Guides
Give an AI Agent a bounded job
Choose the questions the agent is expected to answer and the approved knowledge it should use. Keep sensitive decisions and unusual cases on a human path. Test a known question, an ambiguous question and a question the agent cannot answer.
Engagivo editorial team ·
Review source material for accuracy and relevance.
Agree when the agent should stop and involve a teammate.
Review the conversation and revise the guidance before expanding the scope.
Prerequisites and screens
Prepare an enabled AI provider, approved source material, permission to configure agents and an escalation owner. Open AI Agents (/ai) to choose the agent’s goal, tone and knowledge; connected-account reply settings control whether automated replies are enabled.
Define the job and limits
Use a narrow question set, such as explaining opening hours from an approved source. Keep payments, bookings and other business actions outside the reply’s claimed outcome unless an authorized action has real evidence.
Safe example and expected result
Use fictional messages in Test Agent: a known question, an unknown question and ‘Please let me speak to a person’. Review sources and handoff results. Expected result: a source-supported answer for the known question and an explicit fallback or human path for the others.
Troubleshooting and stopping point
If knowledge is missing or a provider fails, check the source and provider status. Keep account replies in team handling until the test passes. An administrator must separately authorize live reply enablement; testing alone does not enable it.