If you want AI to help you write on Windows, the obvious answer is Copilot. It is in the taskbar, it is in Word and Outlook, and there is a key for it on new keyboards.
It is also a particular shape of tool, and that shape does not fit every kind of writing.
What Copilot is, concretely
Copilot on Windows comes in layers that are easy to confuse.
- The Copilot app and taskbar entry. A chat window: you ask, it answers, you copy the result somewhere. Free at a basic tier.
- Copilot in Microsoft 365 apps. Drafting, rewriting, summarising and tone changes inside Word, Outlook, Excel, PowerPoint and Teams, increasingly surfaced as suggestions on a selection so you can highlight a paragraph and ask for a tighter version. This needs a Microsoft 365 subscription that includes Copilot, and it works inside those applications.
- Copilot+ PC features. A separate thing again: hardware-gated capabilities that run on a local neural processor — Recall, Click to Do, Live Captions with translation, Windows Studio Effects, improved local search. Microsoft's baseline for the badge has been a neural processor of roughly 40 TOPS plus 16 GB of memory, and the on-device story has been broadening beyond that hardware. Note what is not in the list: none of these is inline writing prediction in arbitrary applications.
Three layers, three answers to "do I have Copilot", and the writing help most people mean is in the second one.
The shape of it: request and response
Copilot's writing help is asked for. You stop, select text or open a pane, describe what you want, wait, read the result, accept or edit it. That suits work where the unit is a document: a draft you are producing, a long mail you are composing, a summary of a thread.
It suits the other kind of writing badly, and for most people that is the majority of it: forty short messages a day in a chat window, a two-line reply, a field in a web form, a comment on a ticket, a note in an app that is not Microsoft's. There, stopping to ask costs more than typing the sentence.
The two also differ in where the text goes. Copilot's generative features run in Microsoft's cloud. For business tenants that comes with commercial data protection commitments, which is a real and meaningful thing, but it is a statement about how a server handles your text rather than a statement that your text stayed on your machine. In some settings that distinction is the whole question.
The gap
Put the two together and the gap is specific:
- short messages in Slack, Teams chat, WhatsApp Desktop, Telegram
- anything in a browser that is not a Microsoft web app: CRM notes, support consoles, admin panels, forms
- third-party editors: Notepad++, VS Code, Obsidian, note apps
- non-Microsoft mail clients
- any typing while offline, on a plane or a bad connection
- any typing that cannot leave the machine for regulatory or contractual reasons
For a lot of people that is where most of the day's words are typed, one keystroke at a time, with no assistance at all.
The other shape: inline prediction
The alternative is not a chat window. It is prediction that appears next to the caret as you type, offering the rest of the sentence, accepted with one key or ignored by carrying on.
| Copilot in Microsoft 365 | Inline local prediction | |
|---|---|---|
| How you invoke it | You ask | It offers; you accept or ignore |
| Unit of output | A paragraph, a draft, a rewrite | The next few words |
| Where it works | Microsoft 365 apps | Any app that takes text input |
| Where text is processed | Microsoft's cloud | Your machine |
| Works offline | No | Yes |
| Cost model | Subscription | Varies; one-time purchase exists |
| Best at | Producing something you have not written | Finishing something you have started |
| Bad at | Short, fast, repetitive messages | Long-form drafting from nothing |
That last row is the honest part. These are not substitutes. Asking a prediction engine to draft a memo is the wrong tool, and asking a chat assistant to help with a two-line reply costs more time than it saves.
Is a small local model good enough for this?
For continuing a sentence you have already started, largely yes, and for a reason about the task rather than the model.
Predicting the rest of a sentence is a heavily constrained problem. Subject, tone, register and vocabulary are already fixed by the words in front of the caret, and the required output is short. That is a very different job from producing an argument from a one-line prompt, and it does not need the same capacity.
The practical constraint is latency rather than quality. A suggestion that arrives after you have typed the words is worthless however good it is, which puts a hard ceiling on model size — and a model small enough to answer in well under a second on a laptop is small enough to run on it.
Typeahead is built on that trade: a compact model running locally through llama.cpp, suggesting the rest of your sentence next to the caret, accepted with Tab. It works system-wide in ordinary Windows applications rather than in one editor, needs no account, and requires no connection once the model is downloaded. It is a one-time purchase rather than a subscription, and snippets and dictionary completion work before any model is downloaded at all.
Which to use
Copilot, if you write documents, you already pay for Microsoft 365, your work lives in Word, Outlook and Teams, and what you want is drafting, rewriting and summarising.
Inline local prediction, if most of your typing is short and spread across many applications, you want help without stopping to ask for it, you need it offline, or your text cannot go to a server.
Both, plausibly. They occupy different parts of the day and do not conflict. If you already have Copilot, the question is not whether to replace it. It is whether the four hundred short messages it does not touch are worth anything to you.