AI moves into the phone, the document and the record

The through-line today is that AI is leaving the demo stage and settling into the surfaces where work already happens: the phone in your pocket, the document you are editing, and the systems that hold your customer and financial records. For a business choosing tools, that shifts the question from "can it do something impressive in a launch video" to "does it act where my data actually lives, and can I see and control what it did afterwards."

Agents that carry quota change what using software means

The most consequential item is not a single feature but a pattern. A recap of this week's SaaStr AI event reports that "The main stage at SaaStr AI 2026 in San Mateo ran end to end on agents" [1], describing agents that are in production, carrying quota and writing to systems of record. That last phrase is the important one. There is a large difference between an assistant that drafts text you read and paste yourself, and an agent that writes directly into your CRM, your ledger or your support queue without a human keystroke in between.

The first kind of tool fails safely. If the draft is wrong, you notice before it goes anywhere. The second kind fails silently unless you have built the controls to catch it. When software can change a record on its own, three questions decide whether it helps or hurts: what exactly is it allowed to change, who approved that scope, and where is the trail that lets you reconstruct what happened. None of those are AI questions. They are the same governance questions you would ask of a new employee, and they matter more, not less, when the actor is fast and tireless.

This is why some decisions should stay with a person no matter how capable the automation becomes, a point worth reading in full in decisions automation should never make. Before you let any agent write to a system of record, decide deliberately what it may and may not do, as covered in what AI should and should not do in your business. In 360REV, actions that change data run against per-seat permissions and leave an audit entry, so an automated write is scoped and reviewable rather than invisible.

Generation is moving inside the document

Alongside the agent story, generation is moving into the tools where you already produce work. Google announced that "You can now create a full, multi-slide presentation using Gemini in Google Slides" [2], grounded in content you already hold and able to match the style of an existing deck.

The concept here is worth separating from the marketing around it. For years, generated output arrived on a blank canvas, disconnected from your own material, and the work of making it usable was copying it back into the real document. Generation inside the editor closes that gap: the output starts as an editable, native artifact rather than something you rebuild by hand. The trade-off is that a first draft that looks finished invites less scrutiny than a blank page did. A deck that is ninety per cent right and ten per cent confidently wrong is more dangerous than a blank one, because the ten per cent is easy to miss when the whole thing already looks polished. The discipline that matters is treating generated artifacts as drafts to be checked, not as finished work because they arrived formatted.

AI search arrives on the phone

Two Google Workspace updates bring existing web features to mobile. Drive's AI overviews are moving to the apps, with the note that "We're now bringing this feature to the Drive Android and iOS apps" [3], and the same is true for the assistant view, where "Ask Gemini in Drive offers you a dedicated, immersive workspace designed for deep focus" [4].

Where you can search changes how often you search. A capability that only exists on a desktop is a capability you use when you are already sitting at a desk with the file open. Moving it to the phone means the question you would have shelved until Monday gets answered from the car park. For a small team, that reduces the number of small tasks that pile up because the tool was on the wrong screen at the wrong moment. The consideration to keep in mind is that mobile search over your own documents is only as good as the documents themselves. If files are misnamed, duplicated or scattered across accounts, an AI overview will summarise the mess faithfully. The value depends on the same housekeeping that made desktop search useful.

Compliance for the code you did not write

A quieter item deserves attention from anyone whose product includes software they did not write themselves. GitHub published a piece describing its own practice, inviting readers to "Explore how the Open Source Program Office uses GitHub's new license compliance product to manage open source dependencies at scale" [5].

Every modern application is built on open source components, and each of those components carries a licence with obligations. At a small scale you can read them by hand. At any real scale you cannot, and the risk is not that a licence is dramatic but that it is ignored until an acquirer, an auditor or a customer's procurement team asks you to account for it. Treating dependency licensing as something you can track continuously, rather than reconstruct in a panic, is the same instinct behind keeping any audit trail nobody thinks about until the day it is needed. If you are evaluating tools this week, ask each vendor what it knows about its own dependencies, because you inherit whatever it does not.

Delegation belongs everywhere you work

Finally, a permissions update: Google is extending delegated Gmail access to mobile, noting that "The feature should be available to all domains on the Rapid Release track in the coming days, followed shortly by the Scheduled Release track" [6].

Delegation is how a business lets an assistant act in a manager's inbox without handing over the password, and the distinction matters. A shared password gives one person another person's entire identity. Delegation grants a specific, revocable, logged permission to act on defined mailboxes. Extending it to mobile removes the reason people fall back to sharing credentials, which is usually that the proper mechanism only worked at a desk. This is the everyday face of a larger principle covered in why permissions matter as a team grows: access should be granted narrowly, tied to a named person, and easy to take back.

The common lesson across all six items is that impressive demonstrations are no longer the deciding factor. What separates a tool worth adopting is whether it acts where your work already lives and leaves you able to see and undo what it did, a test spelled out in choose software worth using. Judge this week's announcements by that standard rather than by the launch video.