In high-stakes work, the costly move is adding the record after the agent is already running. I’d build the AI audit trail before the agent runs at all, because some mistakes get fixed and some get read back to you. A wrong figure sent to a client costs a corrected email. The same slip in work a regulator reads, or a reporter can quote, puts your name in the story.
Reading that trail each week is its own discipline, and a reviewer on autopilot is no check at all. In Alberta, the Personal Information Protection Act already expects firms to protect the personal information they hold, and a trail full of client detail is personal information in its own right.
Most planning starts with the tool: what it can reach, how fast it can work, which steps it takes off a desk. Those are real questions, and in ordinary work they come first. In high-stakes work they come second. The first set is about the trail, because the trail is what you get judged on when something goes wrong.
Why is the same mistake priced differently in high-stakes work?
Because the cost of an error is set by its audience, not by the minutes it takes to fix. A wrong draft caught inside your office costs a redraft. A wrong draft that leaves the building in regulated work can cost a filing redone under scrutiny, a client relationship explained in writing, or a headline with your firm’s name in it.
The work and the tool are the same. What the work touches changed, and that is where the risk sits.
Risk follows the use, not the tool. The same agent sorting internal notes on Monday is a different risk on Friday, drafting material a regulator will read. Planning from the tool misses it. Planning from the use, and what one mistake costs in that use, gets the order right.
In ordinary work you pay for the fix. In high-stakes work you pay for the audience.
Why can’t the record be added once the agent is running?
Because a record added later gets shaped by whatever the tool happens to save. You end up auditing a trail the vendor chose, with gaps exactly where your design should have been, and the day you need the trail is the day you learn which steps were never logged.
Adding it later means the agent has settled into its working pattern, and changing what it exposes becomes a fight with a system people now rely on.
Design order is the whole decision here. If the record comes first, it sets the terms: the agent is built so that each step lands somewhere a person can read. If a planned step cannot be recorded in that form, that is a design finding. It belongs in the planning meeting, not the incident review. You record the step another way, hand it to a person, or drop it from the agent’s reach.
A record bolted on afterwards describes the tool. A record designed first describes your standard.
What does it mean to build the agent to the record?
It means the spec starts with the trail. Every action the agent takes has to land in a record a person can read: what it opened, what it changed, what it sent, who or what was on the other end, in order.
The list is written before the build, in plain words, and the build is judged against it. A tool that cannot produce it for a given use does not get that use, however strong the demo looked.
This is the design end of the discipline I call a habit, not a binder: the record has to serve the firm every week, not sit in a file for an inspector. Once the agent is running, the standing question is if it can show last Tuesday on demand. That question only has a good answer if the record was in the spec before the first run.
If a step cannot be recorded, the agent does not get that step.
What else gets written down before anything ships?
Two more lines on the same page. The first is who can read the record, by name. A trail holding client detail is itself sensitive, so access is part of the design: who reads it in a normal week, and who reads it the day something goes wrong.
The second line is the error rate that ends the trial. Decide, while everyone is calm, how many wrong actions in how much work ends the experiment, and what counts as wrong, because a wrong send and a wrong draft are not the same weight. I’d set that number before anyone feels attached to the tool.
Set that number after a bad week and the mood of the week sets it. One bad day and the trial ends in an argument. One forgiving month and a sloppy agent becomes how the office runs. A figure written at design time turns the end of a trial into a reading, not a debate.
A trial with no end written down is not a trial. It is a habit forming.
My take: design the trail first, or you will end up auditing whatever the vendor chose to save.
What goes on your one-page agent spec this week?
Take your highest-stakes planned agent use and write one page before anything ships. Three things go on it: the record the use must keep, who can read it, and the error rate that ends the trial. If the page cannot be written, the use is not ready, and that is news worth getting this week.
That one-page spec is the first deliverable in AI Automation Systems work.
Write the three lines while the decision is still yours.
Can you write that one page for your riskiest planned agent today? If you can’t, the use isn’t ready yet. More specs like that run weekly in the AI Newsletter with ZAK.
More on AI audit trail
What is an AI audit trail?
A readable record of every action an AI agent takes. What it opened, what it changed, what it sent, who or what was on the other end, and in what order. In high-stakes work it belongs in the spec before the build, so the agent is designed to produce it.
Why design the record before building the agent?
A record added later is shaped by whatever the tool happens to save, with gaps where your design should have been. Written first, it sets the terms. If a planned step can’t be recorded, you learn that in planning instead of in an incident review.
When should an AI agent trial end?
When it crosses an error rate you wrote down before it started. Decide how many wrong actions in how much work ends the trial, and what counts as wrong, while everyone is calm. A number set after a bad week gets set by the mood of the week.
5 Comments