Start here: your agent, your business information
AnswerAct lets you build AI agents for your business's Instagram and Facebook Messenger conversations and supported page comments. You supply the information, choose a model and configure the actions the agent may take. The useful result is a customer who gets a correct answer and a clear next step—not merely a message sent automatically.
This handbook follows the journey from account setup to daily operation. It covers the customer-facing features currently available, with separate settings for message agents, comment replies and follow-up bots. Channel access and limits depend on your plan, connected assets and the platform's permissions.
1. Set up the account, plan and working credit
Open Pricing, compare the available plans and check the number of pages and agents, the subscription period and any included AI credit. Use the live price shown for your country and currency. Complete registration and the payment steps displayed for your selected method. A manual payment awaiting review is not an active paid subscription.
After signing in, check Subscriptions and Usage. AI replies, prompt generation, tests, follow-ups and media analysis can consume your account's AI credit. A successful login does not establish that a payment was approved or that an agent is ready to run. Review the current plan and credit before testing a new workflow. For a password reset, use the login page's recovery flow and the code sent to your email; never share that code with another person.
2. Fetch the right pages, then connect them
Open Pages in the dashboard and choose Fetch pages from Meta. In Meta's authorization window, use the account that has access to the business assets you want to connect. Select the relevant Facebook pages and Instagram accounts, finish authorization and return to AnswerAct. Allow the popup for the site if the browser blocked it.
Review the discovered page list and select Connect beside the page you will use. Fetching an asset list and connecting a page are separate steps. Check the resulting status. If a page is missing, first verify which Meta account and asset permissions you selected. When a connected page requests renewed access, complete that renewal before diagnosing the agent's prompt. A prompt change cannot repair an expired page connection.
3. Create an agent and assign its channel
Open Agents and create an agent with a name your team can recognize. Choose a message agent for private conversations or a comment agent for public replies. In Channels, assign the connected page to the appropriate surface. Message and comment agents can serve the same page because their jobs are separate.
Follow the wizard in order: basics, channels, data sources, applicable actions, model and memory, instructions, timing, pause controls, then Test and publish. The available steps differ by agent type. If the target is already assigned to another agent, review the existing assignment rather than publishing a second competing message agent. Begin with a clear task, such as answering delivery questions for one shop page.
4. Add facts in readable data sources
In Data sources, enable the sources the agent needs and add the corresponding records. For a shop, that may include product names, variants, prices and delivery terms. For a service business, start with service descriptions, requirements and opening hours. Save each record and confirm it appears in the source list.
Put changing business facts in their fields and put behavior in the agent instructions. “Delivery to this area costs X” is a fact; “ask for the area before quoting delivery” is an instruction. Test an unavailable item and a missing answer. Tell the agent how to handle an unknown fact instead of asking it to guess. Updating business information is part of operating the agent; it does not require claiming that the underlying AI model was retrained.
5. Configure actions and verify their results
For order recording, choose the fields your team actually needs to fulfil a request. Explain which details are required and when the agent should ask the customer to confirm them. After a test, inspect the saved order as well as the reply. A recorded request, a paid order and a confirmed booking are different outcomes.
Use HTTP call when a workflow needs an external system and you have a working, authorized endpoint. Configure its method, input fields and authentication in the supplied settings, with help from the system's developer when needed. Keep credentials out of the agent prompt and customer messages. In Notifications, configure the available email or Telegram destination and test that the intended team receives the message. Enable only actions you have checked end to end.
6. Choose a model for the conversation you expect
Select an available provider and model in Model and memory. Check its advertised capabilities against your intended text, image, audio and tool tasks, then test representative examples. A model being available in a provider's catalogue does not by itself establish that every capability is usable through your current setup.
Where the selected model supports an incoming media type, the agent can process it directly. Otherwise, the configured media-analysis path supplies context for the agent to use in its answer. Test the result with the text that accompanies an image or voice message: a good description alone is not an answer to “Do you sell this?”. Review AI usage for the whole exchange, including analysis and subsequent replies. Keep memory settings sufficient for useful context and review the cost of unusually long conversations.
7. Write instructions your team can test
Specify the agent's job, allowed sources, tone and next step. An example for a shop is: “Answer the question first. Use our current product and delivery information. Ask one question when a required detail is missing. Do not promise a payment, booking or delivery confirmation unless the relevant action has actually succeeded.” Adapt the example to the actions you configured.
Choose the customer-facing languages and dialects your audience uses. Arabic and English are dashboard languages, not the full list of possible reply languages. Include examples of natural customer wording and mixed-language messages. Test each language with your selected model. If you generate a prompt, read the generated text before saving; generation is an AI operation and can use credit.
8. Give public comments their own rules
Open comment management for the connected page. Choose whether the workflow uses one page-wide reply setup or the available post-specific configuration. The “Enable one reply setup for the whole page” action means a shared page-level configuration; it should not be mistaken for assigning a private message agent.
Keep answers suitable for everyone reading the post. Do not ask customers to publish phone numbers, addresses, payment details or order identifiers in a comment. For personal follow-up, direct them to the appropriate private conversation or your team. Test a genuine comment on the intended page after configuration. Verify any configured private reply separately; a successful public reply does not prove a private message was delivered.
9. Set up a relevant follow-up with page context
Open Follow-up bots, select the Facebook or Instagram pages and choose a fixed message or AI mode. Set the delay in the format shown by the form. The follow-up system works inside its 24-hour window from the customer's last inbound message; choosing a late delay does not extend that window or guarantee delivery.
For AI mode, select each page's active message agent and one readable business data source, such as products or services. This links the correct business information to that page's conversation context; it does not grant the bot access to every source in your account. Choose the model, languages, dialect and prompt, then review the overlap setting before activation. If you change the page's agent or remove a source, revisit that association. Fixed messages do not incur an AI-generation charge; AI follow-ups and prompt generation use customer credit.
10. Test a conversation, an action and the real channel
Use Test and publish for a short sequence: a straightforward product question, a follow-up referring to the previous answer, missing delivery information, a changed quantity and an out-of-scope request. If the workflow records an order or sends a notification, check its destination and fields. Do not judge an action solely by the agent saying it completed it.
Configure Smart wait when customers send a question in several short messages. Set the available pause/resume controls so your team knows when automated replies should stop and how to restart them. Publish after correcting observed failures, then send a new message or comment through the actual connected page. The wizard test and a real-channel test check different parts of the workflow.
11. Keep the account and business information current
Assign responsibility for price, stock, service and policy updates. When one changes, update the corresponding data and repeat a question that uses it. Review recorded requests, notifications and Usage alongside the conversations that produced them. Check subscription status and available credit before adding more traffic or trying a larger model.
Limit team access to the tasks each person needs. Share only the minimum customer information needed to fulfil the request, and use the configured credential fields for integrations rather than pasting secrets into instructions. Review the published Privacy, Acceptable Use, Security and Refunds pages for the current policies. Contact support from the official contact page when an account-specific issue needs investigation.
When something stops: check the right layer
Page missing: check the Meta account and asset access, then fetch the pages again. Page disconnected: renew access and verify connection status. Agent silent: check its active/published state, channel assignment, pause controls, subscription and credit. Incorrect answer: inspect the source record and instructions before changing several settings at once.
Follow-up not sent: check selected pages, active message-agent and data-source associations, timing, overlap settings and the allowed window. Tool action failed: inspect the configured inputs, endpoint permissions and the actual saved or external result. Media answer incomplete: check the selected model's capabilities and Usage, then test with the accompanying question. Ask support to review the platform-managed media-analysis configuration if needed. If the issue remains, send support the displayed error code and the time of the event. Never send passwords, recovery codes or API secrets.