Choose one conversation you can trace
An active badge is one checkpoint in a working conversation. A customer message still needs to reach the connected page, match the right agent, pass the current operating rules and receive a delivered reply. Rewriting the prompt before checking those steps can hide the original problem.
Begin with a controlled test. Record the page, approximate time, message type and expected answer. Use an ordinary question with a known answer, such as opening hours. Keep the first test free of orders or other actions that change business data. Change one setting at a time and repeat the same question so you can explain what improved.
Check the page and the agent together
Confirm that the intended Facebook Messenger or Instagram page is connected and selected for the agent. A working connection elsewhere in the account does not prove that this page has the assignment you expect. Check both the agent’s status and its channel assignment; a paused assignment can prevent replies even while you are looking at an active agent.
Then check the conversation itself. A closed conversation, a conversation awaiting review or one under human takeover can follow different handling from a normal open conversation. Keep intentional human control in place while investigating. An agent test inside the dashboard is useful, but also send a real controlled message through the connected page to verify the channel path.
Look for an intentional wait or pause
Review Smart Wait before treating a short delay as a failure. Its purpose is to let a customer finish a sequence of messages before the agent replies. Compare your observation with the settings you actually saved.
Also review the agent’s stop and start settings. A configured stop condition can pause a conversation; the configured start command can resume it. Use the values in your account rather than assuming a universal command. If a pause was intentional, identify who should resume the conversation and when. Do not remove these controls merely to make a test produce a reply.
Verify that the response can be generated
Check the current provider and model selection, saved instructions and available AI credits. The subscription must allow the intended use, and the wallet must be available for AI processing. A small positive balance does not necessarily cover the reservation required for a particular model request. Read the displayed error before changing model settings or adding credit.
If the message says the input exceeds the model’s context capacity, separate essential instructions from repeated material and review how much history you include. A larger prompt limit in the editor does not mean every model can accept that entire prompt together with conversation history, tools and an output allowance. Choose a suitable model or reduce the input deliberately.
Separate missing replies from wrong replies
If a reply arrives but gives the wrong price, inspect the business information the agent can read. If text questions work but a voice note fails, test the attachment and model capability. If an answer says an action happened, verify the actual order or tool result. These are different problems with different evidence.
Use this short test set after the first simple question works:
1. A known business fact, to check the source.
2. A follow-up question referring to the previous message, to check context.
3. An unclear request, to check that the agent asks for clarification.
4. A photo or voice note that you can verify yourself, if you need media handling.
5. A controlled tool action, only after its inputs, permissions and expected result are clear.
A useful success record includes the customer message, the response and the confirmed outcome. A fluent sentence alone is insufficient evidence that an external action completed.
Send evidence that narrows the problem
If the issue remains, provide support with the page, time, visible error, message type and the step that failed. Explain whether the dashboard test worked and whether the customer received anything. Remove access tokens, passwords and unnecessary customer details from screenshots. Avoid repeatedly triggering an uncertain write action while waiting for help; check its actual result first.
Common launch questions
Does “active” guarantee a reply to every message?
No. Page assignment, conversation state, stop rules, subscription, credits and channel conditions still matter. Trace one controlled conversation before changing several settings.
Should I repeat a timed-out order or HTTP action?
First check whether the original action took effect. A missing confirmation does not by itself prove that an external system made no change. Use the observed result to decide the next step.
Use the operating guide as your setup reference
The complete AnswerAct guide covers the setup behind these checks. Keep this diagnostic list for launch day and rerun the relevant examples after changes to pages, models, instructions or tools.