What's the actual sentence?

"This tool may [action], on [data], in order to [outcome] — and nothing else."

Fill in the three brackets for every AI connection on the list from yesterday's audit, and you've turned a vague grant of access into something you could defend to another person. That's the whole method. The value isn't in the sentence's cleverness — it's that writing it forces you to notice when you can't finish it.

Why does writing it down matter more than just knowing it in your head?

Because "I sort of know what it does" and "I could say exactly what it does" are different states, and only one of them survives contact with a problem.

Here's the test: if that tool did something you didn't expect next week, would your scope note tell you immediately whether it was doing its job or exceeding it? A sentence that says "this tool may read my calendar, on event titles and times only, in order to suggest meeting slots — and nothing else" answers that instantly. A shrug doesn't. The permission contract this newsletter recommended for hired agents back in August made the same point at a bigger scale — one page instead of one sentence — but the underlying discipline is identical: write down what "in scope" means before you need the answer, not after.

Doesn't this take forever if I have a dozen tools connected?

It takes about ninety seconds per tool once you've done the first two, because the format doesn't change — only the three blanks do. A dozen tools is fifteen minutes, once, and then it's a document you update only when something changes rather than a task you repeat.

Keep the sentences together in one place — a single note, not scattered across each tool's own settings — because the value is in seeing them side by side. That's when the real problem usually shows up: not one bad connection, but two or three sentences you genuinely can't finish, sitting next to a dozen you can.

What if I literally can't finish the sentence for something?

Then you've found the answer to Tuesday's "keep, check, or cut" without needing to think about it separately — an unfinishable sentence is a cut, not a check. If you can't say what a tool is for, in one sentence, right now, it is not doing a job important enough to justify the access it's holding. Revoke it. Saturday's post is the two-minute version of that action.

The harder, more common case is the tool where you can finish the sentence, but the scope it was actually granted is bigger than the sentence describes — a note-taking tool that only needs to read your calendar but was granted the ability to edit it too, because that was the only toggle on offer. That gap between the sentence and the actual grant is tomorrow's post in full: the difference between delegating a task and handing over a master key, and it's the single most common way access quietly outgrows its job.

Does this replace reading the tool's actual permissions?

No — it sits on top of them, and that's deliberate. The permission screen tells you what a tool can do. The one-sentence scope note tells you what you've decided it's for. Keeping both means you can compare them at a glance, which is the entire audit in miniature: does the grant match the job, or is there daylight between them? Most of the daylight you'll find this week won't be sinister. It'll just be a default nobody had a reason to narrow — until now.

For the guide that finds every place AI already touches your business — free, 5-min read: The 10-Hour Audit.

Related Decodes

• The Agent You Never Approved — this week's hub.
• The 5-Minute Connected-Apps Audit — yesterday's post: the list this sentence gets applied to.
• Your Agent Is Now a Supplier. Give It a One-Page Contract. — the same discipline, one page instead of one sentence.
• Delegate the Task, Keep the Key — tomorrow's post: what it means when the sentence and the grant don't match.

AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.

Not financial or legal advice. A written scope note is a personal accountability habit, not a legal or contractual instrument.

TL;DR: Write one sentence for every AI tool connected to your business — "this tool may [action], on [data], in order to [outcome], and nothing else" — and keep all the sentences together in one document. The exercise takes about ninety seconds per tool once you've done the first two, and its value is comparative: it lets you see, side by side, which connections you can account for and which you can't. Any tool you can't finish the sentence for is a cut, not a "maybe" — it isn't doing a job important enough to justify the access it holds. Any tool where the sentence is narrower than the actual permission it was granted has outgrown its job quietly, through a default nobody bothered to narrow, and that gap is the thing worth fixing next.