A Ban Is Half an AI Policy.

November 11, 2026

Watch the video version:

Late in October, I sat at a roundtable in New York with technology executives. One governance leader there keeps some uses of AI absolutely off limits: nothing touches transaction data. In the same breath, they argued that a bounded area has to exist where people can try things, because large language models have turned employees into hobbyists. People build tools at home, sometimes on open-source models, and then ask to run them at work. That is their account of their own organization. The question underneath it is one every technology leader writing an AI policy is answering right now: what do we do with the tools we didn't give them?

There are three common answers. Ban it until the policy is ready. Approve a short list of vendor tools and nothing else. Or write no rules and trust your people. Each is defensible on its own terms, and each writes one boundary at most. The ban and the vendor list say where AI can't go. "No rules" says nothing at all.

One boundary doesn't stop anyone from building. The people who build at home keep building, and their tools work. With no sanctioned place to try them, the tools reach real work through side doors: a personal account, a script on a laptop, a model running somewhere nobody reviewed. Nobody decides to break the policy. The way work actually gets done drifts away from the documented one, one useful shortcut at a time. I call that the Shadow State: the operating reality that grows alongside the documented one, because the organization adapts faster than it records. Leadership is now watching the policy, not the practice, and the risk the ban was written for, usually data, runs where nobody manages it.

I am one of those hobbyists. Outside my day job, I build my own AI tools, and the business I run alongside it runs on a system of agents I built myself. Earlier this month I used that system to plan how an agent could do test-driven development for an engineering team: take a story that is ready to build, write the tests first, write the code that passes them, and open a pull request for a person to review. I am setting it up at work now. It is going in through the front door, as a pilot, and that is possible because my workplace has both lists in writing: what is off limits, and where experimenting is allowed.

The pilot carries its own limits, written before any agent touches code. One repository. Low-risk stories only, with acceptance criteria that can be tested as written. Tests first. Lower environments only, with no production data anywhere in reach. A kill switch, and a cap on how many agent pull requests can be open at once. People review and merge; no agent ever merges, approves, or deploys. None of those limits is a ban. Each one says where the agent may run, and who decides what leaves the zone.

So write the zone before you write the ban. For every line on your off-limits list, write the matching line for where the safe version can be tried. Then give the zone its own written limits: what work it covers, which environment, what data it can and cannot see, who decides, how to stop it, and the path out to real work. The off-limits list protects what must never be touched. The zone is how you find out what is worth adopting, and it is how the hobbyists' tools reach you instead of going around you. It also answers the tension Monday's piece on AI chargebacks left open: a unit paying for its own AI cuts the slow trials first, and a written zone is where those trials can still run.

Here is the test. If someone on your team built a useful agent at home tonight, what is the path to try it at work by Friday? Name the place, the data it could see, and the person who would say yes. If nobody can answer, that agent is coming in anyway, through a side door, and the first you hear of it will be the day it breaks something. Write the answer down, and you have the second list.

The Edge Case walks through the Shadow State in full: how the way your organization actually runs drifts from the one on paper, and how to see the gap while you still can: https://TheEdgeCaseBook.com

What your newest hire can see about your organization that you no longer can, and why that window closes within weeks.

Anthony S. Jackson

Anthony S. Jackson

Anthony S. Jackson has spent 30 years inside technical organizations. He is the author of the Architecture Protocol Series: three books on the structural problems technical leaders were never told they would face. He writes the LeadershipOS™ Inner Circle, a monthly printed newsletter for CTOs and engineering managers who design teams that hold under pressure.

LinkedIn logo icon
Back to Blog