Choose the workflow before you choose the AI

The first purchase decision is not which AI product to buy. It is which piece of work is worth changing, and what must remain under a person's control. For an agency or MGA, a useful starting point is one bounded task with an identifiable input, a checkable output, and a clear owner.

NIST's voluntary AI Risk Management Framework calls for defining the business context, specific tasks, intended use, and oversight, and for considering viable non-AI alternatives (NIST AI RMF Core). The selection method below is an editorial adaptation of that approach, not an insurance requirement or a validated scoring model.

Write one sentence: “When this input arrives, this person produces this output for this next step.” For example: “When a submission packet arrives, an assistant builds a draft document inventory for an underwriter to review.” That is more testable than “use AI in underwriting.”

List what the tool will not do. In this example, it will not decide whether the account fits appetite, infer missing facts, contact the broker, or alter the policy system. A narrow boundary makes the first pilot easier to explain and its failures easier to locate.

For each candidate, answer five questions:

• Verification: Can a reviewer tell whether the output is correct from an available source?
• Consequence: What happens if an error passes unnoticed?
• Data access: Can the task be tested using authorized information in an approved environment?
• Repeatability: Does the task recur often enough to observe different cases?
• Fallback: Can the team complete the work manually if the tool fails?

Do not add these answers into a reassuring total that hides a serious weakness. A task with easy verification but unauthorized data is not ready. A task with potential time savings but no safe fallback needs more design before a pilot.

An internal document inventory, a first-pass comparison for review, or a draft follow-up list can be useful candidates. These are proposed examples, not claims that a particular product can perform them reliably. Evaluate them against the documents and constraints your team actually faces.

Treat sending messages, changing records, or making consequential decisions as a separate approval question. OWASP recommends limiting an AI system's functionality, permissions, and autonomy, including human approval for high-impact actions (OWASP Excessive Agency).

Name an owner, define a manual fallback, and decide what observation would stop the pilot. Your first useful result may be that the workflow needs cleaner inputs or a simpler non-AI tool. That is a valid outcome, not a failed innovation effort.

Next: Read how to design a pilot and use the pilot worksheet. The objective is a controlled decision, not an impressive demo.