Copilot gets eyes
Microsoft has spent two years teaching Copilot to read what you've written. Vision in Microsoft 365 Copilot, rolling out through July 2026, teaches it to look at what you're looking at. Start a voice conversation, share your screen or point your phone camera at something, and ask a question about it directly — a confusing settings dialog, a dashboard that doesn't make sense, a printed report, a piece of equipment on a factory floor.
On Windows, you share either the full desktop or a single window and talk Copilot through what's on it. On mobile, you point the camera instead. Copilot combines whatever it sees with your existing work context — email, documents, chat history — to answer in the moment rather than asking you to describe the problem in text first.
Why this one is different from the usual feature drop
Most Copilot features extend what it can read. Vision extends what it can watch, in real time, during a live session — and that's a meaningfully different privacy shape. Microsoft's design is reasonably careful about it: once a session ends, Copilot doesn't retain the screenshots or camera frames themselves. But the voice transcript of the conversation does persist in chat history, and whatever you narrated or asked about while sharing your screen is captured in that transcript, along with Copilot's response.
Put plainly: the screen itself isn't kept, but a written record of what was on it, in your own words, might be. Anyone rolling this out should treat a Vision session exactly like an open screen-share in a Teams meeting — nothing you wouldn't want summarised in a chat log should be visible while it's running.
What's actually in the user's control
The feature requires a Microsoft 365 Copilot licence, microphone access, and an active voice chat — it won't switch itself on outside a conversation the user has deliberately started. Admins can turn Vision off tenant-wide without affecting ordinary Copilot voice chat, and users sharing a desktop can choose to share a single window rather than the whole screen, which is worth teaching explicitly: window-only sharing is the safer default for anyone who regularly has email, chat or a second monitor open next to whatever they actually want Copilot to look at.
It's also worth being clear with users about what Vision doesn't do. It analyses; it doesn't act. It can't click a button, fill in a field or drive an application on your behalf — if a user describes a task and expects Copilot to carry it out on screen, that's a different feature (and a different training conversation) entirely.
The training and governance angle
We wrote back in March about the oversharing risk baked into SharePoint permissions — content Copilot can technically already reach. Vision is a different flavour of the same underlying issue: not what Copilot can search, but what a user might accidentally hold up to the camera or leave visible on a shared screen while they're focused on the question they're actually asking. A finance user diagnosing a broken formula with screen-share Vision, with three other tabs open including a draft redundancy list, is a realistic five-minute mistake, not a hypothetical one.
Three things are worth doing before Vision reaches your users rather than after.
1. Decide your admin posture deliberately, not by default. If your organisation handles particularly sensitive visual material — clinical records, unreleased financial results, anything under NDA — decide now whether Vision should be on, off, or restricted to specific groups, rather than discovering the setting after someone's used it.
2. Teach window-only sharing as the house style. It costs nothing and removes most of the accidental-exposure risk in one habit.
3. Add one line to your Copilot training deck. Vision sessions behave like a screen-share in a meeting — anything visible while you're sharing should be treated as seen by whoever reviews the record, including Copilot's own logs. That single sentence will do more good than a page of policy nobody reads.
Where this is heading
Vision is a genuinely useful feature — talking Copilot through a confusing dashboard is a far more natural interaction than typing out a description of it, and the mobile camera use case (pointing at a piece of misbehaving equipment, a whiteboard, a printed invoice) solves a real gap in what text-only Copilot could ever do. The organisations that get the most out of it will be the ones that spend twenty minutes on the privacy habit before the rollout, rather than explaining it after the first uncomfortable screen-share.