After a few months with AI you develop a feel for what it does well and badly. Explaining that boundary as “hard things versus easy things” usually fails, though, because it handles things people find hard remarkably well and gets oddly confused by things people find trivial.
A better explanation is context. AI knows what you gave it, and most of what is true about your business you never gave it.
Three things it cannot see
First, the facts of your business: actual revenue and costs, the work already queued, your history with a particular client, why you decided something the way you did last time.
Second, anything unmeasured. How many seconds your site actually takes to load is knowable only by measuring it. AI can produce a plausible number, but a plausible number is a sentence rather than a measurement — and that distinction becomes decisive the moment it enters a report.
Third, responsibility for the outcome. This is the important one. The cost of a bad decision falls on the business, not the tool. A decision where responsibility does not move has not actually been delegated.
Decisions to keep
Look at the upper half and the pattern is clear: every one of them needs information only you hold. Ask anyway and you will get an answer, but an answer about the general case rather than your case. The general answer might well fit — and deciding whether it fits is, again, yours.
Where AI still earns its place
None of this means “do not use AI”. If anything, a clear boundary makes it easier to use freely.
What the right-hand column shares is that AI is widening the option set. Where you would think of two approaches alone, it lays out six and you choose. Used that way it is genuinely valuable, and being wrong costs nothing.
Keep one gate
The single most effective safeguard in practice is one review step immediately before publishing. We publish AI-assisted articles too, with a gate in between: no unverified figures, categories and links correctly attached, and only promises we can actually keep.
What that gate really does is put the context back. It is the step where a person supplies what the model could not see, which is why unreviewed generation and reviewed AI assistance produce genuinely different kinds of output.
Practical ways to bring AI into daily work are in the AI archive, and the judgement points a person has to own are visible in our published process. Work that depends on measurement and verification is exactly how our optimization program runs.