Harness

A firm is trialling two AI products for drafting suitability reports. Both vendors name the same AI model. The compliance officer builds a test client, Martin Doyle, and leaves out the attitude-to-risk questionnaire on purpose. One product links each figure in its draft to the page it came from, and notes that the questionnaire is missing. The other writes a fluent report in the firm's tone, with a pension value that appears nowhere in the file. It never mentions the questionnaire. If the model is the same, why does only one draft tell the truth about the file?

The model is the part both products share. On its own, it reads what it is sent and writes a reply, and it keeps nothing from one request to the next. The software around it chooses what goes into the model's context, adds instructions, and can check the draft against the file before anyone sees it. The products differ in that software, and so do their drafts.

That software is called the harness, and the word is centuries older than computers. Before engines, a horse in harness pulled the plough. The collar took the load, and the reins let the ploughman keep the furrow straight. An AI harness does the collar's job for a model, putting it to work on one client's file and in the firm's format. Nobody holds the reins while it drafts. Whatever keeps it on course has to be set up beforehand.

An AI harness also sets limits, as a harness and lead do for a strong dog. Some products are only allowed to produce a draft for someone in the firm to check, while others are allowed to write into the back-office system. And like a climber's harness, tied to a rope that holds a fall, it can catch a slip before anyone reads the draft. Taken together, a harness is the software around an AI model. It decides what the model is given, what it may do, and which checks run on its work.

A firm already puts controls around its work, such as the template, the system access, the sign-off and the file check. They apply however a piece of work was drafted, and a harness is the AI's version of them. That comparison breaks twice. The firm can read its own controls, but most of an AI harness is written by the vendor, and the firm usually can't see it. An update can change how the product behaves with the same model underneath. The same test client shows whether the product still flags the missing questionnaire. That test is worth repeating at set intervals, because the firm may not be told when the product changes. A firm's controls also lean on people who notice what no procedure listed, while a harness catches only the slips it was built to look for.