What is model routing in AI Agents? A Guide for compliance officers in lending

Model routing is the process of sending an AI agent’s request to the most appropriate model based on the task, risk level, cost, or policy rules. In lending, it means deciding whether a low-risk customer question goes to a cheaper general model, while a sensitive credit decision or regulated disclosure goes to a stricter, audited model.

How It Works

Think of model routing like a bank’s internal approval chain.

A teller handles routine requests. A branch manager handles exceptions. A credit committee handles high-risk decisions. Model routing does the same thing for AI agents: it inspects the request and chooses the right model before any answer is generated.

In practice, the router looks at signals such as:

• •What the user is asking
• •Whether the request involves regulated content
• •The customer’s risk tier
• •Required accuracy or explainability
• •Cost and latency targets
• •Whether human review is required

A simple example:

• •“What is my loan balance?” → basic model
• •“Explain why this application was declined” → more controlled model with policy checks
• •“Generate adverse action notice language” → highly governed model, often with templates and legal review

For compliance teams, the key point is that routing is not just about performance. It is a control layer. It decides which model is allowed to touch which type of lending activity.

A useful analogy is airport security lanes.

Not every traveler gets the same screening path. TSA PreCheck, standard screening, and secondary inspection exist because different risk levels need different controls. Model routing works the same way: low-risk prompts can move quickly, while sensitive prompts are diverted into tighter oversight.

Why It Matters

Compliance officers in lending should care because model routing affects control design, not just engineering architecture.

• •

  It reduces regulatory exposure

  • •Sensitive activities like credit decisions, adverse action explanations, and complaint handling can be routed to models with stricter guardrails.
  • •That helps limit accidental use of an under-controlled model in a regulated workflow.
• •

  It supports policy enforcement

  • •Routing rules can block certain prompt types from reaching general-purpose models.
  • •This matters when your institution has policies around fair lending, UDAAP, privacy, or use of third-party data.
• •

  It improves auditability

  • •A good routing layer logs why a request went to a specific model.
  • •That gives compliance teams evidence for reviews, audits, and incident investigations.
• •

  It enables risk-based design

  • •Not every interaction needs the same level of control.
  • •Routine FAQ responses can be cheap and fast, while high-impact decisions can require approved models, human review, or deterministic templates.

The main compliance question is simple: if an AI agent makes a mistake, can you show that your routing logic prevented sensitive work from going through the wrong path?

Real Example

A retail bank uses an AI agent in its mortgage prequalification flow.

The agent handles three kinds of requests:

Here’s how it works operationally:

1. •The borrower chats with the agent.
2. •A classifier detects whether the prompt is informational, advisory, or decision-related.
3. •The router checks policy:
   • •Does this involve creditworthiness?
   • •Could it trigger fair lending concerns?
   • •Does it require a legally reviewed disclosure?
4. •The request is sent to the approved path.
5. •The system logs:
   • •Prompt category
   • •Model chosen
   • •Policy rule applied
   • •Any fallback or escalation

If the borrower asks something like “Can you tell me if I’ll qualify before I apply?”, the router may block direct prediction and instead send back a compliant response such as: “I can explain common underwriting factors, but final eligibility depends on your full application.”

That’s model routing doing governance work. It keeps the AI agent inside approved boundaries instead of letting one general-purpose model handle everything by default.

Related Concepts

• •

  Model orchestration

  • •The broader system that coordinates multiple models, tools, and workflows.
  • •Routing is one part of orchestration.
• •

  Prompt classification

  • •The step that labels a request as low-risk, regulated, sensitive, or escalation-worthy.
  • •Often used to drive routing decisions.
• •

  Human-in-the-loop review

  • •A control where certain outputs must be checked by staff before release.
  • •Common in lending disclosures and exception handling.
• •

  Policy engines

  • •Rule systems that decide what an AI agent can do under specific conditions.
  • •Useful for enforcing compliance requirements consistently.
• •

  Audit logging

  • •Records showing what happened, when it happened, and why.
  • •Essential for examinations, incident response, and internal controls.

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.