A short, honest list
I use AI tools constantly, and I am genuinely a fan. Which is exactly why the gap between what they are sold as and what they actually do for me on a hard day is worth being honest about. Here, with affection, are the things it will not do no matter how well I prompt.
It will not decide what is worth building. It will happily build anything, brilliantly, including the wrong thing, exactly as fast as the right thing. The choosing is still mine, and the choosing was always the hard part.
It will not tell me when someone on a call is quietly unhappy. The words say "sounds good." The pause before the words says something else. The model gets the transcript. It does not get the pause.
It will not push back when I am about to do something daft. It will help me do the daft thing faster and more thoroughly, with excellent formatting. Enthusiasm without judgment is a strange sort of colleague, and it has never once said, are you sure about this?
The rest of the list
It will not take responsibility. When something ships and breaks, "the agent wrote it" is not an answer anyone can give a client. A tool cannot stand behind an outcome. Only a person can, and that person is usually me.
It will not know what the client actually meant, as opposed to what they said. So much of delivery is the translation between the two, and that gap lives in context the model was never in. And it will not care whether the project succeeds. It has no stake. That sounds obvious until you notice how much of good delivery runs on someone genuinely minding whether it works out.
And it will not read the thing back to me before I send it. It will generate beautifully and hand it over with total confidence, and whether anyone actually understands what just got made is still, stubbornly, a human job.
It will not hold the context between the lines, either. A client says one thing on a call and something different in an email two weeks later, and the whole game is knowing which one they meant and why it changed. That contradiction is invisible to a tool reading either message on its own. It lives in the history, and the history lives with me.
And it will not decide when good enough is actually good enough. It will polish forever if I let it, or stop dead the moment the thing technically works, and neither is the same as knowing when to ship. That call is a judgment about the client, the deadline and the stakes all at once, and it has never once been the tool's to make.
It will build you the wrong thing flawlessly, at exactly the same speed as the right thing.
The point under the joke
None of this is a complaint. It is the opposite, really. The tools took an enormous amount of genuine drudgery off my plate, and I am not going back. The reports, the boilerplate, the first draft of nearly everything, all faster and often better than I would have managed at four on a Friday afternoon. That is a real gift, and I am not pretending otherwise.
But every one of those things it will not do turns out to be the part of the job that was always the actual job. Deciding, reading a room, owning an outcome, caring whether it works. The machine cleared away the busywork sitting in front of all that, and left me standing right on top of it. Which is a strange gift: it made the job smaller and more important at the same time.
There is something almost freeing in the list, once the panic wears off. If the machine can take the parts of the job that never really needed a human, then the parts left over are, by definition, the parts that did. That is not a threat. That is a fairly good description of work worth doing: the bits a very capable tool still cannot touch, no matter how nicely you ask it.
So if you use these tools as much as I do, I am curious: what is on your list of things it will not do for you, no matter how nicely you ask?