Asking Nemi AI to do something

Nemi AI is an assistant that works inside your account. You ask for something in ordinary words, it does the work with the same tools you would have used by hand, and it tells you what it did. It can find a file, write a document, fill a spreadsheet, change the document you already have open, make a share link, look at a photo and take you to the right screen, and it does all of it as you, under your own permissions.

It is part of the Pro, Max and Business plans. On every other plan the panel opens and says so. See Plans and what is in them.

Opening it#

There is a button marked Nemi AI in the bottom right corner of the screen. It is in every app: Files, Docs, Sheets, Canvas, Forms, Calendar, Gallery, Meet, Beam and Rooms. On a phone it sits just above the row of buttons at the bottom. Press it and a panel opens down the right of the window, so you can keep reading the thing you are asking about. On a phone it slides up from the bottom instead.

Some asks are a session rather than a question. To give the assistant the whole window, press Open full screen at the top of the panel, or choose Nemi AI in the sidebar. Open full screen takes the conversation you were having with it. In full screen your recent conversations are listed in the sidebar under Nemi AI, with New chat at the top, so you can go back to one with a single press. Escape closes the panel, or the list of conversations first if that is open. When it takes you to another screen, it finishes its answer first and the panel comes along on a computer, open on the same conversation; on a phone, press the button again to see it.

What a good first ask looks like#

• "Find the invoices from March and put them in one folder."
• "Make a share link for the video in Client work, with a password, expiring in a week."
• "Write a short project brief from the notes in this document and start a new document with it."
• "Convert every image in this folder to WebP."
• "What is on my calendar tomorrow?"

Naming the folder, the file or the workspace saves it a search and saves your allowance. If you say "this folder" it uses the workspace you have open, which is usually what you meant.

It knows what you have open#

When you send a message it is told what is in front of you: the document, spreadsheet, canvas or form you have open, the folder the file browser is showing, the files you have ticked, the file open in the preview, and the photo you have open in Gallery. That is what makes a three word request work, because "this" and "these" already have something to point at.

• "Convert these to WebP." with four images ticked in the file browser.
• "What is this about?" with a PDF open in the preview.
• "Add a closing paragraph about the deadline." with the document open in front of you.
• "Where was this taken?" with a photo open in Gallery.

Your browser sends the identifiers and nothing else, and every one of them is checked against your own access before it is used, so nothing can be described to the assistant that you are not allowed to see yourself. Naming things still helps when what you mean is not on screen. What this means for your data is in Your conversations and what is sent.

Editing a document while you watch#

Ask it to change a document or a spreadsheet you have open and the change goes into the same live editing session your own typing goes into. It appears where you are looking, everybody else in the document sees it arrive at the same moment, and there is nothing to reload and nothing to save.

• It makes targeted changes. It can add paragraphs to the end, replace an exact piece of wording, rename the document, and read and write individual cells in a spreadsheet. There is deliberately no "rewrite the whole document", because a whole document write would throw away the sentence somebody typed a second ago.
• Its work is marked as Nemi AI. Every paragraph and every cell it writes is stamped in the document's own history, exactly as a colleague's would be. See Version history.
• It cannot change something you may only read. Where your access to a document is read only, it says "You can read this document but not change it, so I cannot edit it for you." rather than trying.

How it answers#

It writes the way somebody helping you over a chat would: usually a sentence or two, plain, no preamble before the answer, and no emoji unless you use one first. A short reply is the intended one rather than a sign that it did not understand. It answers in the language you write to it in, and when a request could mean two things, especially before something is deleted, shared or overwritten, it asks one short question rather than guessing.

Watching what it does#

Every action it takes appears in the answer as its own row, labelled with what kind of action it is, so a page of prose never hides a deletion.

While it is working there is a Stop button in place of Send, and pressing it ends the turn. Anything it had already done stays done, because those actions really happened in your account: stopping it is not an undo.

It acts as you, with your permissions#

• It can only do what you can do. Every action is checked against your role in the workspace it touches, exactly as if you had done it yourself. In a workspace where you may only read, it may only read.
• A refusal says why. If you are not allowed to do something, it tells you that rather than pretending it worked.
• Its edits carry its name. Anything it writes into a document or a spreadsheet is attributed to Nemi AI in that document's own history, block by block and cell by cell. You are never recorded as the author of a paragraph it wrote. See Version history.

Attaching a file#

You can attach up to four files to a message with the paperclip, by dragging them onto the panel, or by pasting a screenshot. An attached file is not uploaded to your workspace. It is sent to Nemi, checked by the same virus scanner every upload goes through, and held for this conversation for 24 hours, outside every workspace: it does not appear in your files, it does not count toward your storage, and nobody else in your workspace can see it. Its size does count toward your own monthly upload volume, the same figure your uploads use, on your own plan and whichever workspace you have open, because storing it costs us the same as storing an upload even though it is deleted after a day. A file the assistant makes for you from it does not count. If your monthly volume is used up, the chip says so and the file is not held for the conversation, the same as any other refused file. That is what lets you say "make this smaller" or "turn this into a PDF" and get the new file straight back as a download, without the original ever landing in your files. A file you drag onto the panel from your files in Nemi is different: it is already in Nemi, so it is attached as that file, the same as picking it with the button beside the paperclip, and nothing is copied.

If you do want it kept, say so: "save this to my workspace" or "put this in the Invoices folder" copies it into your files as an ordinary upload, with the usual checks on the workspace owner's plan and storage. A converted result can be saved the same way. Converting an attached file uses your own daily conversion allowance, because the file belongs to your account and not to a workspace.

The chip under the message box says what state the file is in: sending and scanning, attached, or refused and why. A file that fails the scan is not kept. A refused picture is still shown to the assistant as a smaller copy so it can answer a question about it, and the chip says so; a refused file of any other kind is left out.

Finding a photo to attach#

The button next to the paperclip opens a picker with two tabs, Files and Photos, and it opens on the one you used last on that device. In Photos one search field finds a picture three ways:

• Describe it: red bicycle, dog on the beach, whiteboard. This is the same search as in Gallery. The closest pictures come first, with a line where the looser matches begin.
• Give a year, or a month with its year: 2019 or August 2019, on its own or with a description, as in beach 2019. The date is the one Gallery shows for the photo.
• Type a file name or a camera: IMG_4821, Canon, iPhone. Photos whose file name, camera or lens match are listed under their own line after the pictures that look like what you typed, or first when what you typed looks like a file name.

Describing a photo works exactly as it does in Gallery. The first time on a device the search has to download to your browser once, which is tens of megabytes, and the picker says so while it happens. Photos that have not been prepared for search yet cannot turn up for a description: the picker says how many there are and links to Gallery, where you can prepare them. It never starts that work itself. If your browser cannot run photo search, the picker says so and matches names, cameras and dates only.

One difference from Gallery is worth knowing. What a photo looks like is still matched with numbers your own device makes from your words, so the pictures are never analysed on our side. But to find file names and cameras, the picker also sends the words you typed to Nemi, where they are compared with the file names and camera details we already hold. A year, or a month next to its year, is taken out of the words first and sent as a range of dates instead. Gallery's own search field does not send your words at all. Section 3.8a of the privacy policy says the same.

From the keyboard: the Down arrow moves from the search field into the pictures, the arrow keys move between them, Space or Enter picks one, and Cmd or Ctrl with Enter adds what you picked. Escape closes the picker.

Showing it a picture#

Ask it to look at an image in your files or your photo library, or attach one. Whatever it is looking at, the model is sent a re-encoded copy, scaled down if it is large, and never the original: no location and no camera details go with it. A picture from an earlier message is not sent again on later turns, which keeps a long conversation from getting expensive.

Where to go next#