The monthly AI check: fifteen minutes, four things
Everything this blog has recommended since October, compressed into a recurring appointment that fits in a Monday morning.
Nine months of posts have produced a lot of advice. Almost all of it collapses into four things checked once a month. Here is the whole routine, written so it can be handed to whoever owns this.
1. Run the twenty examples (five minutes)
The test set from the 6 June post, against whatever your automated process currently uses. Compare to last month’s result.
You are looking for drift, not perfection. A score that has moved by two is the signal that something changed upstream — a model update, a vendor change, a new document format from a supplier.
2. Read five outputs end to end (five minutes)
Not the ones flagged as problems. Five ordinary ones, chosen at random, read properly.
This is the step that catches the failure nobody reports: output that is subtly wrong in a way the process does not detect and the person approving it has stopped reading closely. Familiarity erodes review, as the 19 February post put it, and random sampling is the only cheap defence.
3. Check what is new (three minutes)
Ask one person what AI they used this month that they did not use last month. Add anything new to the list.
Features keep arriving in software you already own. The list goes stale by itself, without anybody doing anything wrong.
4. Look at the bill (two minutes)
Not the total — the shape. Did usage jump? Did the expensive tier get used for something routine? A surprise here is always explainable and never explains itself.
Every three months, add the review
Update the page, add one clarification from a grey case, change the date. Twenty minutes, four times a year, as set out on 28 April.
Why this is the whole programme
Because AI governance in a company of thirty is not a framework. It is one person, fifteen minutes a month, and a page with a recent date on it.
Companies that do this will not have incidents that surprise them. Companies that write a policy in January and never open it again will — and the policy will not help, because the problem was never that the rules were wrong.