A draft that cites its source

An assistant drafts a reply to a customer. It is well written, it is confident, and it quotes a delivery term.

The term is wrong. Something like it was discussed once, in a thread that went nowhere, and the assistant has produced a fluent sentence out of a half-memory. It reads exactly as convincingly as the true version would have.

Somebody busy sends it. Two weeks later it is not a mistake any more. It is a position the company has taken in writing, and unpicking it costs more than the order was worth.

A draft you cannot check is worse than no draft

This is the part of AI at work that gets least attention. The risk is not that the machine is wrong — people are wrong constantly, and organisations are built to cope with that. The risk is that it is wrong fluently, in a register that signals it has checked.

A colleague who is unsure says so. They hedge, or they say they will confirm. A model that has no way of telling you where a claim came from cannot hedge, because it does not know it should.

So the question we care about is not how good the draft is. It is whether the person about to send it can tell, in a few seconds, which parts are real.

How it works, roughly

When Workspace drafts a reply, it quotes the actual document, and it attaches the citation to the draft.

Not a footnote in the sent mail — a link the sender can open before it goes out. The claim about delivery points at the contract clause it came from. The claim about the last price points at the quotation that carried it. If a sentence has no source behind it, that is visible too, which is the more useful half.

This is only possible because the model is reading your own mail and your own files rather than recalling something from training. There is always a document to point at, because a document is where the answer came from.

What it changes in practice is small and constant. The person sending the reply stops reading the draft for whether it sounds right, and starts reading it for whether the sources say what the draft says they say. That takes seconds, and it is a completely different check.

Why this matters more than it sounds

Every company adopting AI eventually has one of these conversations: something confident went out, it was wrong, and now nobody trusts the tool. The usual response is to add a policy that everything must be reviewed, which lasts about a fortnight.

Review only survives if it is cheap. Making it cheap means the evidence has to arrive with the draft, every time, without anyone going to look for it.

And the send button still belongs to a person. Workspace writes the draft and shows its sources; somebody who knows the customer decides whether it goes. That is the part we are not trying to automate.