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

Function calling is a way for an AI agent to request that a specific software function be executed instead of trying to answer everything in free text. In practice, it lets the model say, “call this approved tool with these inputs,” so the system can fetch data, calculate something, or take a controlled action.

How It Works

Think of function calling like a wealth management firm’s approved workflow form.

A relationship manager does not walk into operations and improvise a transfer or trade. They submit a request using a fixed form, with required fields, validation checks, and an approval path. Function calling works the same way: the AI agent identifies the right function, fills in the parameters, and hands that structured request to your system.

Here is the flow in plain English:

• •A user asks the agent something like, “Can I rebalance this client’s portfolio?”
• •The model reads the request and decides it needs a tool, not just a narrative answer.
• •It outputs a structured function call such as get_portfolio_exposure(client_id) or check_restrictions(account_id).
• •Your application executes that function against internal systems.
• •The result comes back to the model, which then explains the outcome in natural language.

The important point for compliance: the model is not inventing actions on its own. It is selecting from functions you have already exposed and approved.

A simple analogy is online banking bill pay. The interface may feel conversational, but under the hood you are choosing from predefined payees, limits, and authorization steps. You are not letting someone type arbitrary instructions into the payment engine. Function calling gives AI agents that same controlled interface.

For compliance teams, this matters because you can separate:

• •Interpretation: what the user wants
• •Execution: what systems are allowed to do
• •Approval: whether policy permits it

That separation is where governance lives.

Why It Matters

Compliance officers in wealth management should care because function calling changes how AI interacts with regulated processes.

• •

  It creates auditability

  • •Every tool call can be logged with inputs, timestamps, and outcomes.
  • •That makes it easier to reconstruct why an agent recommended or triggered an action.
• •

  It reduces hallucinated actions

  • •A plain chatbot might claim it “checked” holdings or “submitted” a request when it did not.
  • •With function calling, the agent must use real system functions or admit it cannot complete the task.
• •

  It supports policy enforcement

  • •You can restrict which functions exist at all.
  • •You can also gate sensitive functions behind approvals, entitlements, or human review.
• •

  It improves consistency

  • •The same rule gets applied every time because the agent calls the same approved function.
  • •That matters for suitability checks, restricted list screening, KYC refreshes, and exception handling.

The practical compliance value is control. Instead of reviewing free-form AI output after the fact, you define what actions are possible before anything happens.

Real Example

Consider a private wealth desk handling a client who asks:

“Can I move $250,000 from my cash sweep into municipal bonds?”

A compliant AI agent should not answer by guessing. It should use function calling to check policy and account state first.

Possible functions:

• •get_client_profile(client_id)
• •check_suitability(client_id, product_id)
• •verify_liquidity_requirements(account_id)
• •check_restricted_products(client_id)
• •create_recommendation_ticket(...)

Here is what that looks like in simplified form:

{
  "function": "check_suitability",
  "arguments": {
    "client_id": "C12345",
    "product_id": "MUNI_BOND_FUND_01"
  }
}

If the suitability check returns:

• •client risk tolerance = moderate
• •time horizon = long-term
• •concentration limits = within threshold
• •no restricted status

then the agent can generate a controlled response:

“Based on current profile data and policy checks, this allocation appears suitable. A human advisor still needs to confirm tax implications and final approval before execution.”

If one check fails — for example, liquidity requirements would be breached — the agent should stop there and escalate:

“This transaction would reduce liquid assets below your firm’s minimum threshold. I’ve created an exception review for advisor approval.”

That is the difference between a risky chatbot and a governed AI workflow.

For compliance officers, this pattern is useful because you can map each function to:

• •an approved business process
• •a control owner
• •logging requirements
• •escalation rules
• •retention policies

You are not approving “AI” in general. You are approving specific callable actions with defined boundaries.

Related Concepts

• •

  Tool use

  • •Broader term for models using external systems such as databases, calculators, CRMs, or policy engines.
  • •Function calling is one implementation of tool use.
• •

  Structured outputs

  • •The model returns data in a fixed schema instead of loose prose.
  • •Useful when downstream systems need reliable fields like account ID, risk score, or approval status.
• •

  Human-in-the-loop approval

  • •A control pattern where AI prepares work but does not execute sensitive actions without human sign-off.
  • •Common for trades, exceptions, complaints handling, and customer communications.
• •

  Policy engines

  • •Systems that evaluate rules like suitability thresholds, jurisdiction constraints, or restricted lists.
  • •Often paired with function calling so policy decisions happen outside the model.
• •

  Audit logging

  • •The record of who asked what, which functions were called, what data was returned, and what action followed.
  • •Essential for supervision reviews and regulatory exams.

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.