Other

Images

Overview

Your AI Agent can read images that end users send and use them to resolve inquiries. A vision model produces a text description of each image, and the Agent reasons over that description. The Agent never receives the image itself.

Supported formats and channels

The AI Agent reads images on Email and on the Messaging SDK, which is in early access.

  • Formats: PNG, JPEG, and GIF on both channels. WebP is read on the Messaging SDK only. Other file types, such as HEIC, HEIF, AVIF, SVG, PDFs, documents, and video, are not read.

  • Size: on the Messaging SDK, each image can be up to 50 MB. On Email, the whole message, including every attachment, can be up to 30 MB.

  • Messaging SDK (early access): each image an end user uploads is read. Images sent through the legacy Ada Web Chat are not read.

  • Email: up to 3 images per email are read. The inbox’s attachment handling setting controls the rest:

    • Hand off to a human (default): an email whose only attachments are up to 3 supported images is answered by the Agent instead of being handed off. Emails with more than 3 images, an image in an unsupported format, or any non-image attachment are handed off as usual.
    • AI Agent replies; attachment is kept for handoff: images are read, and every attachment is kept for a later handoff.
    • AI Agent replies; attachment is discarded: images are not read.

    HEIC, HEIF, AVIF, and SVG attachments follow the setting like any other non-image attachment, but they are never read.

What the Agent reads

The AI Agent receives a description of each image. It includes all visible text: order numbers, error messages, codes, dates, and amounts. It also includes the visible content: products, damage, packaging, and the screen shown in a screenshot. Text inside an image is treated as end-user input, never as instructions.

Sensitive images

An image is withheld when the AI Agent detects any of the following anywhere in it. The Agent learns only the document type, not its contents, and the filename is dropped.

  • A payment card with any part of the card number visible
  • A government identity document: passport, driver’s license, national or state ID, residence permit
  • Another credential, such as a Social Security card or a bank statement showing full account numbers

Detection is automatic and best-effort. A card number that detection misses is still masked by the built-in payment card redaction before the Agent sees the description. Any other text that gets through is handled like typed text, including your Redaction settings.

The Agent tells the end user the image could not be processed because it appears to contain sensitive information. It does not ask them to resend it. For a payment card, the Agent may ask for the last 4 digits as text. For an identity document or other credential, the Agent offers a human agent instead of asking for the number. In the dashboard, a withheld image appears as [Image withheld: sensitive content]. A withheld email attachment is not forwarded to human agents on handoff.

Redactions

Redaction settings apply to image descriptions the same way they apply to typed messages.

Retention

Images are deleted 7 days after upload, consistent with other attachments. The text description remains part of the conversation transcript.

Unreadable images

If an image cannot be read, the AI Agent tells the end user and asks them to describe it or send it again.

Have any questions? Contact your Ada team, or email us at .