Let your agent read the files your customers send
Images, voice notes, videos, PDFs, spreadsheets — when a customer drops a file into the chat, your agent can read it and answer questions about its content.
TL;DR
- Open your agent → Brain tab → Multimodal File Understanding.
- Switch it on. Pick how long files stay readable (10 minutes up to 4 weeks).
- Done. From the next inbound file onwards, your agent can answer questions about its content.
OFF by default. Only files received while it's ON are processed — so there's never a surprise bill from a feature you didn't opt into.
1What it does
When a customer sends a file, your agent receives the attachment and the platform extracts a representation it can read:
Documents
PDF, DOCX, RTF, TXT, Markdown
Full text extraction. Tables and paragraphs preserved.
Spreadsheets & structured
CSV, TSV, XLSX, JSON, XML
Row-by-row flattened so the agent can answer questions like 'what's the price of model X?'.
Images
JPG, PNG, GIF, WEBP
AI vision describes what's in the picture and suggests how it might be useful in the conversation.
Audio
MP3, M4A, WAV, OGG
Transcribed end to end. The agent reads the transcript, not the audio.
Video
MP4, MOV, WEBM
Summary plus key scenes. Good enough to answer 'what does the customer want in this video?'
Stickers
WhatsApp / IG decorative
Stored for the inbox view but not analyzed — they carry no useful content.
2Turn it on
Open the agent you want to enable it on. In the editor, navigate to the Brain tab and find the section called Multimodal File Understanding.
Pick the shortest TTL that fits your workflow. Files past their TTL are no longer readable by the agent — they're purged from the AI's context. The customer's original message and attachment remain in your inbox; only the AI-extracted representation expires.
3What happens when a customer sends a file
Customer sends it
WhatsApp, Instagram, or Messenger — the inbound webhook receives the attachment.
Platform reads it
The file is extracted in the background. Usually within 10–30 seconds for documents, longer for big videos.
Agent uses it
On the next reply, the agent has the file's content available and can answer questions about it.
While a file is still being read, the agent sees it as "still processing". If the agent wants to wait for the file before replying, it will. If it would rather acknowledge first and answer in a follow-up, it can — see step 5 below.
4Review what the AI understood
Open any chat (real inbox or Test Chat). On a file message, click the ⌄ chip:
- For a document or spreadsheet: you'll see the extracted text.
- For an image: the AI's description plus suggested uses.
- For audio: the full transcript.
- For video: a summary plus the key scenes.
Spot a mistake? Click Edit.
Fix the text, save, and the agent will use your corrected version from then on. This is the fastest way to recover from a misread invoice number, a transcription error, or a scanned PDF the AI couldn't parse.
5Acknowledge before reading
Bigger files (long videos, hefty PDFs) can take several seconds to extract. With Acknowledge before reading turned on, your agent can send a quick "Got your file, give me a moment to read it." while it waits — so the customer never wonders if you saw their attachment.
The agent decides when to send it — based on the file size and the context of the conversation. It's not forced on every file.
Leave it off if you prefer the agent to just wait silently and reply once with full content.
6Try it in Test Chat
Once the feature is on, the editor's Test Chat panel grows a paperclip button. Click it, pick a file from your computer, and the same upload + extraction flow runs against a private test conversation pinned to that agent.
Real flow, sandbox data
Same bubble, same modal, same AI as production. Ask the agent about the file's contents to see how it would reply to a real customer.
Test files expire fast
Files uploaded through Test Chat live for 1 hour, then they're removed automatically. The agent's configured TTL only applies to inbound customer files.
You need to have at least one inbox connected (WhatsApp / Instagram / Messenger) to use the Test Chat paperclip. We pin the test conversation to one of your existing inboxes for safety reasons — nothing is ever sent outbound.
7Cost note
Only files received while this feature is ON are processed.
Toggling it on doesn't retro-actively re-process past attachments. Toggling it off immediately stops new extractions (existing files still live out their TTL). There's no standby cost — the meter only runs when an inbound message actually carries a file.