KEY TAKEAWAYS
Start here.
- Conversation does not define the architecture.
- Choose the simplest approach that completes the task.
- Test permissions, failures, and human review.
Define the task before the label
A chatbot is a conversational interface. Automation moves work between steps. An AI agent selects and uses permitted tools for a task. These can overlap. Start with a useful job, its completion criteria, exceptions, and responsible people before choosing architecture.
Use a simple decision sequence
Try rules first. Add AI for variable language, retrieval for approved knowledge, and tool selection only when needed.
| Need | Starting point | Boundary |
|---|---|---|
| Move known fields | Rules-based automation | Validate inputs and exceptions |
| Draft or summarize text | AI assistant | Review accuracy and context |
| Answer from internal knowledge | Retrieval-backed assistant | Respect access and freshness |
| Select tools for a bounded task | AI agent | Limit actions and execution |
Example: triage a customer request
A form collects the request, rules validate fields, and AI suggests a category and summary. Approved sources add context. A person reviews before assigning a ticket or sending a response. This workflow does not need unrestricted tool access.
Specify each permitted action separately. Reading records, drafting replies, changing status, and sending messages have different consequences. Permissions must follow the task and user access.
Test failure before expanding
Evaluate realistic cases: ordinary requests, ambiguity, missing information, contradictory sources, and escalation. A successful demonstration alone does not establish readiness for wider operation.
Track operating effort as well as answer quality: review time, usage cost, response time, and repeated issues. Keep an owner for source updates and permission changes. These requirements can affect whether a pilot remains useful after its initial demonstration.
- Task: one job and a completion condition.
- Sources: owners, access rules, and freshness.
- Actions: explicit read, draft, and write permissions.
- Evaluation: examples and acceptance criteria.
- Fallback: unavailable sources or tools.
- Operations: review, issue handling, and change approval.
Review the next step
Check accuracy, unauthorized access, clarification, safe stopping, and investigation effort. Expand sources, users, or actions only after reviewing evidence. Bring a sample workflow and connected systems to Unanimous to scope an AI pilot with clear operating boundaries.