The Patchwork Enterprise

Everyone adopted AI. Almost nothing changed.

Stanford's 2026 AI Index puts organizational adoption at 88 percent, up from 71 the year before. But fewer than one in ten of those organizations have scaled AI in a single business function.

The Federal Reserve measured the same moment from another angle this April: 41 percent of American workers use generative AI on the job. The Census Bureau asks firms a stricter question — whether AI is formally part of how they produce their goods and services — and gets a yes from 18 percent. The figures measure different things: AI used somewhere inside the organization, and AI wired into how the firm works.

Two years ago the worry was that enterprises would be slow to pick AI up. That worry is dead. Workers moved at consumer speed, but that speed never reached the P&L, and most companies still read like nothing changed.

The best agencies in weak states are built as pockets

Development economics has spent fifty years on a version of this puzzle, studied from the other end. In states where almost nothing works, a few agencies do. Nigeria's food-and-drug agency, NAFDAC, was rebuilt in the 1990s from a moribund directorate. Under Dora Akunyili it took on the counterfeit-drug trade and became one of the most trusted institutions in the country. China's Salt Inspectorate was set up in 1913 under Sir Richard Dane, on the ruins of a dead tax bureau. It became the most effective revenue agency of its era. Brazil's development bank ran industrial policy with a professionalism the rest of the state never matched.

Sociologist Erin Metz McDonnell calls these pockets of effectiveness, and her Patchwork Leviathan is the best account of how they happen. The recipe repeats across a century and three continents. Cluster the capable few in one unit instead of spreading them thin, then insulate that unit from the hiring, the schedules, and the favors that run the rest of the system. The pockets that last also grow an identity — the sense, inside, that we do things differently here.

None of her states got fixed wholesale. The parts that worked were built one pocket at a time.

The pocket is the right unit of AI adoption

The enterprise AI playbooks seem to run the opposite way: a transformation office, an org-wide rollout, mandatory training, one platform for all fifty thousand seats. The enterprise itself is the unit of adoption. I think that's a mistake.

Look at what small teams now do. Midjourney reached around $500 million in revenue with a team of roughly forty and no outside investors. Cursor passed $500 million in annual recurring revenue with fewer than sixty people. The products differ; the shape is the same. Everyone sits close to the same problem and can change how they work without asking permission.

A large enterprise can't turn itself into a startup, but it can carve out the same conditions one team at a time — a pocket with its own budget and its own tuned workflows. The people closest to a problem can see where AI fits, and that view rarely survives a rollout designed three levels up — the shallow-adoption numbers are what that looks like.

The enterprise even holds an advantage those weak states never had. Weak states could build only a few pockets because capable, motivated administrators were scarce and had to be hoarded where they mattered most. An enterprise starts from a fuller shelf — much of the workforce already uses AI on its own. The people who could run a pocket are already inside, already warmed to the idea.

Any budget holder can start a pocket

What counts as a pocket? The simplest test: a team whose leader controls a budget. That's the closest the org chart comes to real autonomy. Procurement and security review still gate what a team can buy, but a budget holder can fund a tool and shift the team's time without a steering committee forming first. When an experiment fails, everyone knows whose call it was.

That boundary solves a governance problem too. A company cannot see what ten thousand individuals do with a chatbot, and an org-wide AI policy is too blunt to govern much of it. A team with a named owner is different. There is a single point of contact and a single place where accountability lands. The pocket is the smallest unit that can move on its own and still be answered for.

Inside it, the leader's job mirrors a founder's: keep the team aligned while it moves fast, and keep the org's goals reflected in what the team builds.

The org's job is gardening, not engineering

A pocket runs on people who want it, and you can't order that into existence. What the center can do is raise the odds that pockets emerge on their own — free AI budgets for any team that asks, training and lunch-and-learns, loud recognition where adoption sticks. Most teams will take the budget and build nothing that lasts, and that's the cost of refusing to mandate. It's worth paying.

The center also sets the guardrails — what safe use means — function by function. Claims processing and marketing need different rules.

And a small enablement group, there for any team that asks. Every pocket that lasts has a protector — someone senior who keeps the org's antibodies off it while it's still fragile. The enablement group is that protection, made routine, and it never goes looking for work.

Your org chart predicts the curve

This approach runs on one tension, and no organization ever fully settles it. You want teams autonomous enough to innovate — to pick their own tools and decide how the work should go — but pointed at the same goals and inside guardrails that keep the risk contained. Lean too hard on control and you're back to the org-wide rollout that changed nothing; leave teams alone and a dozen of them build the same thing a dozen incompatible ways. Companies have been making that call, imperfectly, for as long as they've existed, and AI only raises the stakes: the gap between a team that's aligned and moving and one that's waiting for sign-off used to be small, and now it compounds every month.

It also makes a prediction you can check. The companies that adopt AI fastest won't be the ones with the biggest budgets or the best models, but the ones already built to let teams decide — flat, distributed, used to pushing calls down. There's precedent: when computers spread through firms a generation ago, the technology paid off where workplaces had already pushed decisions down, and the gains landed where the structure was ready for them. AI looks like the same story, faster.

The real constraint is the shape of the company, and a centralized firm can't buy its way past that. The ones that spent years pushing decisions out to their teams have been building an AI advantage the whole time, without ever calling it that, while the firms that still run everything from the center have the hardest adjustment ahead of them.