Start with the job, not the conversation
A chatbot can answer a product question or collect a message. Operational work begins when the system must identify the customer or order, check a source, apply a rule, update another system, and preserve what happened for the next person or step.
The useful design question is therefore not whether the system sounds human. It is whether the job, inputs, allowed actions, stopping conditions, and responsible owner are explicit.
Require state, tools, and authority boundaries
A governed AI role needs a task record that survives beyond one chat message. Tool access should be limited to the systems and actions required for the job, while monetary, policy, and data boundaries determine what may happen automatically.
NIST's AI RMF emphasizes defined human-AI roles, documentation, ongoing governance, and oversight. Those practices turn human review into an operating control instead of an emergency fallback.
- Name the event that starts the job
- List the sources the system may trust
- Separate read access from permission to change data
- Record actions and source references
- Route exceptions to a named role
Choose the smaller system when conversation is enough
Not every business question needs an agentic workflow. If the result is simply an answer, a search experience or well-designed chatbot may be the safer and less expensive choice.
Use an AI Employee pattern only when completing the work creates enough value to justify integration, testing, monitoring, and human accountability.