Start with the required action
A search interface can help people find an answer. A workflow can move agreed fields between systems. An agent can choose among available actions based on context. These categories overlap, and the label does not establish reliability or commercial value.
Do not add reasoning where a rule is enough
If the task is to copy approved CRM fields to a project tool, explicit mappings and validation may be easier to test than a model deciding what to do. Consider AI for a genuine requirement, such as extracting information from unstructured text, with an evaluation that can fail.
Treat tool calls as requests to validate
A model proposing an API call is not authorization to execute it. Check input, the actor's permissions, and the intended record on the server. Approval steps, duplicate handling, and a visible error path still matter.
Retrieval from company documents does not eliminate unsupported answers. Design a way to decline or ask for review instead of treating confidence as evidence.
Demonstrate a workflow, not a conversion promise
Use synthetic examples to test successful requests, missing information, unauthorized access, and failed connections. Conversion and cost outcomes require measurement in the actual operation; a conversational interface alone does not prove them.
