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

Function calling is a way for an AI agent to ask your system to run a specific action, like checking a loan application, pulling account data, or calculating affordability. It lets the model return structured instructions instead of free-text answers, so the agent can trigger real business workflows safely and predictably.

In lending, that means the AI is not just “chatting” with a customer or analyst. It can decide, based on the conversation and available data, which internal tool or API should be called next.

How It Works

Think of function calling like a loan officer using a checklist and then handing tasks to the right back-office team.

The officer does not manually verify income, pull credit, and calculate debt-to-income ratio themselves. They gather the right information, then route each step to the correct team or system. Function calling works the same way:

• •The user asks something like: “Can I qualify for a $25k personal loan?”
• •The AI agent interprets the request.
• •Instead of answering from memory, it selects a function such as:
  • •get_customer_profile
  • •pull_credit_report
  • •calculate_dti
  • •estimate_loan_offer
• •Your application executes that function.
• •The result comes back to the model.
• •The model turns that result into a customer-facing response or next step.

The key point: the model does not directly access your database or core banking system. It emits a structured request, usually JSON, and your application decides whether to execute it.

Here’s what that looks like in practice:

{
  "name": "calculate_dti",
  "arguments": {
    "monthly_debt_payments": 1450,
    "monthly_income": 5200
  }
}

That structure matters. Product teams get:

• •Predictable behavior
• •Clear audit trails
• •Easier compliance review
• •Better control over what actions an agent is allowed to take

For engineers, function calling is basically controlled tool use. For product managers, it is the difference between an AI that “sounds helpful” and an AI that can actually move a lending workflow forward.

Why It Matters

• •

  It turns conversation into action

  • •A borrower can ask one question in plain English, and the agent can kick off real checks behind the scenes.
• •

  It reduces manual handoffs

  • •Instead of bouncing customers between chat, forms, and support agents, the AI can collect missing details and trigger internal workflows.
• •

  It improves compliance control

  • •You decide which functions exist, what inputs they accept, and when they can run. That makes it easier to enforce policy boundaries.
• •

  It supports better decisioning

  • •The agent can combine conversation context with live systems like credit bureaus, LOS platforms, KYC services, or pricing engines.

For lending products, this is especially useful because many customer journeys are procedural:

• •pre-qualification
• •document collection
• •affordability checks
• •status updates
• •exception handling

Function calling gives you a clean way to automate those steps without letting the model freestyle its way through regulated decisions.

Real Example

Let’s say you run a digital personal loan product. A customer types:

“I earn about $5,000 a month and want to borrow $18,000. Can I qualify?”

A well-designed AI agent should not guess. It should do three things:

1. •

   Ask for missing inputs if needed

   • •monthly debt obligations
   • •employment status
   • •consent to check credit
2. •

   Call internal functions

   • •get_credit_score(customer_id)
   • •calculate_affordability(income, debt, expenses)
   • •check_product_eligibility(loan_amount, term_months)
3. •

   Return a controlled response

   • •“Based on the information provided, you may qualify for loans up to X.”
   • •“We still need consent to run a credit check.”
   • •“Your estimated monthly payment would be Y.”

A simplified flow might look like this:

The important part is governance. The model should never directly approve or decline a loan on its own unless your policy engine explicitly allows it. In most lending setups, function calling should support decisioning—not replace it.

A good production pattern is:

• •AI handles conversation and data collection
• •Rules engine handles policy checks
• •Core systems handle source-of-truth actions
• •Human review handles exceptions

That keeps the experience fast without losing control.

Related Concepts

• •

  Tool use

  • •The broader pattern where an AI model calls external systems like APIs, databases, calculators, or search tools.
• •

  Structured outputs

  • •Constraining model responses into fixed schemas so downstream systems can process them reliably.
• •

  Workflow orchestration

  • •Coordinating multiple steps across services such as KYC, credit pull, underwriting rules, and notifications.
• •

  Guardrails

  • •Policy controls that restrict what the agent can say or do in regulated flows.
• •

  Human-in-the-loop

  • •Escalating edge cases or high-risk decisions to an operations team or underwriter instead of automating everything.

If you are building lending products, function calling is one of the cleanest ways to make AI useful without making it unpredictable. It gives product teams a practical bridge between natural language and real system actions.

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.