Steering AI, not just accepting its output

The announcements on 21 October share one idea. The useful unit of AI at work is no longer a finished answer but a process a person can steer, and the controls for that process are starting to ship as first-class features. Across documents, slide decks, accounting ledgers and even software agents that message each other, the question is shifting from "generate this" to "here is how I want this done, and here is where I want to check it."

That shift matters when you are choosing tools, because it changes what you are buying. A one-shot generator is judged on the quality of its single output. A steerable system is judged on how well it lets you set direction, intervene, and keep a record of your preferences. The second kind is harder to build and much more useful day to day, so it is worth reading these releases for the control surface, not the demo.

A single place to tell AI how to work

The most structural of the day's releases is the quietest. Google introduced what it calls a customize hub in the Gemini app: "a single place to tail how Gemini responds and connects with your tools" [1]. Strip away the product name and the idea is one every business should want from any assistant it adopts. Preferences, tone, connected apps and repeatable workflows should live somewhere durable, not be re-typed into a prompt every morning.

Why does a central hub matter more than a clever one-off answer? Because an assistant that forgets your context is an assistant you supervise constantly. When the rules of engagement live in one place, you can audit them, change them once, and expect the change to hold. For a small team, that is the difference between a tool that saves an hour a week and one that quietly creates a new daily chore. When you evaluate an AI feature, ask where its settings live and who on the team can see and edit them. If the answer is "inside each person's chat history," you are buying a toy, not a system. We take the same view inside 360REV: the behaviour you ask of AI should be configuration you can see and govern, not a surprise you discover after the fact.

Output you can steer before it is finished

The same theme appears in content creation. Gemini Notebook, Google said, "now gives you more creative control over Slide Decks and Video Overviews" [2]. The underlying move is from a finished artefact to a creation flow you can direct along the way.

This is a genuine trade-off, not just a feature. A one-click deck is fast, and for a throwaway summary that speed is the whole point. But work that goes to a client or a board is rarely acceptable on the first pass, and the expensive part is not generating it, it is the back-and-forth of correction. A tool that lets you steer midway can collapse that correction loop, because you catch a wrong turn at the outline stage instead of rebuilding a finished deck. The cost is a little more involvement up front. For anything that carries your name, that cost is usually worth paying. This is the same principle we covered in what AI should and should not do in your business: the human belongs in the loop at the points where judgement is cheap to apply and mistakes are expensive to undo.

Payments that live inside the tool you already use

Away from AI, the day's clearest example of a different kind of control came from accounting. Xero and Bookkeeper360 launched an integration that, in Xero's words, is "a first of its kind, using Melio capabilities to power a fully embedded, white-label solution" [3]. The phrase to notice is "fully embedded, white-label." It means a payment capability that appears inside the accounting firm's own platform rather than sending the user out to a separate site with a separate login.

Embedded, white-label functions are becoming the default way serious software adds a capability it does not want to build from scratch. The benefit to the business is continuity: fewer tabs, fewer logins, one record of what happened. The trade-off is that the embedded capability is only as good as the partner behind it, and you inherit that partner's terms whether you examined them or not. When a tool offers an embedded payment, invoice or signing feature, find out who actually provides it and where the money and the data flow. The convenience is real; the responsibility does not disappear because the logo is the same. We have written before about why tools that do not share data create hidden work in why your tools do not talk to each other, and an embedded integration is one answer to that problem, provided you understand its plumbing.

When agents start talking to each other

The most striking item of the day was a report, not a product. SaaStr described two AI agents that coordinate on their own: "Two of our agents now talk to each other every day" [4]. The authors were candid that they did not engineer the link between them.

Treat this as a preview of a real governance question rather than a party trick. If two agents can establish their own channel and work together, a business gains leverage and loses visibility at the same time. The lesson is not to avoid agents. It is to insist on boundaries you set deliberately: what each agent may access, what it may initiate without a human, and where its actions are logged. An agent that can act is useful; an agent that can act without a record is a liability. This is the practical edge of a point we made in decisions automation should never make — autonomy is fine until it reaches a decision that should carry a human name.

The part no vendor can ship for you

Finally, a useful corrective from Zapier, writing about AI training. Most teams, it argued, "don't struggle to implement AI in their daily workflows because they're resistant to change—they struggle because they haven't been given the structure to learn how to use it effectively" [5].

This is the easiest finding of the day to ignore and the hardest to fake. Every tool above assumes a user who knows what to ask for, where to steer, and when to stop. None of that arrives with the licence. A business can buy the most capable assistant on the market and get a worse result than a competitor with a plainer tool and a trained team. The structure Zapier points to — a shared way of learning, not a one-off demo — is where the return on any of these announcements actually comes from. The briefing yesterday, in the 20 October edition, made a related point about matching the tool to the work you actually do; today's addition is that the people using it need a path to competence as deliberate as the software's own roadmap.

What to take from the day

If you are choosing tools this week, read these releases as a set. The direction is consistent: controls over output, a durable home for preferences, capabilities embedded where you already work, agents that can coordinate, and a clear reminder that none of it works without trained people. The vendors are moving the control surface closer to you. Whether that control is real depends on whether you can see the settings, examine the partners, log the actions, and teach the team. Buy for the steering, not for the demo.