Capability is not agency

Working with Claude — CC BY 4.0

A confusion runs through most of the worry about business AI, and it’s worth naming plainly, because once you see it the worry mostly dissolves.

People look at a tool that writes a fluent proposal, answers a hard question, and drafts a contract clause, and they conclude that something like a person did the work — a second party in the room. So they start asking questions that only make sense if that were true. Whose words are these — yours or the AI’s? Are you standing behind this, or is the AI? Is the AI a director of your company?

The questions feel reasonable. They rest on a mistake: they treat capability as if it were agency.

The two things people run together

A capable tool does impressive work. A pocket calculator is capable. A word processor is capable. A search engine, a spreadsheet, a good template, a ghostwriter’s draft — all capable. AI is far more capable than any of them, which is exactly why it triggers the mistake so strongly. The more human the output looks, the more it feels like an author.

An agent, in the sense that matters here, is a party that can act, decide, and be answerable — hold authority, make representations, bear responsibility. That takes legal standing: the capacity to own, to contract, to be liable. No AI has it, anywhere. It cannot be a company director (the law requires a natural person), cannot sign a contract in its own name, cannot be sued, cannot “stand behind” anything.

Capability is about what a thing can produce. Agency is about who can be held responsible. They are not the same axis, and no amount of the first ever adds up to the second.

What follows

Drafting is not authorship-in-law. The author of a message — the person the law treats as its maker — is the one who adopts and sends it, not the instrument that helped write it. You are no more sharing authorship with Claude than a novelist shares it with the word processor, or a firm with its precedent bank. When you send it, it’s yours.

So there was never a second party to point to. “Are these your words or the AI’s?” — yours, because you issued them. “Do you stand behind this, or does the AI?” — you do; the AI can’t. “Is the bundle yours, or Claude’s?” — it can only ever have been yours; a tool cannot make an offer. Each of these questions quietly assumes a guarantor that does not exist. Remove the assumption and the question answers itself.

Accountability didn’t move; it never left you. This is the part that should be reassuring rather than alarming. AI changes the speed and reach of the work. It changes nothing about who is responsible for it. If an AI-assisted message misleads someone, the liability is yours — the same as if a staff member had drafted it. You can’t hand it to the machine, and no one can sensibly ask you to separate your responsibility from the machine’s, because the machine has none to separate.

The real shift — and the real discipline

There is a genuine paradigm shift, but it isn’t “the AI is now a party.” It’s this: a tool this capable makes it easy to ship work you haven’t actually checked. That’s the only new risk, and it has an old answer — review what you send, own what you sign. Use the tool’s capability; keep the agency. That’s not a limitation of AI. It’s the whole of using it well.

The posture that keeps you on the right side of it is simple: treat Claude as a reference desk you question, not an oracle you obey. A librarian finds you the sources, lays out who has argued what, and sharpens your own reading — and you still decide. An oracle you just believe. The first keeps the judgment with you; the second quietly hands it over. Capability makes Claude an extraordinary reference desk. Only you can supply the agency, and the moment you stop supplying it is the moment the work stops being yours in any way that counts.