What Actually Changed in Google Workspace in September 2026?
On 9 September 2026, Google announced that Gemini now acts as a cross-app orchestrator inside Workspace. From the side panel in Gmail, Docs, Slides, Chat or Drive, you can ask it to create a document, run deep research, draft and send email, schedule a meeting, or file a to-do. The gradual rollout started on 2 September 2026.
Before this, the side panel was app-bound. Generating a formatted document required being in Docs. Building a spreadsheet required being in Sheets. If you were reading a client thread in Gmail and wanted a deck out of it, you copied, switched app, pasted, re-prompted.
That boundary is gone. Google calls the underlying layer Workspace Intelligence, and the practical effect is simple: the app you happen to be standing in no longer decides what Gemini can make for you.
Five capability groups shipped together, per Google's announcement on the Workspace Updates blog: create content (Docs, Sheets, Slides saved to your Drive), deep research across folders and long threads with source attributions, draft and send emails, schedule meetings and resolve calendar conflicts, and log to-dos straight into Google Tasks.
How Do You Get Gemini to Build a Deck Without Leaving Gmail?
Open the Gemini side panel in the app you are already in, then name the output artefact explicitly in the prompt. Say "in a new spreadsheet" or "as a slide deck", not "summarise this". The artefact noun is what routes the request to the right generator and saves the file to your Drive.
This is the single behavioural change that separates people who get value from this release and people who conclude it does nothing. The side panel will happily answer you in the panel. You have to ask for a file.
Google's own published examples make the pattern obvious. From Gmail: "Create a tracker for this project in a new spreadsheet." From Docs: "Turn this proposal into an easy to read slide deck for my director using @presentation as a style reference." From Drive: "Create a customer insights deck from project Zebra using my project files including sheets, reports, and emails."
Three things are doing work in those prompts. The artefact noun (spreadsheet, slide deck). The audience (my director). The reference (@presentation, my project files).
The @ reference matters more than most people realise. Attaching an existing deck as a style reference is how you stop every output looking like a default Google template. If your team has a deck that already passes internal review, that deck is now a reusable style asset.
What Is the Human-in-the-Loop Preview Card, and Why Does It Matter?
For actions that leave your account, sending an email or committing calendar time, Gemini does not execute directly. It presents an interactive preview card. You review the drafted content, edit it inline, and confirm before anything sends. Google describes this as a Human-in-the-Loop control and it is on by default.
This is the design decision that makes the feature usable at work rather than terrifying. An assistant that can compose email from your document is only useful if it cannot send email from your document without you looking.
Two other guarantees ship alongside it. Your data is not reviewed by humans and is not used to train Gemini models. And Gemini inherits your existing permissions exactly: if you cannot open a document, neither can Gemini when acting as you.
The practical implication for your workflow is that you should treat the preview card as an editing surface, not a speed bump. The draft that appears is usually 80% right and wrong in a predictable place, typically tone and the specific ask in the final paragraph. Fix that one paragraph in the card rather than regenerating the whole email.
How Do You Write a Cross-App Prompt That Actually Works?
A reliable cross-app prompt names four things: the source material, the artefact to produce, the audience, and the structure. Miss the structure and you get a generic output you then spend twenty minutes reorganising. Naming all four in one prompt is the difference between a usable first draft and a rewrite.
Here is a template you can paste into the Gemini side panel from Gmail, Docs or Drive. Replace the bracketed parts.
Try this prompt:
Using [this thread / these files in the Project X folder] as the source, create a [slide deck / spreadsheet / document] for [my department head, who has not followed this project].
Structure it as: [1. Where we are now, 2. The three open decisions, 3. What I need from you, 4. Dates].
Use @[name of existing deck or doc] as a style reference.
Keep it to [8 slides / one page]. Use plain language and no jargon. Where a number is uncertain, mark it as an estimate rather than stating it as fact.
The last sentence is the one people skip and the one that saves you. Cross-app generation pulls figures out of threads where they were being debated, not confirmed, and presents them with total confidence. Asking for uncertainty to be flagged converts a silent risk into a visible one.
One more refinement is worth building into the habit. Cross-app generation is strongest when the source is bounded and weakest when it is vague. "These files in the Project X folder" produces a tighter synthesis than "my project files", because the second phrasing lets the model decide what counts as relevant and it will decide generously.
If the output keeps drifting, tighten the noun before you rewrite the instruction. Naming a folder, a thread or a specific document fixes more problems than adding adjectives to the request.
If your prompt library is still a scratch file of one-liners, the structure above is worth saving as a reusable block. Our walkthrough of Anthropic's free interactive prompting tutorial covers the same discipline applied to chat models, and the habits transfer cleanly.
Where Does This Break Down?
At launch the feature is English only. Google has said support for more languages will be added, without a date. For Hong Kong teams working in Traditional Chinese, that is the constraint that decides whether this is usable today, and it is worth testing before you build a workflow on it.
Four more limits are worth knowing before you promise anything to your team.
Plan gating. Availability covers Business Standard and Plus, Enterprise Standard and Plus, and on the consumer side Google AI Pro and AI Ultra, with AI Pro excluded from the scheduling functionality. Frontline Plus gets scheduling only. If a colleague says it is not there, check the edition before you troubleshoot.
Rollout timing. The gradual rollout started 2 September 2026 with up to 15 days for feature visibility. Two people on the same domain can legitimately see different things in the same week.
Usage limits. Advanced AI features across Workspace apps are subject to usage limits. A generation-heavy afternoon can hit a ceiling you did not know existed.
Permission-shaped blind spots. Because Gemini respects your access exactly, a research prompt across "my project files" silently excludes anything shared with your team but not with you. The output looks complete. It is not. If a synthesis feels thin, check access before you blame the model.
None of these are reasons to skip the release. They are reasons to test it on one workflow before you announce it to a team, because the failure modes are quiet rather than loud. Nothing errors out. You simply get a confident document built on partial information.
What Should You Test in the Next Twenty Minutes?
Run one real task end to end rather than five toy ones. Pick a live email thread with a genuine outcome attached, ask for a file, inspect the preview card, and compare the result against what you would have produced by hand. Twenty minutes gives you a defensible answer about whether this belongs in your week.
A concrete sequence that works:
--- Open a client or project thread in Gmail with at least eight messages.
--- In the side panel, run the four-part template above and ask for a one-page status document.
--- Read the output for one thing only: are the facts traceable to the thread, or did it fill gaps?
--- Then ask, from the same panel, to draft the follow-up email to the client.
--- Stop at the preview card. Edit the final paragraph. Send or discard.
What you learn in that loop is your own error profile, not the model's. Most people discover that the document is reliably good and the email is reliably too eager, or the reverse. Knowing which one you are is worth more than any prompt you copy.
The Takeaway
The upgrade here is not that Gemini got smarter. It is that the wall between apps came down, and the confirmation step in front of anything irreversible stayed up. That combination, capable and interruptible, is a good description of what a working AI tool should feel like.
There is also a cost to over-adopting. If every internal update becomes an auto-generated eight-slide deck, your colleagues will start skimming decks the way they skim mass email. The feature removes the friction of making an artefact, and friction was doing some quiet quality control. Use it where a real artefact was already warranted.
The practitioners who get the most out of this release will be the ones who stop asking for answers in a panel and start asking for artefacts with structure, an audience and a style reference. The tooling has moved. The prompting habit has to move with it.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold. Twenty-eight years of watching technology arrive in Hong Kong offices has taught us the same lesson every time: the tool is never the hard part, the workflow around it is.
Reviewed by the UD AI team.
Test How Far Your AI Skills Actually Go
You have the prompts. The next question is where your blind spots are.
UD's AI IQ Test benchmarks your practical AI ability in minutes, and we'll walk you through every step from there, from workflow design to deployment.
