“With fewer permission gates, everyone becomes a founder ... A tool someone builds for themselves can become a team’s agent, then a shared pattern, then an internal product.”
Vanessa Meyer is describing companies. Government agencies run under different rules about who may act.
Meyer, who writes the newsletter The Coordination Layer, argues that AI-native organizations are replacing bureaucratic coordination with infrastructure. In her account, that infrastructure moves decisions through an organization without meetings, status updates, or approval chains [1]. By her reasoning, an experiment that costs a few tokens and a day doesn't need a committee. The controls still exist, built into the tools people use.
Her evidence is real. In a field experiment with 776 professionals at Procter & Gamble (P&G), people who worked with AI reported more excitement, energy, and enthusiasm than people who worked without it. They also reported less anxiety and frustration. Working with AI lifted their mood about as much as working with a human teammate did, and sometimes more [2]. In a 2026 survey of tech workers by Lenny's Newsletter, 82 percent said AI was making them measurably more productive. Asked how AI had changed their view of themselves, 49 percent chose "amplified," which the survey defined as able to do more, and better [3].
Those findings come from workplaces where the authority to act is already settled. Someone in the organization can say yes, so the only open question is how fast they get to it. That describes a company far better than it describes a government agency.
Government work runs on two clocks
Government work runs on two clocks, which most writing about AI treats as one. The first is decision latency, the time between when a decision-maker has the authority to decide something and when they decide it. The second is authorization latency, the time between when a problem becomes visible and when someone is legally allowed to act on it.
In a company, the two clocks are almost always the same clock. A CEO can move budget, a product manager can ship a feature, and an engineer can push code. Delay usually comes from work nobody has finished yet. AI shortens that delay by taking on more of the work, including the coordination. The person at the gate already holds the key.
In government, where the clocks run at very different speeds, AI touches only the first one.
Statute, appropriation, delegation, and the chain of legal authority that ties each action to the public trust all limit authorization latency. An employee at the Department of Health and Human Services who sees a problem can't simply build a tool to fix it, even a tool that would work. First they have to establish several things:
- They have the authority to build it.
- Any information it collects from the public clears the Paperwork Reduction Act.
- The result is accessible under Section 508 of the Rehabilitation Act.
- Buying it followed federal competition requirements.
- Its records follow a records schedule approved by the National Archives.
- The decision to use it can survive review by an inspector general, Congress, and members of the public who file records requests.
Together these requirements form the legitimacy architecture of public work.
An agency can't move these controls into a software pipeline, because the agency didn't create them. They come from Congress, the courts, and statutes such as the Administrative Procedure Act. The controls exist because the money belongs to the public, the authority comes from statute, and people who weren't in the room can review every decision. Shortening authorization latency is a question for democratic institutions, which are designed to move slowly here.
Companies assume the authority that agencies must establish
Meyer's argument persuades because the organizations she describes, among them Block, Lovable, and Every, operate on assumed authority. When a "directly responsible individual" (DRI) at Block takes ownership of a decision, nobody questions whether the decision is legitimate. The DRI's authority comes from the org chart, which answers to the board, which answers to shareholders. Because the builder answers to the same boss as the company's governance, the company can build that governance into the tool.
That's why intrapreneurship works inside companies. An employee who takes an idea to production over a weekend is using delegated authority that everyone already agrees exists. In the worst case, a failed experiment costs a sprint. The company learns from the loss.
A failed experiment in government can end in a front-page story, an inspector general's report, a Government Accountability Office (GAO) finding, or harm to a constituent from a system deployed without proper authority. Government employees feel that risk as accountability anxiety. AI can make it worse, because a failure can now happen faster, at larger scale, before anyone checks the authorization gate.
The P&G finding inverts across this boundary. The employees in that study were building something they were already allowed to build, so their only question was whether they could build it well. In government, the first question is whether anyone may build it at all.
Many delays come from authority nobody can find
Some government delays have no basis in statute. Call them phantom gates. A large share of what looks like red tape comes from missing information.
Most government employees work without a clear map of what they personally can decide. The delegation memo from 1987 sits in a filing cabinet, on a shared drive, or in the memory of a career executive who retired in 2019. For someone who wants to stay on the right side of an audit, the rational move is to send the decision up. Escalation becomes the default because nobody can read the delegation, even where the law doesn't require it.
So a large category of government work sits in authorization queues for no legal reason. Making authority readable is work that AI does well.
Picture a system that answers four questions before any work begins. Can this be decided? By whom? Under what authority? Right now? The system maps the decisions an employee can make. It reads the delegation memos, appropriations language, program statutes, regulations, and precedent. Then it tells the employee: "You can decide this. Here is the authority. Here is the boundary. Here is what requires escalation."
Such a system would make government clearer. Where phantom gates take more time than real ones, clarity speeds up the work.
AI tools fail when they skip the authority question
AI fails in government when it brings decision-latency speed to authorization-latency problems. A system that drafts a decision memo in an hour doesn't know whether the decision could be made at all. It writes confident prose about actions that may require congressional notification, a new appropriation, a regulatory change, or a public comment period under the Administrative Procedure Act.
Years of federal AI experiments show a pattern. Agencies adopt tools that work inside clear authority boundaries, such as tools that draft responses to Freedom of Information Act (FOIA) requests, sort grant applications, or answer routine questions from constituents. They shelve tools that reach into the permission layer: systems that determine eligibility, allocate resources, or deal with the public in ways that raise due process questions. Federal policy draws the same line. In March 2024, the Office of Management and Budget told agencies to put minimum safeguards on AI that affects people's rights or safety by December 1, 2024, or stop using it [4].
Shelving those tools is a correct judgment. A tool that speeds up decisions nobody is authorized to make raises an agency's liability faster than it raises productivity.
The people who build these tools often come from the world Meyer describes. In their experience, the bottleneck has always been coordination between people, so they assume it still is. Few have worked inside an agency where the bottleneck is the distance between a program officer and the statute that created the program 40 years ago. Fewer still have sat in a room at the Centers for Medicare and Medicaid Services (CMS) trying to decide what a machine learning model that predicts hospital readmissions is. It could be a clinical decision support tool under Food and Drug Administration (FDA) jurisdiction, a program integrity measure under CMS authority, both, or neither. Agencies face questions like this one every day.
Agencies need tools that make authority readable
If companies need a coordination layer, government needs a permission layer: a way for employees to see the rules that already apply.
The most valuable AI in public service answers one question before the work starts: Who may do this work? It maps authority. It surfaces the 1987 memo. It flags the statutory boundary. It can tell a decision that needs a signature from one that needs an act of Congress.
This kind of tool looks little like the ones built for the private sector. Its output is clarity, which is hard to show in a demo. That also makes it hard to fund. It fits the way public work is built.
Agencies already keep the raw material. Many publish delegations of authority manuals that record who may decide what, under which authority. The General Services Administration, for example, keeps a Delegations of Authority Manual whose Federal Acquisition Service chapter lists the authorities that service's commissioner has delegated [5]. An AI system needs records like these to be useful. One that could read, maintain, and query them would make government legible to itself.
Legibility has effects further down the line. Employees who know what they can decide decide faster. Employees who know what requires escalation escalate more precisely. Once the phantom gates are gone, people can find their way through the real ones. The system becomes more coherent, with every safeguard still in place.
In government, participants share a structure of authority
Meyer's article cites John Vervaeke's distinction between perspectival knowing, which is experiential and embodied, and participatory knowing, which is relational and contextual. She argues that the coordination layer is the infrastructure that makes participatory knowing possible at scale. The article doesn't develop the idea far.
The idea assumes that the participants already share a context. In a company, they do. The DRI, the engineer, and the designer share a mission, a budget, and a tolerance for risk. Their participatory knowing is coordinated because their perspectives overlap.
In government, the perspectives don't overlap that way. The program officer, the inspector general, the congressional appropriator, the GAO auditor, the federal judge, and the member of the public who files a FOIA request all take part in the same system without sharing a context. They share a structure of authority. A coordination layer can't make them see the same thing. Each is supposed to see something different, so that the system can settle the differences between them.
That makes the permission layer a question of legitimacy, which works only when it's stated openly.
Conclusion
AI in government will follow a slower path than AI in companies, as institutions learn to see themselves.
The agencies that get this right will invest in mapping what they're already allowed to do, ahead of buying the most powerful models. They'll make authority legible before they automate. Of every proposed AI system, they'll ask whether anyone already has the authority to make the decisions it speeds up.
If nobody does, the tool isn't ready. Public trust sets the pace, whatever the technology can do.
Meyer is right that coordination is the bottleneck for companies. Government is a different kind of institution, one where the rules are the point. The permission layer is the foundation of public work. The most important AI tools for public service will help people see it clearly.
Notes and sources
- Vanessa Meyer, How AI-Native Companies Are Rebuilding the Management Function, October 5, 2026, in The Coordination Layer
- Fabrizio Dell'Acqua, Charles Ayoubi, Karim Lakhani, Hila Lifshitz, Raffaella Sadun, Lilach Mollick, Ethan Mollick, Yi Han, Jeff Goldman, Hari Nair, and Stewart Taub, "The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise," 2025, as summarized in Ethan Mollick, The Cybernetic Teammate, March 22, 2025
- Noam Segal and Lenny Rachitsky, How tech workers are feeling in 2026: a workforce splitting in two, July 7, 2026
- Office of Management and Budget, M-24-10: Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence, March 28, 2024
- General Services Administration, Delegations of Authority Manual (FAS), order 5450.1 FAS D, March 11, 2014