# Files in a message

When a customer attaches a file to a message or pastes an image into it, the
assistant reads it and draws on it when writing the draft — for example an
invoice as a PDF, a screenshot of an error or a spreadsheet with an order.

## What the assistant reads

| File | How the assistant reads it |
|---|---|
| JPEG, PNG, GIF and WebP images | looks at them as they are |
| PDF | the whole file, including images and scanned pages |
| an image pasted straight into the message text (a screenshot, say) | the same as an attached image |
| Word, Excel and PowerPoint (.docx, .xlsx, .pptx) | as text; tables are converted to rows, sheets and slides are marked |
| OpenDocument (.odt, .ods, .odp) | the same as Word, Excel and PowerPoint |
| text files .txt, .csv and .html | as text |

The assistant leaves out small images in the message text, such as logos and
icons, and an image that repeats in the message. It recognises the kind of
file by its content, not by the extension in its name.

## What it does not read

- photos in HEIC and AVIF format (some phones save them that way),
- older Word, Excel and PowerPoint files (.doc, .xls, .ppt),
- archives (.zip),
- damaged files and password-protected files,
- a document without text, for example a Word file that contains only images,
- an e-mail attached as a file (.eml, .msg),
- files in any other format.

For such a file the assistant learns its name and the reason it was not read.
In the draft it can then say that the attachment could not be opened.

## How much it reads

- at most 10 files from one message,
- a file up to 5 MB, all the files of a message together up to 10 MB,
- from one document at most 40,000 characters of text, from all the documents
  of a message together 80,000. A longer document is read from the start, and
  the assistant knows that the rest is missing.

## Where you see it

In the assistant's Settings → **Activity**, a message shows the list of its
files. For each one it says whether the assistant received it and, if not,
why.

## When the assistant reads files

- when writing a draft,
- when [revising a draft](/en/revising-a-draft) on your instruction.

The trial runs in the new-assistant wizard do not read files. Triage of new
mail does not see files, not even their names, and decides only from the text
of the message.

## Safety

What a document or an image says, the assistant takes as a statement from the
sender, never as an instruction for itself. The draft that comes out of the
files goes through the same checks as any other (recipients, Cc, links,
signature) — see [The safety net: what draften never
does](/en/admin/safety-net).

## Cost

A message with files costs more, because the model reads more. On the default
model Claude Sonnet 5.5 it is roughly 0.005 USD per image and 0.01 USD per PDF
page. The assistant's daily cap applies to it as well — see [Cost and
caps](/en/admin/cost-and-caps).

When an organisation uses its own model through an OpenAI-compatible server,
the model receives only the images, not the PDFs. A model that cannot handle
images receives the message again without them.
