Article
24 Oct 2026
How much should your AI do without asking?
Once AI can act in your other systems and spend your credits, the question changes. A practical way to decide what it should do without asking first.

What I've noticed, watching people connect AI to their other systems for the first time, is that the most consequential decision of the whole session usually takes them under a second: a small box pops up halfway through something they're enjoying, and they answer it with the same practised reflex we've all built up over years of cookie banners. They click "always allow". Most of us would do the same, because stopping to read feels like losing momentum, and yet ask people afterwards what exactly they agreed to, and for how long, and you'll usually get a slightly uneasy pause.
That little box is one of the most important governance decisions a business makes about AI, and most businesses make it by accident.
What the permission settings do
When you connect Claude to another system, each tool inside that connection gets its own setting. "Needs approval" means Claude asks you every time it wants to use that tool. "Always allow" means it goes ahead without asking. "Blocked" means it can't use the tool at all. When an approval prompt appears mid-conversation, you can approve that one action or allow the tool for the rest of the chat. You'll find the settings under Customize, then Connectors, by opening a connector and working down its list of tools, and on Enterprise plans admins can set them by role across the business.
I saw this play out when I connected Claude to Higgsfield, an AI image and video platform, and asked it to turn a photo of me into a black and white caricature. Claude asked permission before generating anything, and I allowed it for that chat. When I then asked for the caricature to become a short animated video, it asked again, because generating video is a separate tool with a separate cost. That second prompt is the system doing exactly what you'd want: a new kind of action, so a new decision.
Your business already has the answer
In most businesses, a new starter can order stationery, a manager can sign off a supplier invoice up to a limit, and anything bigger lands on a director's desk. Those limits exist because the cost of a mistake rises with the size of the decision, and most of us have a story about the purchase order that should never have gone through. The businesses I've seen handle AI permissions well apply exactly that thinking to their tools.
A tool that only reads information carries very little risk, one that creates drafts carries a little more, and anything that spends money, sends something to a client or changes a record other people rely on deserves the scrutiny you'd give anyone making that call for the first time. Start those tools on "needs approval", watch how they behave, and move individual actions to "always allow" once they've earned it. For most businesses, a sensible opening position puts tools that only read information on "always allow", anything that creates on approval for the first few weeks, and anything that sends, deletes or spends on approval for good. Allowing a tool for the rest of a chat works like a temporary sign-off limit for one project, which ends when the project does.
I leave most generation tools on "needs approval". Partly that's habit, and partly I like to see what Claude is about to do before it does it, the same way you'd glance over a colleague's first few client emails in a new role before trusting them with the whole inbox.
Autonomy comes with a price tag
Every image or video Claude creates through Higgsfield draws on paid credits, and the cost per second of video varies widely between models, which is new territory for leaders used to a flat monthly licence. I had several videos to make that month, so I named the model in my prompt: Kling 4.0, which produces excellent results at a sensible price, around 25 credits for five seconds of 720p footage on Higgsfield. Leave every tool on "always allow" with no model specified, and a creative afternoon can turn into an expensive one before anyone notices (the finance director usually notices first).
Name the model you want, or ask Claude which model it plans to use before it starts, and keep approval switched on for anything that spends money until you know what normal looks like.
Paying attention is where the value shows up
Early in that session, Claude struggled to pass my photo across to Higgsfield and asked whether it could use a different upload route. I attached the image again, it carried on, and the caricature that came back was a good match for the brief: a classic big head on a small body, drawn in the style of a pencil-sketch street artist.
For the video, I let Claude invent its own story for the cartoon version of me and asked for seven seconds. The first result was fine, and in its response Claude pointed out that seven seconds was a tight window for everything in the story it had planned, so I asked for twelve. The second version was excellent, the kind of thing no small business would ever have commissioned for an about-me slide, and it took one prompt, a short follow-up and about fifteen minutes of waiting.
Neither moment was dramatic, and both depended on someone reading what came back and making a call: whether the result was good enough, whether it suited the slide it was for, and whether another round justified the credits. Switch everything to run unattended and nobody makes those calls at all.
Where to start
If your team has already connected AI tools to other systems, spend ten minutes this week checking which ones people have set to "always allow" and asking whether anyone chose that on purpose. Move anything that spends money or talks to clients back to "needs approval" until you've seen it behave.
The most useful thing that happened in that whole session was Claude telling me my brief was too tight. Permission settings cover what your AI may do without asking you. So what have you set up for the moments when it needs to tell you that you've asked for the wrong thing?
Where to go next
Our M365 AI Safety Audit checks what AI can reach across your Microsoft 365 tenant and what it shouldn't. For keeping a lid on spend, read The decision inside every AI request.