What is state machines in AI Agents? A Guide for CTOs in lending

State machines are a way to model an AI agent as a set of explicit states, with rules that control when it can move from one state to another. In AI agents, a state machine defines what the agent is doing right now, what it is allowed to do next, and what event triggers the transition.

How It Works

Think of a loan application workflow like a bank branch queue.

A customer starts at Submitted, moves to KYC Check, then maybe Income Verification, then Risk Review, and finally Approved or Rejected. The customer cannot jump from Submitted straight to Approved unless the required checks have happened and the rules allow it.

That is the core idea of a state machine:

• •You define a finite set of states
• •You define valid transitions between those states
• •You define events or conditions that trigger each transition

For AI agents, this matters because the agent is not just “chatting.” It is making decisions across multiple steps:

• •Ask for missing documents
• •Validate inputs
• •Call external systems
• •Escalate to a human
• •Retry on failure
• •Close the case

Without a state machine, agents tend to become messy. They may repeat actions, skip checks, or behave differently depending on prompt wording. With a state machine, the agent behaves like a controlled workflow engine with intelligence attached.

A simple lending example:

In practice, the AI can still use LLMs for extraction, summarization, and classification. The state machine just keeps the process disciplined.

Why It Matters

CTOs in lending should care because state machines solve problems that show up fast in regulated workflows:

• •

  They reduce agent chaos

  • •Lending flows have hard business rules.
  • •A state machine prevents an agent from “freestyling” its way through underwriting.
• •

  They make compliance easier

  • •You can prove why an action happened.
  • •Every transition can be logged, audited, and tied to a policy.
• •

  They improve reliability

  • •If document extraction fails, the agent can move to a retry or escalation state.
  • •That is better than letting the agent stall or hallucinate next steps.
• •

  They separate policy from intelligence

  • •The LLM handles interpretation.
  • •The state machine handles process control.
  • •That separation is what makes systems maintainable.

For engineering teams, this also means cleaner integration boundaries. Your orchestration layer knows exactly when to call credit bureaus, when to ask for bank statements, and when to stop.

Real Example

Let’s say you are building an AI assistant for SME lending origination.

The borrower uploads:

• •Business registration documents
• •Bank statements
• •Director IDs
• •Tax returns

The AI agent’s job is not just to summarize these files. It must guide the application through a controlled pipeline.

State flow

1. •

   Received

   • •Application lands in the system.
   • •The agent creates a case ID and stores raw documents.
2. •

   Document Check

   • •The LLM extracts document types and checks completeness.
   • •If tax returns are missing, transition to Need More Info.
   • •If complete, move to Verification.
3. •

   Verification

   • •The system calls OCR, identity checks, and bank statement parsers.
   • •If names do not match across documents, move to Manual Review.
   • •If consistent, move to Risk Scoring.
4. •

   Risk Scoring

   • •The AI summarizes financial signals for an underwriting model.
   • •A deterministic rule engine applies policy thresholds:
     • •revenue minimums
     • •debt service coverage ratio
     • •blacklist checks
5. •

   Decision

   • •If all rules pass: Approved
   • •If borderline: Escalated
   • •If failed: Rejected
6. •

   Notification

   • •The agent generates customer-facing messaging based on final state.
   • •No matter how many times the LLM runs internally, only valid outputs are allowed from this point.

Why this works

The LLM is useful inside each step:

• •extracting values from PDFs
• •classifying document quality
• •drafting explanations for underwriters

But the state machine controls the sequence:

• •no skipping verification
• •no approval before risk scoring
• •no repeated calls after finalization

That gives you predictable behavior in a domain where mistakes are expensive.

Related Concepts

• •

  Workflow orchestration

  • •Similar goal: control multi-step processes.
  • •Difference: workflows often focus on task execution; state machines focus on allowed states and transitions.
• •

  Finite State Machines (FSMs)

  • •The classic computer science version.
  • •Useful when your agent behavior can be expressed as discrete states with clear transitions.
• •

  Rule engines

  • •Good for policy decisions like approval thresholds or exception handling.
  • •Often paired with state machines in lending systems.
• •

  Agent memory

  • •Stores context across turns or tasks.
  • •Memory answers “what happened before,” while state machines answer “what phase are we in.”
• •

  Human-in-the-loop review

  • •Critical in lending when confidence is low or policy requires manual intervention.
  • •State machines make escalation paths explicit instead of ad hoc.

If you are building AI agents for lending, start by mapping your process into states before you write prompts. That one design choice usually determines whether your system behaves like production software or like an impressive demo that breaks under real cases.

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.