One agent. A different toolbox in every app.
Eli is not a chatbot bolted onto a product. It's a single agent surface built into the platform shell, so it appears in the same place, with the same interface, in every app — and picks up a different set of tools depending on where you opened it.
In a document
Eli gets tools for finding text, typing, applying patches, inserting tables, charts and images — and edits the page in front of you.
In a spreadsheet
The same panel, now with cell-level tools — read a range, write formulas, build a chart from the data it just calculated.
In your business apps
Ask for an income statement or a list of overdue invoices, and Eli calls the app's own registered tools rather than guessing.
It keeps a notebook in every app
Eli writes things down as it works — what you decided and why, where a file lives, a gotcha worth not repeating — and reads them back the next time you open that app. So the second conversation does not start from nothing, and you are not re-explaining your own business every morning.
One notebook per app, per person. Yours is not visible in a colleague's run, and when it grows long it is summarised down rather than truncated, so the early decisions survive instead of falling off the end.
Eli only sees what the app it's open in can do
It's not one catch-all assistant with everything switched on everywhere — what Eli can touch in a given app matches what that app is for. Open it in a document and it can find text, type, and insert a table or chart; open it in a spreadsheet and it can read a range, write a formula, build a chart from real data. In TextWork, GridWork and StageWork, Eli reads the actual content of a file you attach. In Email, Messages and Forms, attachments exist but Eli can't yet see inside them — it knows a file is there, not what's in it.
It does the errand,
not just the sentence
Most AI writing tools can only give you words back. Ask Eli for something real — "go through last quarter's sheet, find the three stores under target, and rewrite section 4 with the actual numbers" — and it will go and look, then come back and do it. A single request can take it dozens of steps.
- It reads your actual file rather than guessing from the title
- It can search the web when the answer isn't in your documents
- It works on several parts of a long document at once
- Anything that would change your work can be set to ask you first
You can stop it mid-run, and you can always see what it changed.
- You ask"Find the stores under target last quarter and rewrite section 4 with the real numbers."
- It goes and looksReads the attached spreadsheet file and pulls the real figures — it does not guess them.
- It does the workRewrites the section in place, in your document, while you watch.
- You stay in chargeEvery edit is attributed and reversible, and you can stop it at any point.
Your colleagues are in
the document too.
The same live editing that lets Eli type into a page is what lets two people work on it at once. Everyone sees everyone else's cursor, and nobody's paragraph quietly overwrites anybody else's.
Nothing gets clobbered
Two people in the same document see each other typing. In a spreadsheet, the cell someone is editing is held while they work on it, and you get a nudge instead of a silent overwrite.
You can see who did what
Every change carries its author — colleague or assistant. If a paragraph changed and you didn't change it, you can find out who did.
Carry on from your phone
The work happens on the server rather than in one browser tab, so you can open the same document on another device and pick up exactly where it is.
The part most
platforms won't
write down.
AI is genuinely useful and genuinely nosy. Most platforms — Microsoft 365 Copilot and Google Workspace Gemini included — will tell you AI is on or off at the org level. Far fewer will tell you what happens for one specific person, on one specific request. That's the level this is built at.
Taken per person, not per company
Each person accepts the AI terms once, individually, in plain language — before any AI feature runs for them. Decline and it stays off for you alone; everything else works as normal.
A cap per person, not just per company
An admin sets a company quota and, separately, one per person. Each resets on its own schedule, and you get a warning at 80% rather than finding out from an invoice.
Nobody is reading over your shoulder
Your administrator can see that AI was used, by whom, and what it cost — never the questions or the answers. There is no admin screen that shows them. A conversation belongs to the person who had it.
We train nothing on your work
No models of our own, and nothing sold to anyone. Eli does reach outside providers to answer you, and what each retains is governed by its own terms — so the two controls below are the real answer, not a promise from us.
An off switch that means it
An admin can turn AI off for the whole company, or allow it but forbid it from reading document contents. Both are enforced before a request is sent anywhere — not filtered afterwards.
Or keep it in your building
Point the platform at a model running on your own server and nothing reaches an outside vendor at all. The trade is that you host it, and it runs slower than the hosted models.