x
The category shifted from advice to action. For three years the useful question was "what should I write here?" Now the tools can open the folder, read the files, change them and save. That is a different relationship, and it deserves a different set of habits.
The capability is genuinely useful and the marketing around it is genuinely overheated, which makes this hard to think about clearly. So strip it back: what changed is not intelligence. It's reach. The same system that could give you a wrong answer can now give you a wrong answer and act on it.
Because a wrong answer you read is a wrong answer you can reject. A wrong answer that has already renamed forty files is an afternoon of your life.
The error rate didn't necessarily get worse. The cost per error did. That's the whole shift, and it's why "is it accurate enough?" is now the second question rather than the first. The first question is: what can it touch, and what happens if it's wrong there?
Two failure modes are worth naming, because they behave differently:
Reversible. It drafts something, files something, tags something. If it's wrong, you undo it. Annoying, not expensive. Most useful agent work lives here, and it's where you should start.
Irreversible. It sends, deletes, pays, publishes, or replies to a customer in your name. If it's wrong, the wrongness is now outside your control and in someone else's inbox. This is where the caution belongs — and it's where most of the demos live, because irreversible actions look more impressive.
The practical rule is unglamorous: give reach for reversible things, withhold it for irreversible ones. Draft, don't send. File, don't delete. Prepare, don't publish. You keep nearly all the time saving and give up almost none of the safety, because the review step on a draft takes seconds and the review step on a sent email doesn't exist.
Three things, in order, none of which take long.
Inventory the reach you've already granted. Most people have connected more than they remember — a calendar here, a drive there, an inbox integration from a trial they didn't cancel. Half an hour with your account settings is the highest-value AI work available to most solo businesses this month, and it's the same instinct behind the mid-year stack review: what's connected, what's earning its access, what should be revoked.
Separate the folders. Agent access to a working folder is fine. Agent access to everything is a decision you should make deliberately rather than by default. Client material, financial records and anything covered by a confidentiality obligation belong outside the reachable area — not because the tool is malicious, but because a mistake there is expensive and a mistake in a drafts folder isn't.
Write the escalation rule. Same four lines as Tuesday's post, applied to reach: what does this agent do when it isn't sure what it's looking at? "Stop and ask" is a valid answer, and a system that asks twice a day is far cheaper than one that guesses twice a day.
Here's what can be verified: from 2 August 2026, EU AI Act transparency obligations apply — including disclosure that content is AI-generated and that a person is dealing with an AI system. Machine-readable marking for systems already on market follows from 2 December 2026, with a lighter route for smaller businesses. If your agent produces anything customer-facing, that date is the one in your calendar. (Not legal advice — check your own position.)
What isn't checkable is most of the productivity claim. Vendor figures for time saved are marketing until they're your figures, measured on your work. Run it for a fortnight on something reversible, count the minutes honestly, and score it the way you'd score anything else — minutes back per pound. Your number is the only one that means anything.
Open the permissions screen on whatever you've connected and read it as though someone else had set it up. Revoke what you don't recognise. That single half-hour is worth more than any tool you could add.
For the shortlist we'd actually connect, the free 5-min read is here: 10 AI Tools That Save 10 Hours a Week.
Related Decodes
TL;DR: July's agent launches changed reach, not intelligence — the same system that could give a wrong answer can now act on it, so the cost per error rose while the error rate didn't necessarily fall. Grant access for reversible actions (draft, file, prepare) and withhold it for irreversible ones (send, delete, pay, publish). Inventory what you've already connected, keep client and financial folders outside the reachable area, and note that EU AI Act transparency duties apply from 2 August 2026.
