AI governance for non-lawyers: what every product and ops team needs before they ship agents

You already have the agent in staging. It can read the ticket queue, draft a reply and, if someone flips a flag, send that reply with no person in the loop. The team is pleased. Someone has booked legal for a month out, when a lawyer has time. On a Thursday, with an ugly backlog, someone flips the flag and the customer reply goes out.

This is the ordinary path. Product and operations move faster than policy, and they should know enough to pause themselves. Governance for people who skip the law degree is that pause, written as questions a squad can answer before an agent may act. NIST, ISO 42001, the EU AI Act and the sector rules are how the organisation files the answers afterwards. They make a poor script for the standup where the flag is about to change.

Draft, recommend or act

The demo looks the same in all three cases, which is why teams blur them. A draft waits for a person. A recommendation pushes a person towards a choice. An action changes the world. Who may take that action is the question in who may let an internal agent send. It sends, pays, grants, denies, files or tells a customer something they will hold you to.

Most of the value in the first months sits in drafts. Most of the harm sits in actions that people described as drafts until the flag changed. Write, on one line, which of the three this system is. If it can act, name the verb a customer would recognise. Send the refund. Close the ticket. Email the client. If you cannot say the verb, you are running ahead of the page.

Before the flag

A squad can answer these without a lawyer in the room. The NIST AI Risk Management Framework is the public document behind the page. Show the page to someone who owns the risk. Ship nothing on a shrug. What data does it read, under the rule you already have? Who can see the output, including vendors, logs and the person who maintains the prompt? What can it do without a person, and what is the blast radius of one wrong action on a busy day? How does a human stop it, including after hours? If the output is wrong in public tomorrow morning, who owns that call? Who feels the action and never chose to join the experiment? What will you review after two weeks, and what result means you turn it off?

Seven answers, on one page, dated, with names. If a question is still "we will sort that out", the agent stays in draft mode. Draft mode means a person sends. Someone who is theoretically available, somewhere in the building, will miss the reply that went out on Thursday.

The person in the loop

Specify the loop. Does the person see every action, a sample, or only the failures? Do they have enough context to judge, or only a suggested button? Did you staff them for the volume, or does the loop collapse at peak?

A loop nobody has time to use is a sentence on a slide. If the volume is too high for a full review, you are choosing a sample and a blast radius. Say so. Cap what the agent may do until the sample looks boring. Boring is what you want from an action that touches a customer who never asked to join the experiment.

Where lawyers still matter

People outside legal can stop a reckless ship. They cannot invent the company's position on privacy, employment, consumer promises or a regulator. Take the page to the people who own those risks before an agent acts on anyone outside the squad. Ask them to mark it, and spare them the blank document. You get a faster answer. They get a concrete system.

The AI governance workshop is for the leaders who have to turn those marked pages into a rhythm. They decide what to approve, who owns a use, and how a rule stays short enough that a squad will follow it on a Thursday. They also decide what to refuse. What that day has to settle is in AI governance training for leaders.

A rule the standup can remember

Drafts can move with a named checker. Actions that affect a person outside the room need the page signed, and a way to stop the system the same day. New tools follow the data rule you already published. A personal account stays out.

Audit the agents that act the way you audit a sample of access rights. You can skip most prompts. You need the ones that send, pay or file. If the page is missing, the flag goes back. A rule that lives only in a policy folder will lose to a backlog every time the queue looks ugly. A rule a squad has to answer before they ship will slow the actions and leave the drafts alone.