State Machines vs. Checklists vs. Case Management: How Should an AI Track Work in Progress?

An AI agent can understand a request, call the right tools, and still lose control of the work between initiation and completion. The underlying problem is often not reasoning quality. It is the representation of progress.

Consider a customer asking to cancel a subscription and receive a refund. The agent may need to verify identity, inspect contract terms, calculate eligibility, request approval, execute two transactions, and confirm the result. If the process pauses overnight or encounters an exception, the agent must know more than what was said. It must know what has happened, what remains valid, what is blocked, and what can safely happen next.

There are three practical ways to encode that operational state: a state machine, a checklist, or a case-management record. Each creates a different kind of discipline. Choosing the wrong one produces brittle automation, invisible omissions, or expensive administrative overhead.

The Three Models at a Glance

The distinction is not merely technical. It changes what the agent treats as truth. A state machine trusts the current status and transition rules. A checklist trusts completed obligations. Case management trusts the assembled record and the decisions attached to it.

State Machines: Strong Control for Predictable Work

A state machine represents work as a defined status, such as received, verified, approved, executed, or closed. The agent can move the item only through permitted transitions. A refund cannot move from received directly to executed if verification and approval are mandatory.

This model is effective when the operating path is stable and the transitions have meaningful business consequences. Payments, account provisioning, order fulfillment, and standardized approvals often fit this pattern.

What the mechanism enforces

• Sequence: execution cannot precede authorization.
• Transition conditions: an approval identifier may be required before entering an executable state.
• Ownership: different states can assign responsibility to the agent, an operator, or a specialist.
• Terminal outcomes: completed, declined, canceled, and failed can be distinguished rather than collapsed into “done.”

Suppose an AI handles vendor onboarding. After collecting documentation, it attempts to move the vendor from documents pending to risk review complete. The transition validator checks whether the tax form is current, the bank account has been verified, and the sanctions screening produced a result. The agent cannot talk its way around a missing artifact; the transition fails.

The cost is rigidity. Real work produces suspended requests, reopened decisions, partial execution, disputed evidence, and dependencies on external parties. Modeling every combination as a status creates transition sprawl. Modeling too few combinations hides material differences. A generic blocked state, for example, may conceal whether the agent is waiting for a customer, a regulator, or an internal approval.

Use a state machine when invalid sequencing is a larger risk than process variation.

Checklists: Flexible Coverage Without a Rigid Path

A checklist represents work as obligations rather than statuses. Items may be required, optional, conditional, or not applicable. The sequence can remain flexible: an agent might confirm contact information before or after collecting supporting documents, provided both are complete before submission.

This model fits work where practitioners know what must be covered but do not always perform it in the same order. Quality reviews, launch readiness, incident response, due diligence, and content approvals often have this shape.

What the mechanism enforces

• Completeness: required steps remain visible until resolved.
• Conditional work: selecting “customer is regulated” can add enhanced-review items.
• Explicit exceptions: an item can be marked not applicable with a reason.
• Parallelism: independent tasks can proceed simultaneously.

Imagine an AI preparing a product launch. The checklist includes legal approval, support documentation, analytics instrumentation, rollback instructions, and stakeholder notification. The agent can gather approvals while drafting support material. Before launch, a gate verifies that every mandatory item has acceptable evidence.

The central weakness is semantic shallowness. “Complete” can mean the agent sent an email, received a reply, validated the reply, or merely created a draft. A checked box may record activity rather than outcome. Dependencies can also remain implicit. If rollback instructions are valid only for the final deployment configuration, completing them before that configuration is frozen creates false confidence.

A checklist therefore needs completion criteria. “Obtain security approval” is weak. “Store the security reviewer’s decision, scope, timestamp, and unresolved conditions” is operationally testable.

Use a checklist when omission is the dominant risk and task order legitimately varies.

Case Management: Durable Context for Exceptional Work

Case management treats each unit of work as an evolving record. The record can contain the request, participants, evidence, correspondence, decisions, tasks, deadlines, approvals, and actions. Rather than forcing every situation through one path, it gives the agent a structured workspace in which to manage variation.

This approach is strongest for long-running or judgment-heavy processes: complex customer complaints, insurance claims, employee relations, investigations, contract exceptions, and regulatory inquiries.

Consider a customer disputing an account closure. The original decision involved risk signals, but the customer supplies new documentation. A case record can preserve the original evidence, the policy version used, the rationale, the appeal, the new evidence, the assigned reviewer, and the final determination. A single workflow status cannot express that history. A checklist alone cannot explain why the second decision differs from the first.

What the mechanism preserves

• Provenance: where evidence came from and when it was obtained.
• Decision lineage: which facts and policy versions supported each determination.
• Multiple threads: legal review, customer communication, and operational remediation can proceed together.
• Reopening: a closed matter can resume without erasing its earlier resolution.

The trade-off is governance. If fields are optional, notes are inconsistent, and documents are attached without classification, the case becomes a digital folder that only its original owner understands. AI can summarize it, but summarization does not repair missing provenance or ambiguous decisions.

Effective case management still requires structure: typed evidence, named decisions, assigned owners, due dates, and explicit closure criteria. Flexibility should absorb legitimate variation, not excuse poor records.

Use case management when reconstructing why something happened matters as much as knowing its current status.

Worked Example: Handling a Contract Exception

Assume a sales team requests nonstandard payment terms for a strategic customer.

With a state machine, the request moves through submitted, finance review, legal review, approved, and incorporated into contract. This creates strong control, but trouble begins if finance approves conditionally while legal requests revised language. A single state may not represent both threads accurately.

With a checklist, the agent tracks commercial rationale, credit assessment, finance approval, legal language, and final contract verification. Finance and legal can work in parallel. However, the agent must understand dependencies: changing the payment period after credit approval may invalidate that approval.

With case management, the agent stores the original request, negotiated alternatives, credit evidence, each reviewer’s conditions, contract drafts, and the final signed terms. It handles negotiation well but demands more disciplined record keeping.

A practical design might use a case as the durable container, a checklist for required reviews, and a small state machine for commitment control. The request cannot enter approved until the checklist passes; the case retains the evidence and rationale. This hybrid works because each model has a distinct job, not because all three are layered indiscriminately.

Selection Criteria That Matter

Choose based on the dominant failure

If the worst failure is an agent executing before approval, favor a state machine. If it is forgetting a required review, favor a checklist. If it is losing the reasoning and evidence behind a contested decision, favor case management.

Measure workflow variance

Do not ask whether exceptions exist; every process has exceptions. Ask whether exceptions change the path, the obligations, or the interpretation of evidence. Path variation pressures state machines. Obligation variation pressures static checklists. Evidence and judgment variation points toward cases.

Separate progress from truth

A status such as approved is not the approval itself. A checked item is not proof that its criterion was satisfied. A case note is not verified evidence. Whichever model you choose, store the underlying artifact or machine-verifiable reference where consequential claims are made.

Design for resumption

An agent should be able to resume after a delay without rereading an entire conversation. The work record should identify the current objective, completed actions, unresolved dependencies, latest valid evidence, next permitted action, and any deadline. If the model cannot provide those elements cleanly, it is not adequate for autonomous operation.

Which Model Should You Pick?

• Pick a state machine for high-volume, repeatable workflows where sequence, authorization, and terminal outcomes must be tightly controlled.
• Pick a checklist for work with known obligations but flexible order, especially when omissions are more dangerous than sequencing errors.
• Pick case management for long-running, contested, or evidence-heavy matters where exceptions and decision history are central.
• Pick a deliberate hybrid when a case needs flexible context but contains a small number of irreversible transitions. Use the case as the record, the checklist as the coverage mechanism, and the state machine as the control boundary.

The strongest design is not the one that models the most detail. It is the one that makes the next action safe, the missing work visible, and the eventual outcome reconstructable. For an AI agent, that is the difference between remembering a conversation and managing an operation.

This post was drafted with AI assistance and reviewed against our editorial policy before publication. Corrections are made at the source, on the page, with the date shown.