What gets built
- An agent with defined scope and channels
- Rule-bound qualification and scheduling
- Permission-limited internal actions
- An evaluation set and action log
Guide controls and deliverables
- A permissions matrix and permitted actions
- A contextual hand-off path
- Evaluation samples before changes
Operational signs to review
- The same questions, quotes and follow-up tasks consume much of the team's day
- Requests remain pending when the team is busy or offline
- Opportunities go cold because a request is not logged or followed up in time
How an AI agent works
Channel
The agent connects to an authorized channel, such as WhatsApp, web chat, or email, with a defined hand-off rule.
Knowledge
It answers from authorized company documents and shows the source when one exists. If it cannot find support or a request falls outside scope, it should hand the request to a person.
Tools
It can take actions: check availability, book, create a CRM record or send an email. Every action has permissions and limits.
Evaluation
Before activation, it is tested with a representative set of conversations and expected answers. Later reviews document failures, changes, and hand-offs.
Workflow description
Before
The same questions, quotes and follow-up tasks consume much of the team's day
After
The agent receives a request handling and follow-up request, checks an authorized source, performs a permitted action, or hands it off when it falls outside the rules.
Consulted source, authorized action, hand-off reason, and conversation record.
What the company needs before starting
- The 20 most frequent questions or requests, each with its correct answer
- Access to the information the agent needs: catalog, schedule, policies
- One channel to start with, usually WhatsApp or web chat
- A person who reviews conversations during initial operation
Common mistakes with agents
Letting the agent answer everything
A useful agent has a written scope and says 'a person will review that' when appropriate.
Not connecting it to real data
Without access to the schedule or inventory, the agent only chats; it doesn't resolve.
Launching without evaluation
Testing with representative conversations before publishing helps detect visible errors and refine the rules.
Forgetting the human hand-off
The hand-off to a person must be immediate and carry the full conversation context.
What to review during operation
- request handling and follow-up requests with a response or hand-off
- Hand-off reasons
- Actions blocked by permissions
- Sources without available support
Terms worth knowing
- Agent
- An AI system that, beyond conversing, performs actions with tools within defined rules.
- RAG
- A technique where the agent searches the company's documents and data before answering, so it doesn't make things up.
- Tool
- A concrete action the agent can perform: look up, book, register or send.
- Evaluation
- A set of test conversations with expected answers, used to measure the agent before and after each change.