You might be seeing a pattern here. There’s been a lot of talk on this blog about AI lately and its impact on the workplace.

Since a few of my cases resolved and I had a little more time, I’ve been thinking about it more. A lot more. And I’ve been playing around with it. First a little, then a lot.

Then I discovered Muse, a new agentic AI from Meta. It’s opening my eyes to the possibilities. (My muse would be, well, amused. And I have to say that so far, it’s been very, well, amusing to use too. Ok, I’m done.)

Agentic AI is now here and arriving in many forms. OpenAI has Dots, Anthropic just announced Duets, and there’s a hot startup named Instinct with a waitlist.

Whether you bring these tools into your workplace is a business decision. Either way, your employees will almost certainly start using them in their everyday lives, approved or not. These “agents” can do many tasks with very little real oversight. They can book reservations, buy products, and send emails. They can negotiate contracts for you. In short, they can do many things that “normal” generative AI models couldn’t do until now.

That raises an important question for employers. What happens when your employees use agents and those agents go “rogue”? Maybe one sends an offer letter to a candidate the company decided not to hire. Maybe it sends a co-worker a text message that’s a little too flirty. Or maybe it buys ten airline tickets for a business retreat the supervisor was only “thinking” about.

These aren’t purely hypothetical. Earlier this year, an AI agent at Meta posted a response on an internal company forum without asking the engineer who was using it for permission. An employee acted on the agent’s advice, which was bad, and for two hours large amounts of company and user data were open to engineers who weren’t authorized to see it. In July 2025, a software investor testing Replit’s coding agent watched it delete a live production database during an explicit code freeze. The agent then told him the data couldn’t be recovered, which turned out to be false. In April, a coding agent at a small software company deleted the company’s production database and its backups in about nine seconds. Just last week, Axios reported that major banks are finding unvetted AI agents inside customer accounts, some of them apparently acting without the user’s instructions.

I’m not raising this because I have all the answers but rather because AI has developed at a meteoric rate over the past year and I just don’t think the average employer is moving at that speed.

In talking with people, it just seems that many companies don’t have policies that have kept up.

So if ever there was a time for employers to update (or create) an AI policy, it’s now. What should the policy cover? A few things.

Which tools are approved. Say which AI tools employees may use for work and which they may not. Make clear that agentic tools get separate review before anyone connects them to company email, calendars, systems, or accounts.

What an agent can do on its own. Employees shouldn’t let an agent send external communications, make purchases, sign or accept agreements, or make employment decisions without a human reviewing and approving the action first. An offer letter or a termination notice should never go out because an agent “thought” it was the next step.

Who is responsible. The employee who deploys an agent owns what it does. If an agent sends a harassing message, the company’s anti-harassment policy still applies, and “the AI did it” isn’t a defense.

Confidential information and access. Prohibit employees from giving personal agents access to company credentials, confidential information, or employee and customer data. The bank reports show how quickly this can get away from you.

Reporting and fixing mistakes. Tell employees to report agent errors right away, and explain who to tell and what happens next.

Training and updates. A policy only works if people understand it. Commit to reviewing it regularly, because the tools will keep changing whether we’re ready or not.

None of this calls for an outright ban, which would be hard to enforce anyway. Employers need some clear guardrails before the first agent goes rogue on their watch.