An agent cannot repair an undefined process. It can only hide the uncertainty for a while.
Process readiness
Can the team explain the workflow, inputs, permitted actions and successful output? If experienced staff handle every case differently, document the decision before asking software to reproduce it.
Data readiness
Identify the systems of record, missing fields, duplicated records and information the agent is allowed to access. Useful context must also be current, attributable and retrievable.
Action readiness
List every tool the agent may use and the permissions each action requires. Reading a CRM note and issuing a refund are different risk classes. Use least privilege and explicit approval gates.
Evaluation readiness
Collect realistic test cases, including ambiguous inputs and known failures. Define acceptable behaviour before choosing a model. Without an evaluation set, improvement is guesswork.
Operational readiness
Decide who receives alerts, reviews traces, handles exceptions and can disable the workflow. Agents are production systems; ownership after launch matters as much as the demonstration.
If these five areas are clear, an agent build may be sensible. If they are not, the readiness work is the project.