Connected isn't constant: when your assistant actually reaches into your inbox

Overview

Through 2026 the link between your assistant and your other apps has quietly become a default rather than an exception — Microsoft’s admin docs (updated May 2026) say Microsoft 365 Copilot’s personalisation and memory are on unless someone turns them off, and the consumer assistants make connecting a one-tap opt-in. This makes it worth being precise about what “connected” actually does, because it’s easy to get wrong in both directions. It does not mean the assistant reads all of your email on every answer. It means it can retrieve specific details when your question calls for them — Google describes Gemini’s version as “retrieving specific details from, say, an email or photo to answer your question.” Selective, not constant.

By the end you’ll have a truer read on what a connection buys, and the one habit that matters for the answers that have to stay clean.

The content

The obvious read is binary: either a connected app sits dormant until you point at it, or it’s an open book the assistant pages through every time. Neither holds. Connecting a source puts it on call — there for the assistant to draw on when it judges your question relevant, and left alone when it doesn’t. Google’s framing for Gemini is retrieval to answer a specific request, not a running feed of your inbox. The enterprise tools that keep a memory of you, like Microsoft 365 Copilot with personalisation on by default, apply it the same way: as context when it fits the task, not a recitation poured into every reply.

So the real issue is narrower than “it’s reading everything”, and it’s mostly about visibility. Because the assistant decides, question by question, whether a connected source is relevant, two things follow. You often can’t tell from an answer alone whether it leaned on your personal data or not. And relevance is its call, not yours — so once in a while it reaches into a source you wouldn’t have thought to involve: a market scan that quietly weights a supplier sitting in your inbox, a “fresh” plan already shaped by last quarter’s calendar. Not often. But you won’t see it coming, and you can’t assume an answer stayed clean just because the topic looked unrelated. It matters more the less of the source you chose yourself. In an enterprise tool answering from a shared knowledge store — where one connector can place hundreds of documents within the assistant’s reach — what’s on call is far larger and less familiar than your own inbox, and a confident answer can lean on material you’d never have thought to include.

That turns the whole thing from worry into two ordinary controls, and neither is dramatic. First, scope what’s on call: connect the apps whose contents you’re content for the assistant to reach when relevant, and leave off the ones you’re not — fewer sources on call, fewer surprises. Second, when an answer genuinely has to be uncoloured by your history, don’t trust that it stayed out; take it out of play. Gemini lets you switch personalisation off for a single prompt or drop into a temporary chat; the temporary or incognito modes in other assistants do the same job. The connection is fine to keep. The discipline is per-question, and only on the questions that need it.

Try it

Two moves, both small. First, scope what’s on call. Open your assistant’s personalisation or connected-data settings — whatever your tool names it — and read what’s linked. For each source, ask one plain question: am I content for the assistant to reach into this whenever it decides a question is relevant? Keep the ones you’d genuinely want consulted; disconnect the ones you linked once and forgot.

Second, on a real question you’d want answered without your own history leaning on the scale, take it out of the equation rather than hoping it stayed out. Before you trust the result, you can also ask the assistant to show its working:

For the question below, first tell me: will you draw on any of my connected
apps, files or saved memory to answer it — and if so, which specific items?

Then answer it twice: once using them, and once as a cold read that uses
none of my connected data or history. Tell me where the two differ.

My question: [your real task]

Where it breaks: this is a check, not a guarantee — a model can’t reliably report everything that shaped a reply, so a clean-looking self-report isn’t proof it stayed out. The dependable control is the setting, not the prompt. If an answer truly must be uncoloured, switch personalisation off for that question or run it in a temporary chat — and keep the list of connected sources short enough that what’s on call is only ever what you’d consciously invite in.

Additional reading

• Personal Intelligence: Connecting Gemini to Google apps — Google (14 Jan 2026) — the selective-retrieval framing (“retrieving specific details from, say, an email or photo to answer your question”), the opt-in model (“off by default: you choose to turn it on”), and the per-answer escape hatches (regenerate without personalisation, temporary chats).
• Manage Copilot personalization and memory — Microsoft Learn (updated 22 May 2026) — the enterprise default: “Enhanced personalization” is turned on by default (the feature is in preview and subject to change).
• Gemini rolls out free Personal Intelligence & Memory — 9to5Google (27 Mar 2026) — the per-prompt off switch in the Tools menu, and the rename of “Past Gemini chats” to “Memory”.
• Use Personal Intelligence & Memory in Gemini Apps — Google Help — the live settings path, and the check you can ask directly: “Did you use any info from past chats?”

Editor’s note

This module is worth keeping in the back of your mind. It’s one that is unlikely to hurt you in an obvious way, but has the chance to degrade your output unnecessarily. It’s especially relevant where you are using a tool that calls on a context library that you didn’t personally curate, which happens all the time in enterprise environments. If your AI tool is using a connector that can deliver 400 documents to an agent, you might need to think a bit deeper about what’s actually needed.