What is function calling in AI Agents? A Guide for compliance officers in lending

Function calling is a way for an AI agent to ask software systems to do specific tasks, like checking a customer’s loan status, pulling policy data, or calculating an affordability ratio. In practice, it means the model does not guess the answer itself; it selects a structured action and passes the right inputs to a trusted system.

How It Works

Think of function calling like a loan officer using a checklist and handing a form to the right department.

The AI agent is the front desk. It understands the user’s request in plain language, then decides whether it needs help from another system. If it does, it calls a function with structured fields instead of free text.

Example flow:

• •A borrower asks: “Can I qualify for a personal loan with my current income and debts?”
• •The AI agent interprets the request.
• •It sees that an affordability calculation is needed.
• •It calls a function like calculate_debt_to_income() with inputs such as income, monthly debt payments, and requested loan amount.
• •The system returns a result.
• •The AI agent explains the result in plain English.

That is the key difference from a normal chatbot. A chatbot may generate an answer from its training data. A function-calling agent can trigger real business logic in approved systems.

For compliance teams, this matters because the model is not making policy decisions on its own. It is routing work to controlled functions that can be logged, reviewed, tested, and restricted.

Here is a simple example of what the model might produce:

{
  "function": "calculate_affordability",
  "arguments": {
    "gross_monthly_income": 6500,
    "monthly_debt_payments": 1800,
    "requested_loan_payment": 420
  }
}

The backend system executes that calculation. The model only handles interpretation and response formatting.

Why It Matters

Compliance officers in lending should care because function calling changes how AI interacts with regulated workflows.

• •It reduces hallucination risk
  • •The model can be forced to call approved systems instead of inventing figures like APRs, income thresholds, or underwriting outcomes.
• •It supports auditability
  • •Every function call can be logged with inputs, outputs, timestamps, and decision paths. That makes reviews and investigations much easier.
• •It helps enforce policy boundaries
  • •You can restrict which functions are available for different users or use cases. For example, an agent may retrieve policy guidance but not approve credit.
• •It improves consistency
  • •If every affordability check uses the same calculation service, you reduce drift between teams, channels, and models.
• •It creates clearer human oversight
  • •Compliance can review which actions were taken by the agent versus which were executed by deterministic business rules.

The practical point is simple: function calling turns an AI assistant into a controlled workflow participant. That is much easier to govern than free-form text generation.

Real Example

A lender deploys an AI agent to support loan officers handling consumer credit applications.

A borrower asks through chat: “What documents are still missing from my mortgage application?”

The agent should not guess. Instead, it calls two functions:

1. •get_application_status(application_id)
2. •get_required_documents(product_type, applicant_profile)

The first function retrieves the current application state from the loan origination system. The second checks policy rules for required documents based on product type and applicant profile.

A simplified flow looks like this:

{
  "function": "get_application_status",
  "arguments": {
    "application_id": "MORT-20481"
  }
}

{
  "function": "get_required_documents",
  "arguments": {
    "product_type": "mortgage",
    "applicant_profile": {
      "employment_type": "salary",
      "first_time_buyer": true
    }
  }
}

The backend returns:

• •Application status: “Under review”
• •Missing documents:
  • •Proof of income
  • •Bank statements for last 3 months
  • •Proof of address

The AI then responds to the borrower in plain language.

Why this is useful for compliance:

• •The response comes from source systems and policy rules, not model memory.
• •The document list can be tied to documented underwriting requirements.
• •The interaction can be logged for complaint handling or file review.
• •If policy changes later, updating the rules engine updates future responses without retraining the model.

This pattern also works in insurance lending-adjacent workflows such as premium financing or collateral verification. The important part is that the AI requests data or actions through approved interfaces rather than improvising.

Related Concepts

• •Tool use
  • •A broader term for allowing models to interact with external systems through APIs or utilities.
• •Workflow orchestration
  • •Coordinating multiple steps across systems such as CRM, LOS, KYC tools, and decision engines.
• •Policy engines
  • •Rules-based systems that determine what actions are allowed under specific conditions.
• •Human-in-the-loop review
  • •Requiring staff approval before high-risk actions like adverse action notices or exceptions.
• •Audit logging
  • •Capturing inputs, outputs, user identity, timestamps, and decision traces for governance and exam readiness.

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.