Formalizing the insurgency

I spent seven years as a Navy SEAL fighting insurgencies from the incumbent's side of the line. Since December 2025 I have been running one. This is what I learned from being on both sides, written for the people who are on the losing side right now and do not know it: chief executives.

You feel it

You have not named it, but you feel it. A competitor ships something in six weeks that your organization has been piloting for a year. Your board asks about AI every quarter and you answer with a training program. Your CISO says no to everything and you have the uneasy sense that she is right and wrong at the same time. Tool spend goes up. Nothing changes. Three of your people are doing things on personal accounts that you cannot see, and the work they turn in on Monday is suspiciously good.

That feeling has a name. It is what it feels like to be the incumbent in an insurgency.

What an insurgency actually is

Strip the politics away and an insurgency is a force that does not follow the incumbent's rules, does not respect its permissions, tolerates risk it will not, and iterates faster than it can respond. It lives in the seams. It does not need to win any single engagement. It needs the incumbent to keep paying the cost of being the incumbent until the incumbent cannot afford it.

Read that definition again and tell me it does not describe frontier AI in the hands of your most capable employees. It does not follow your rules. It does not respect your permissions. It accepts risk your governance would never sign. It iterates in minutes. And the value it finds is on the boundaries between your departments, precisely where your structure forbids anyone from working.

AI has no intent. The insurgents are your people. The weapon is AI.

How incumbents lose

In Iraq and Afghanistan the conventional force carved the map into areas of operation with clean lines and never moved them. The enemy learned the lines in a week. The bombs were never in the middle of a sector. They were on the boundaries, where nobody was responsible and nobody looked.

Your org chart is the map. Your permission structure is the lines. Everything that made your organization safe and legible for the last thirty years is now the place the value hides and the place your structure cannot reach.

The incumbent's instinct is to fight it: lock the tools, block the accounts, route every request through a committee. This works. Your three insurgents leave for a company that lets them operate. Your twenty quiet adopters go quieter. Your competitors keep shipping. You are safe right up until you are irrelevant.

The other instinct is to ignore it and hope the pilots mature. Ungoverned cells running on personal credentials with no shared goal eventually put customer data somewhere it should not be or ship something that lies to a customer. Then you do option one, late, in public.

How insurgencies win

Not, as the field manuals promised, with hearts and minds. I watched the Taliban win. They won with unity of goal, sanctuary, patience, brutality when it served the goal, and a decision cycle the coalition could not match. The population did not convert. It complied, and then it adapted to the new government.

Speed was the real weapon, and speed came from where the decisions were made. The fastest force I ever served in pushed judgment to the lowest validated operator and treated approval above him as a rubber stamp. Its risk appetite was higher because the person making the call had proven his judgment repeatedly, under conditions that punished bad judgment immediately. Your approval chain is not a safety mechanism. It is a decision to be slow.

The third choice

You can fight it and lose slowly. You can ignore it and lose fast. Or you can do the thing no incumbent wants to do: recognize the insurgency, give it a commander, and formalize it.

Formalizing an insurgency means six things, and a CEO signs every one of them.

One goal, one sentence. Not "adopt AI." Something a cell can test its work against, like: this company makes decisions at the speed of its best operator. Work that cannot be traced to the sentence stops.

Sanctuary, signed. The commander reports to you. Not through the CIO, not to a steering committee. You sign an acceptance of risk that names what may go wrong and states that you have chosen it over the alternative. Inside the defined area, for the defined time, security and IT advise. They do not veto. No insurgency survives without a place the incumbent's forces cannot reach, and inside a company only one person can grant it.

Cells with common doctrine. Each operator owns an area and acts inside it without asking. Every operator runs the same governance. In my shop it is called Ridgeline: the plan is attacked by adversarial models before any code exists, the finished work is attacked again by different models, deterministic checks are the only thing that can hard-stop a release, and anything the checks cannot judge goes to a signed human decision. It has governed more than two thousand pull requests. I have never hand-written a line of the code it governs.

Rules of engagement, enforced by the pipeline. The data classes and systems that are never touched are not in a policy document. They are in the checks. Models advise. Checks decide. A human judges what the checks cannot.

Logistics up front. Tokens, tools, and permissions are budgeted before the first cell deploys. An operator throttled to a hobbyist's monthly token budget is a soldier issued two magazines.

Phases with criteria, and the endgame said out loud. Latent, guerrilla, conventional. Each transition is your decision with named criteria. And on day one you say the quiet part: the insurgency becomes the government. The way the cells work becomes the way the company works.

Casualties

It cannot be made safe. Roads will be mined. Some projects will fail in ways that cost real money and a few careers. People who built their standing on the old permission structure will lose it. Anyone who tells you there is a version of this without casualties has never fought one, and is selling you option two with a nicer deck.

Here is the honest arithmetic. About three percent of your people will go all the way in. Twenty percent will use the tools and use them badly. Eighty percent want the benefit without touching the terminal, and no training program on earth changes that, because it is human nature, not a skills gap. The eighty do not convert. The insurgency's job is to make things happen for them invisibly, inside the tools they already use, so the work of the week changes without their being asked to change.

And the truth underneath, which the leaders of every frontier lab have been saying in public for years: if a person cannot direct AI to do their work by the end of this decade, their options will be devastatingly limited. There is no pretty version of that. War is not pretty either. The only decent response is to put the capability in more hands and let the record of how it is governed be visible to the people it governs.

Why me

I am the insurgency. I lead it with forward-deployed operators, and I built the doctrine they run on because the insurgency in most companies right now is unchecked, unbounded, and losing for lack of a goal. I have fought this kind of war from the incumbent's side and I know how it ends when the incumbent refuses to name what is happening.

This site is where I put words to the thing you feel. It is not where anything is sold. When you are ready to formalize, the company that does that work is Uncommon Waters. Until then, read the doctrine, take the qualifier, and argue with me.