Over the past year I started taking on work I would once have handed to a contractor, a consultant, or a domain expert. Not because I learned to do all of it myself. Because a draft stopped costing money and time.
Something quietly went missing along the way. There used to be a person between the draft and the decision, someone I could ask what this rests on. They could be vague, they could be wrong, but the question always had an address. Now the draft arrives at once: assured tone, tidy structure, references. There is nobody left to ask but myself.
This is not limited to small companies. The House Office of Legislative Counsel, the lawyers who write the actual text of bills, is buried in AI-assisted drafts: 5,623 drafting requests in the first sixty days of this Congress, 72% above the year before. The office is the same size it was. By the lawyers' own account, repairing what arrives takes longer than writing the text from scratch.
Producing text became several times cheaper. Checking it did not. The bottleneck moved, and almost nowhere is that shift reflected in how control is built.
Control almost always watches the process: who signed off, on which template, in what order, whether there is a signature. While volume stays low, process and result agree and there is no need to tell them apart. Once volume rises they separate, and control shows nothing, because from where it stands everything is fine: the procedure was followed.
Checking the result works the other way round. It looks at what the document says rather than how it was made, and it requires every claim to be tied to a source. In my case that came down to three conditions, and all three had to become machinery instead of a paragraph in a policy.
Every claim carries a named source and the date of that source. Sources rank against one another: an owner's decision outranks a contractor's document, a later reply outranks an earlier one. And the person who wrote it does not check it. Someone, or something, that took no part in preparing it does, before the document goes out, against the primary source.
The third condition is obvious to anyone who has ever approved a document. The second one is not, and I underrated it myself. The most assured error does not come from invention. It comes from a genuine document that was overridden two weeks ago by the next email. The reference exists, the paper is real, the conclusion is wrong, and that is harder to argue with than an obvious fabrication.
A mechanism of this kind works, but it is not a one-time setup. It has to be re-tested and corrected on a schedule: any check drifts with time and starts measuring something other than what it was put there for. The policy it replaced died in exactly that way.
Which decision in your company rests on a document nobody has opened in the past two weeks?