The AI employee reads documents, and voice messages can be transcribed
How to let the AI employee read files from a customer, which formats and sizes it understands, and how to transcribe a voice message with a chat button or a scenario block.
Documents for the AI employee
A customer sends a price list, a contract or a spreadsheet, and the AI employee answers from the content of the file. AISAR parses the file itself, and only the extracted text goes to the AI provider. The feature is off by default: without it the employee does not open the file and hands the conversation to people.
How to turn it on
- Open the AI employee, the "Schedule and behavior" tab, the "Attachments from the customer" card.
- Turn on the "Read customer documents" switch. It sits between "Voice messages" and "Photos".
- Save the settings.
What is read and within what limits
| What | Limit |
|---|---|
| Formats | pdf, docx, xlsx, csv, txt. The format is detected from the content, not from the file name. |
| Size and count | Up to 10 MB per file, up to 3 files in one message. |
| The first 50 pages. | |
| xlsx | The first 3 sheets and 2000 rows. |
| Text volume | Up to 20,000 characters per message, the rest is cut off. From earlier messages in the conversation up to 4,000 characters per file are used. |
| Not read | Scans and images inside a pdf (no text layer), password-protected and damaged files, doc, ppt and other formats. |
If a file cannot be read, the employee does not invent an answer: it skips the turn and the conversation goes to people through the usual hand-off. If a message has several files and some were read, the employee answers from the read ones and notes that the others could not be read. The employee treats the file content as customer data, not as instructions: commands inside a file are not executed.
Transcribing voice messages
- Voice transcription is part of the Business plan. On other plans there is no button or block.
- The limit is shared for the company by the button and the scenario block: 120 minutes of voice per day. A day is counted in Astana time and resets at 00:00.
- One voice message: up to 5 minutes and up to 6 MB.
- A finished transcript is reused and does not count against the limit a second time, even if it was made by the AI employee, a scenario or a colleague.
- An incoming voice message is recognized. Outgoing voice messages are not transcribed by this button.
The "Transcribe" button in the chat
Under an incoming voice message press "Transcribe". "Transcribing…" appears, then a "Transcript" block with the text and a copy button opens under the player. Everyone who can see the message sees the transcript, and it stays in the conversation. If a transcript already exists, it is shown right away and nothing has to be pressed. If there is no button, transcription is not available on your plan and a hint with a link to plans appears nearby.
| What you see | What to do |
|---|---|
| Today's transcription limit is used up | Wait until 00:00 Astana time or use a transcript made earlier. |
| Too often | Wait a minute: at most 20 presses per minute per employee. |
| The voice message is longer than 5 minutes | It cannot be transcribed, listen to it manually. |
| There is no speech in the voice message | There is nothing to transcribe. |
| The transcription service did not answer, "Could not transcribe" | Press "Retry". If the failure repeats, contact support. |
The "Transcribe voice" block in a scenario
To transcribe voice messages automatically, put the "Transcribe voice" block in a scenario: it puts the text into a variable, for example {{vars.voice_text}}, and the "If contains" condition can check it further. Example: "Inbound message", "Transcribe voice", "If contains the word 'order'". The exits, the error reasons and the limits of the block are described in Automations: basics.
If you work through the API
The transcript is stored in the message, in the message_content.transcript field, and the message_content.text key of such a message is always present (null when the voice message has no caption). The button corresponds to the POST /v1/messages/{id}/transcribe request, details in REST, messages. The message.updated webhook is not sent when a transcription completes: read the transcript through GET /v1/messages.