligible, the agent can proceed to a separate approval workflow. If not, it explains why and offers alternatives.

This is where product decisions matter:

For banking products, this pattern is useful in customer service bots, RM copilots, claims assistants in insurance, KYC workflows, and internal ops tools. The common thread is simple: the model handles language; your systems handle truth and action.

Related Concepts

• •

  Tool use

  • •Broader term for letting an AI agent call external systems.
  • •Function calling is one implementation of tool use.
• •

  Structured outputs

  • •Useful when you need predictable JSON from the model.
  • •Often paired with function calling for validation-heavy workflows.
• •

  Agent orchestration

  • •The logic that decides when to think, call tools, ask follow-up questions, or stop.
  • •Important once you have multiple functions and branching flows.
• •

  RAG (Retrieval-Augmented Generation)

  • •Used when the model needs documents rather than actions.
  • •Good for policy lookup; not enough for executing transactions.
• •

  Human-in-the-loop approval

  • •Required for sensitive steps like payments, underwriting decisions, or customer profile changes.
  • •Keeps automation within governance boundaries.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.