How to Build a underwriting Agent Using CrewAI in TypeScript for retail banking

An underwriting agent in retail banking takes a loan application, gathers the right facts, checks them against policy, and produces a decision package for a human underwriter or an automated approval flow. It matters because underwriting is where you control credit risk, regulatory exposure, and turnaround time; if you get the workflow wrong, you either lose good customers or approve bad risk.

Architecture

• •

  Application intake layer

  • •Normalizes applicant data from CRM, LOS, PDF statements, and bureau responses.
  • •Validates required fields before any agent work starts.
• •

  Policy retrieval component

  • •Pulls product rules, eligibility thresholds, and exception policies from a controlled source.
  • •Keeps the agent aligned with current underwriting policy.
• •

  CrewAI orchestration

  • •Uses Agent, Task, and Crew to split work into analysis steps.
  • •One agent can assess income stability, another can evaluate risk flags, and a final one can compile the decision memo.
• •

  Tooling layer

  • •Exposes bank-approved tools for bureau lookup, affordability calculation, KYC checks, and document extraction.
  • •Keeps external access auditable and constrained.
• •

  Decision output service

  • •Produces a structured underwriting recommendation: approve, refer, or decline.
  • •Stores rationale for audit and model governance.
• •

  Compliance logging

  • •Captures inputs, tool calls, outputs, timestamps, and policy version.
  • •Supports audit trails, explainability reviews, and dispute handling.

Implementation

1. •

   Install CrewAI and define your banking tools

   In TypeScript projects, keep your tools explicit. For retail banking, each tool should map to an approved internal service or a read-only external dependency.

   import { Tool } from "@crewai/core";
   
   export const calculateDTITool = new Tool({
     name: "calculate_dti",
     description: "Calculate debt-to-income ratio from verified income and obligations",
     execute: async ({ monthlyIncome, monthlyDebtPayments }: { monthlyIncome: number; monthlyDebtPayments: number }) => {
       if (monthlyIncome <= 0) throw new Error("monthlyIncome must be > 0");
       return {
         dti: monthlyDebtPayments / monthlyIncome,
       };
     },
   });
   
   export const kycCheckTool = new Tool({
     name: "kyc_check",
     description: "Check KYC status using internal compliance service",
     execute: async ({ customerId }: { customerId: string }) => {
       // Replace with internal API call
       return {
         customerId,
         kycStatus: "passed",
         sanctionsHit: false,
         pepHit: false,
       };
     },
   });
   
   export const bureauFetchTool = new Tool({
     name: "fetch_credit_bureau",
     description: "Fetch credit bureau summary for underwriting",
     execute: async ({ ssnLast4 }: { ssnLast4: string }) => {
       return {
         score: 721,
         delinquencies12M: 0,
         utilization: 0.28,
         inquiries6M: 2,
       };
     },
   });
   

2. •

   Create agents for analysis and decisioning

   Use separate agents for separation of concerns. That gives you better traceability than one giant prompt that does everything.

   import { Agent } from "@crewai/core";
   import { calculateDTITool, kycCheckTool, bureauFetchTool } from "./tools";
   
   export const riskAnalyst = new Agent({
     role: "Retail Banking Risk Analyst",
     goal: "Assess applicant risk using verified financial data and bank policy",
     backstory:
       "You are a conservative retail lending analyst who prioritizes policy adherence, affordability, and fraud signals.",
     tools: [calculateDTITool, bureauFetchTool],
     verbose: true,
   });
   
   export const complianceReviewer = new Agent({
     role: "Bank Compliance Reviewer",
     goal: "Validate KYC/AML status and identify policy exceptions",
     backstory:
       "You ensure every recommendation is compliant with retail banking rules and auditable.",
     tools: [kycCheckTool],
     verbose: true,
   });
   
   export const decisionWriter = new Agent({
     role: "Underwriting Decision Writer",
     goal: "Summarize findings into an underwriting recommendation with rationale",
     backstory:
       "You produce concise decision memos for human underwriters and audit teams.",
     verbose: true,
   });
   

3. •

   Define tasks with structured outputs

   Keep task outputs structured enough to store in your loan origination system. For banking workflows, free-form text alone is not enough.

   import { Task } from "@crewai/core";
   import { riskAnalyst, complianceReviewer, decisionWriter } from "./agents";
   
   export const assessRiskTask = new Task({
     description:
       "Analyze applicant affordability and credit quality using verified income of $8500/month and debt obligations of $2450/month. Return DTI analysis plus key risk flags.",
     expectedOutput:
       "A JSON-like summary containing dti, credit observations, and recommendation inputs.",
     agent: riskAnalyst,
   });
   
   export const complianceTask = new Task({
     description:
       "Verify KYC status for customerId CUST-10492. Identify any sanctions or PEP concerns.",
     expectedOutput:
       "A JSON-like summary containing kycStatus, sanctionsHit, pepHit, and compliance notes.",
     agent: complianceReviewer,
   });
   
   export const memoTask = new Task({
     description:
       "Write the final underwriting memo using prior findings. Recommend approve, refer, or decline with reasons.",
     expectedOutput:
       "A concise underwriting memo suitable for audit review.",
     agent: decisionWriter,
     context: [assessRiskTask as any, complianceTask as any],
   });
   

4. •

   Run the crew and persist the result

   The core pattern is Crew -> kickoff() -> persist output. In production you should wrap this behind an API endpoint or queue consumer.

   import { Crew } from "@crewai/core";
   import { assessRiskTask, complianceTask, memoTask } from "./tasks";
   
   async function main() {
     const crew = new Crew({
       agents: [],
       tasks: [assessRiskTask, complianceTask, memoTask],
       verbose: true,
       processType: "sequential",
     });
   
      const result = await crew.kickoff();
   
      console.log("Underwriting result:", result);
      // Persist result + policy version + request metadata to your audit store here
    }
   
    main().catch((err) => {
      console.error(err);
      process.exit(1);
    });
   

Production Considerations

• •

  Data residency

  • •Keep applicant PII inside your approved region.
  • •If CrewAI calls external services or LLMs are involved downstream, ensure model hosting stays within your regulatory boundary.
• •

  Auditability

  • •Log every task input/output pair with timestamps and policy version.
  • •Store tool invocation metadata so reviewers can reconstruct how a decision was reached.
• •

  Guardrails

  • •Hard-block unsupported decisions like “approve despite failed KYC.”
  • •Add deterministic rules outside the agent for minimum score thresholds, max DTI limits, blacklist checks, and exception routing.
• •

  Monitoring

  • •Track approval rate drift by product segment.
  • •Alert on abnormal referral spikes because they usually indicate policy mismatch or upstream data quality issues.

Common Pitfalls

• •

  Letting the agent make final credit decisions without hard rules

  • •Avoid this by enforcing deterministic policy checks before any recommendation is accepted.
  • •The agent should recommend; your rules engine should decide whether that recommendation is admissible.
• •

  Passing raw sensitive data into prompts unnecessarily

  • •Avoid sending full account numbers or unmasked identifiers.
  • •Use tokenized customer IDs and only the minimum attributes required for underwriting.
• •

  Skipping versioned policy context

  • •If you do not stamp each run with a policy version, audit becomes messy fast.
  • •Store the exact eligibility rules used for every request so disputes can be traced later.

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.