Teams answer the same questions every day: opening hours, directions, preparation steps, check-in details or what happens next. AI is useful when it retrieves approved information and prepares a consistent answer. It is risky when missing context is presented as fact.
MirageCHAT supports account-level AI modes where the feature is enabled and operational. Suggestion mode prepares a draft for an agent to review. Automatic mode is a separate opt-in. Off means off for that account. Availability still depends on the configured module, source information and service health.
Start with a controlled source
Add current business information to the approved knowledge source. Keep ownership and review dates visible for facts that change, especially prices, availability, opening times and policies. Do not copy private runbooks, credentials, customer records or unsupported provider statements into the public or AI source.
Choose the safest mode for the task
Begin with suggestion mode for questions where an incorrect answer would affect money, access, health, travel or a commitment. The agent reads the original message and the draft together, edits anything that is incomplete, and sends only the final reviewed answer. Use automatic replies only for a narrow, verified question set with an explicit recovery path.
Hand off uncertainty
When the approved source does not contain the answer, the workflow should say that a person needs to help. It must not invent a price, availability, appointment, reservation or provider approval. The conversation remains in the team inbox so an assigned agent can continue with the complete thread.
Measure quality without collecting message content for marketing
Review sampled outcomes inside the controlled product workflow: whether the source supported the answer, whether an agent changed the draft, and whether escalation happened when expected. Organic acquisition analytics must not receive message text, phone numbers, email addresses, tenant IDs, user IDs or client records.
Limits
No AI system should be described as hallucination-free, 100% accurate or a replacement for agents. If service health or quota is unavailable, the public copy and UI must not imply an answer was successfully generated. Human takeover, source maintenance and permission boundaries remain part of the operating model.
Last reviewed · checked against MirageCHAT release 891edef · MirageCHAT product team. Availability can depend on plan, provider approval and the permissions of a connected account; the status shown inside the workspace is authoritative.