AI with access to everythingisn't modernity

There are AI assistants that promise total access: your email, your WhatsApp, your calendar, your files. Everything connected, everything “on autopilot.”

And one thing needs saying before criticizing: it works. An assistant that reads your calendar, cross-checks it with your email and gives you back a free slot saves real time. Whoever installed it wasn't naive. They were solving a real problem for someone who works alone and has no assistant.

The problem isn't the promise. It's what it asks for in return — and the fact that the request arrives on a permission screen nobody reads, in the middle of a busy afternoon, with a blue button that says “Allow.”

The intern on day one

Think about it with me. A new person joins your business today. Competent, willing, recommended by someone you trust.

Do you hand over the passwords to your email, WhatsApp, bank and calendar on day one, unsupervised?

Probably not. And it isn't because you distrust them. It's because they don't yet know what is normal in your operation. They don't know that customer can't get messages after 8 p.m. They don't know that supplier double-charges on purpose. They don't know the price spreadsheet has an old tab that no longer applies.

They'll learn. In two weeks, in a month. And then their access grows — because then they have context.

With AI the logic is the same, with one difference: AI doesn't learn what is normal in your operation unless someone writes it down for it. And it won't stop in the middle of a task to ask “are you sure?” the way a person would.

Minimum access isn't distrusting the tool

When we talk about minimum access, the common reaction is: “so you don't trust AI?”

I do. I use it every day. The question isn't trust, it's how much damage is done when something goes wrong — and something always goes wrong at some point, with any software.

If the AI that answers your WhatsApp only has access to WhatsApp, a mistake it makes is a messaging mistake. Annoying, fixable, the customer understands.

If the same AI has access to WhatsApp, email, Drive and the calendar, the same mistake can become the wrong message to the wrong customer with an attachment that was never supposed to go out.

It's the same failure. What changes is the reach.

Security in a small operation is rarely about stopping the genius attack. It's about making sure the silly mistake — and the silly mistake will happen — stays small.

What changes when the data is the customer's, not yours

There is one line that changes everything: the moment the data the AI accesses stops being yours and becomes your customer's.

Your email is your problem. A patient's WhatsApp conversation, a buyer's delivery address, a student's contract — that isn't yours. It is in your care.

The practical difference is simple. In the first case, if it goes wrong, you decide whether to let it slide. In the second, the person who trusted you decides — and they decide whether to remain a customer.

That's why it's worth separating access by that criterion, not by tool. Before connecting any AI, the question is: will this touch third-party data? If so, minimum access and mandatory review, even if it's more work. If not, you have more freedom to experiment.

It's the same tool. What changes is who pays if it goes wrong.

Setting it up per task costs one afternoon, once

The honest objection to minimum access is operational, not philosophical: it's more work.

It is. Setting permissions per task means going into each tool, understanding what it calls a scope, ticking three boxes instead of one and testing whether the flow still works afterward.

It's one afternoon. Once.

And there is a side effect almost nobody expects: when you list what each AI needs to access, you discover you no longer use half of them. The permission inventory turns into a subscription clean-up.

What doesn't work is the in-between version — giving total access “for now, I'll sort it out later.” “Later” never comes. Nobody goes back to a permission screen that is already working.

Where this argument stops holding

Two situations in which minimum access is bad advice, and it's worth saying so out loud.

If you work alone, with data that is only yours. A freelancer who connects an AI to their own personal account, with no customer data involved, is accepting a risk that is theirs and nobody else's. Demanding permission governance in that case is bureaucracy. Broad access there is a legitimate choice.

If control gets so tight nobody can stand using it. That's the opposite mistake, and it's common. When asking for access becomes a three-day process, the team doesn't become safer — it becomes creative. It starts using the personal tool, on the personal account, with company data inside. Then the data has truly left your control, and now you don't even know where it is.

Minimum access works when it's easy to ask for more. If it isn't, you didn't create security. You created a workaround.

There is no screen that warns you access is too broad. No tool sends a notification saying “you gave me more permission than I need.”

So the test is different, and it's uncomfortable: open the list of connected apps on your main account right now and try to say, out loud, what each one accesses and why.

Whatever you can't explain probably shouldn't be there.

Modernity without access limits is just convenience until the first incident.