At a Bloomberg finance conference in New York in October, Steve Randich, FINRA's chief information officer, described what he wants from AI at the regulator. He wants the way a police department works: the same number of officers and less crime. The useful part of that comparison is what it says about reading.
Numbers were never sampled, but words were
FINRA has used AI for about ten years on hundreds of billions of market events a day, because no team of people could read that much. Trades are numbers, and software has always been able to read numbers. Complaints, emails and videos are language, and only people could read language. So those were sampled, and the rules were written to allow it.
Language models remove that limit. FINRA's investor complaints alone run above 10,000 a year. At fifteen minutes each, that is about 2,500 hours of reading, more than a year of one person's working time, spent before anyone has decided anything. A summary tool turns that year into a queue of cases ranked by how serious they look.
The same staff finds more, and firms behave differently
If a regulator keeps its staff and each person can cover far more, the number of cases it finds goes up. Randich's goal is that the number of violations goes down, and that follows from the first. Under sampling, a sloppy habit could go unseen for years because nobody happened to pick that file. Once everything is read, luck stops protecting anyone, and firms fix the habit before it is found.
The reverse is also true for firms that use AI themselves. FINRA's position is that its rules were written to be neutral about technology, so a firm answers for what its AI writes just as it answers for what an employee writes. Randich wants to know whether firms keep their models accurate and up to date, and whether a named person is looking at the results.
What a small business can do with this
Take a small lender that, every quarter, has someone read 50 of its 3,000 customer emails to check that nobody promised a rate or a result they should not have. An AI can read all 3,000 in an afternoon and list the 40 or so that need a person's eyes. The person still decides what each one means, but the afternoon replaces a week of reading, and the sample becomes the whole set.
The same applies to a clinic's patient letters, a recruiter's job adverts, or a food business's temperature logs. In each case an inspector has so far looked at a few, and an inspector with AI will not have to.
So the first complete reading of your own records should be yours. What it finds is a list of fixes you can make before anyone else sees the problem, and that list is the main thing an AI can give a business that expects to be inspected.