When not to use Claude
Five categories of work I tell senior teams to keep away from it, including one that costs them nothing measurable and matters more than any of the others.
I am paid to train people to use Claude well, so an article about when not to use AI is not obviously in my commercial interest. It is also the part of a session people quote back to me a year later. A tool with no stated limits gets used everywhere for a fortnight and then trusted nowhere, which is a worse position than never having started. So here are the five places I tell people to stop.
When the thinking is the job
Some documents are not deliverables. They are the mechanism by which you work out what you believe: the strategy paper you would have argued yourself into and out of twice, the difficult note to a co-founder, the position you will have to defend on a Tuesday morning.
Taking a first draft on those saves an hour and costs you the reasoning you would have done during it. You end up editing a position rather than forming one, and editing is a much weaker form of thinking. Nobody can tell from the finished page. You will find out in the meeting, when someone asks the second question.
Decisions about people, and anything you cannot check
Hiring, performance, promotion, redundancy, discipline. Claude can be useful upstream of these: tidying interview notes, checking a policy reads consistently, asking what you have not considered before you go in. The decision itself, and the reasoning that supports it, should belong to a named person who can still explain it eighteen months later.
That is partly a fairness argument and partly a practical one. Nobody has ever been comforted by the news that a process was efficient.
My working rule: if you cannot tell within a minute or two whether the output is right, and being wrong carries a real cost, do not use it there without a human who can. Anthropic is admirably blunt about this in its published advice, which lists the techniques that reduce errors and then states that they do not eliminate them, and that critical information should always be validated.
That is the company that makes the thing. It seems reasonable to take them at their word rather than argue upwards from optimism.
Confidential material, and work whose value was that you bothered
Anything covered by an obligation you have not verified stays out: client data under a confidentiality agreement, candidate information, unannounced numbers, whatever a regulator would ask about. Your position might be perfectly fine. Might is not a basis. Ten minutes with whoever holds the contract settles it permanently. This matters more than it used to, because Claude now sits inside browsers, desktop apps and Microsoft 365 rather than in one window you consciously open.
The last category is the one people underestimate. The condolence note. The thank you to someone who stayed late three nights running. The apology that should carry a cost. The content was never the point of those. The point was that a person you know sat down and spent nine minutes on you. Automating that is not a time saving, it is a small theft, and the recipient usually senses it even if they cannot say why.
What this leaves
Rather a lot, as it happens. Ruling out five categories makes the rest more usable, because people stop hedging on everything and start delegating on the work that suits it.
For the mechanics underneath the third gate, there is a plainer account of why a confident answer can still be wrong. If you are weighing up whether any of this needs formal teaching, four questions will settle it in about four minutes.
Write your own version of the list before somebody needs it at half past four on a Friday. It takes an afternoon at most, and the argument you have while writing it is the valuable part. I have never seen a team agree on the fifth category first time, and I am no longer sure they should.