Your AI agents will inherit your worst habits

You’ll hear that an agent learns your firm from your policies and your stated values. My view is that it learns from your records, the approval waved through because the client was waiting and the fix that lives in one person’s head. That’s why I think AI agent governance has to start before the agents do. An ordinary week in your firm is a training set now.

Canada’s privacy commissioners make the same point from the legal side. Their principles for generative AI say accountability for decisions stays with the organization, not the automated system, so a habit an agent copies is still yours.

Every firm runs on two sets of rules. One set is written down, the way work is supposed to move. The other set is what people actually do when the day gets busy, and it lives in the records the firm keeps without thinking about them. Your messages, your tickets, your approvals, your shared files. An agent joining your firm studies the second set, because the second set is the one with examples in it.

Where does an agent learn how your firm works?

From the record you already keep, exactly as you keep it. An agent learns a job by studying how the job has been done, the finished work, the threads around it, the calls people made when the steps ran out.

It reads the approved version and the version somebody shipped at 6 pm on a Friday with the same attention. Nothing in the record is labelled official or improvised. There is only what happened, and how often.

That is the part owners miss. You cannot hand an agent your intentions. You hand it your history, and your history includes every workaround your team built to survive a busy season and never took apart again.

An agent cannot tell your standard from your shortcut. It only knows which one shows up more.

Which habits does it copy first?

The repeated ones, and in a busy firm the repeated ones are usually the workarounds. The approval that gets granted before anyone reads the attachment, because the reader trusts the sender. The estimate that gets padded a little, because the last honest one caused an argument.

The job that leaves through a side door, a personal inbox, a quick call, because the front door takes three days. The answer sent to a client before anyone checked if it was true, because the client was loud and the day was full.

None of these habits survives because your people are careless. Each one solved a real problem on a real day, and then it stayed. A person carrying a habit still has judgement around it. They know when the shortcut is safe and when the day calls for the long way. The agent inherits the move without the judgement that used to travel with it.

Whatever your firm does on its busiest day is the behaviour you are about to scale.

Why does this land before the purchase?

Because after the purchase, the habit stops being a habit and becomes output. One person waving through an approval is a bad afternoon. An agent trained on a year of waved-through approvals does it every run, at machine speed, in your name, and it does it with complete confidence because your record taught it that this is how approvals work here.

The harm was never the tool. The harm is the habit, scaled past the point where anyone notices it happening. I’d fix the habit first and buy the agent second.

This is why I keep governance at the level of practice. I have argued that compliance holds up as a habit, not a binder, and inheritance is the same argument one step earlier. The trail that lets you ask an agent to show last Tuesday is the same trail it studied on its way in. If the trail embarrasses you, the teaching embarrasses you, and no setting on the purchase order changes what the record says.

Scale does not fix a habit. Scale repeats it until it reads as policy.

Which habit should you fix before the agent arrives?

Pick one process an agent will touch in your firm this year, the invoicing, the follow-up, the intake, whichever is first in line. Then write down, in one sentence, the behaviour in that process you would be ashamed to see scaled.

Not the worst thing that ever happened. The ordinary thing, the corner your team cuts every busy week, the step everyone skips and nobody wrote down as skipped.

Then you have one move, and it comes in two shapes. Fix the habit first, run the process the written way long enough that the record shows the written way, and let the agent learn from that. Or fence the agent off the process, keep it on work your record can defend, and build the record before you build the agent, until the habit is fixed and the trail proves it. What you cannot do is buy the agent, point it at the old record, and act surprised when it behaves like your firm on a busy day. Reading a firm’s own record before its agents do is how my AI Automation Systems work starts.

Short answers on AI agent governance

My take: fix one habit this month, then let the agent learn from it. More of these fixes land every Friday in the AI Newsletter with ZAK.