Start with the question behind the attachment
A customer sends a photo of a bag and asks, “Do you have this in black?” Another sends a voice note asking for delivery to a particular district. The useful outcome is a clear answer based on the attachment and your business information. A detailed description of the photo, by itself, does not answer whether you sell that item.
In AnswerAct, the route depends on the capabilities of the model selected for the agent. Understanding that route helps you choose a model, write better instructions and test the complete customer experience. It also makes failures easier to locate: receiving the attachment, understanding it and answering the business question are separate checkpoints.
When the selected model supports the attachment
If the selected model supports image input, an eligible customer image goes directly to that model. The equivalent applies to supported audio input. The agent continues with its business instructions and conversation context; it does not receive a separate instruction to become a media-analysis assistant.
Check the capabilities shown for the actual model you select. A provider name or a model family name does not establish support for every member of that family. Format, size and provider requirements still apply. A model marked as capable is not a promise that every file a customer can upload will be accepted.
What fallback analysis contributes
When the agent’s selected model lacks the relevant capability, AnswerAct can use the platform’s configured fallback analysis. An image produces an understanding of its visible content; audio produces a transcript. That derived information is then supplied to the agent with its origin identified.
The agent remains responsible for the final customer-facing answer. It uses the analysis alongside the conversation, business instructions and permitted tools. Text inside an image or a recording is customer-provided content, not permission to change the agent’s role, reveal information or claim an order succeeded.
For example, a transcript containing “I want two black bags” is evidence of what the customer requested. It is not evidence that an order has been recorded or stock reserved.
Work through one photo request
Use a test photo whose contents you know. For a bag shop, send a picture of a blue shoulder bag with the caption “Is there a black version of this?” Then check three things.
1. Recognition: does the reply identify the relevant item without inventing details that are not visible?
2. Business grounding: does it consult the information available to the agent before stating that a black version exists?
3. Next step: if several products match, does it ask for a product name or link instead of guessing?
A useful instruction is: “Use attachments to understand the customer’s request. Confirm product facts from our available business sources. If a match is uncertain, ask one short question.” Test a second photo with an unclear label. The response should express that uncertainty naturally.
Test voice notes with details that matter
Record a short test message with a product, quantity and delivery area. Listen to it yourself, then compare the agent’s answer with what was actually said. Include a realistic ambiguous detail, such as two similarly named districts. The correct behavior is to clarify before recording an order or promising delivery.
Media recognition can be imperfect, especially with noise, overlapping voices, small text or blurred images. If the original attachment is unavailable to the channel, ask the customer to resend it normally or describe it. Do not instruct the agent to pretend it saw missing content.
AI processing uses the customer workspace’s credit balance. A fallback route can include an analysis request and an agent-response request, so review usage rather than assuming one attachment always means one model call or a fixed cost.
Questions about model choice
Does fallback run for every photo or voice note?
No. Routing first considers the selected agent model’s relevant capability. Fallback is for a missing capability, subject to the configured processing path and input requirements.
Should I copy the analysis into the customer reply?
The agent should answer the customer’s business question using the available understanding and verified information. A transcript or description is context; it may need clarification or a permitted tool action.
Review model, data and testing together
The AnswerAct operating guide connects model selection with instructions, tools, page assignment and usage checks. Keep a small set of known photo and audio examples to rerun whenever you change the model or the agent’s job.