How do you decide which AI tool a piece of work belongs in?

Three properties of the work decide this and none of them is a property of the tool. What the material is: whether it belongs to somebody else, whether it identifies a person, whether the business is under an obligation about it. What the work must reach: whether it needs your files, your inbox, your records, your codebase, or nothing at all. And where the output goes: to you, to a colleague, to a customer, to a public page, or into a system that acts on it. Answering those three puts most tasks in a tool before capability is discussed.

Sensitivity rules tools out, and it does so absolutely rather than on a sliding scale. Material held on somebody else's behalf, anything the business has contracted not to disclose, anything a professional rule governs — that work belongs only in an arrangement whose terms permit it, and no amount of capability makes a non-permitted tool the right one. The useful move is to establish this once per category of material rather than per task, because per-task judgement is what fails at four o'clock on a Friday.

Access rules tools in, and it is the constraint people underweight. A task that requires reading twelve files in a repository, or the last three months of a mailbox, or a live query against records, can only happen in something connected to those. Most disappointment with these tools comes from asking a disconnected one to do connected work and receiving a plausible answer assembled from nothing. Where the work needs context the tool cannot reach, the answer is not a better prompt.

The destination decides the review, which in turn decides how much the placement matters. Work that returns to you and gets checked as you use it can go almost anywhere, because you are the verification. Work that goes to a customer, a public page, or a system that acts on it needs a review step, and where that review will actually happen is part of the tool decision — a tool that produces output somewhere nobody looks has placed the work badly however good the output is.

Only then is capability worth discussing, and it is a real question rather than a formality. Long, structured reasoning; large volumes of material at once; code that has to run; reliable formatting; retrieval over a corpus. These differ genuinely between products, and choosing among the tools that survived the first three filters on those grounds is sensible. Choosing on those grounds first is how sensitive material ends up in a consumer account because it produced better prose.

The practical output of all this is a placement rule, not a decision each time. Something like: client material and anything under an obligation goes in the sanctioned business tool; anything touching our own systems goes through the connected assistant; public research and drafting with no particulars can go anywhere; nothing that reaches a customer goes without a named reviewer. Written down, that takes a page, survives a change of vendor, and removes the case-by-case judgement that is where placement actually goes wrong.