The Pentagon Learned to Transform Itself with AI. Companies Keep Repeating Its Previous Mistakes
There is a data point that should make any executive who approved an artificial intelligence budget over the last two years deeply uncomfortable: the United States, the country that builds the most powerful models in the world, ranks 24th in global AI adoption. Its rate is 28.3%. Singapore stands at 61%. The United Arab Emirates, at 54%. Goldman Sachs recorded that investment in AI contributed "practically zero" to U.S. GDP growth during 2025. The problem is not technological. It never was.
Drew Cukor knows this better than almost anyone. As a retired Marine colonel and founder of Project Maven — the program through which the Department of Defense integrated AI into its most complex operational workflows — Cukor built from the inside what many considered impossible: demonstrating that commercial software could function within the largest bureaucracy on the planet and produce results that previous systems could not achieve. Today, from his firm TWG AI, he watches as American companies commit, point by point, the same mistakes the Pentagon almost made before Maven worked.
His diagnosis, published in Fortune on May 11, 2026, coincides with a singular moment: two days later, the Department of Defense announced agreements with eight technology companies — Amazon Web Services, Google, Microsoft, OpenAI, SpaceX, NVIDIA, Reflection, and Oracle — to deploy frontier AI capabilities across its classified networks at the IL6 and IL7 levels. The Pentagon is not debating whether to adopt AI. It is executing.
The Gap That Separates Building from Integrating
The Stanford 2026 Index on artificial intelligence does not describe a research failure. It describes an implementation failure. The United States leads in benchmarks, in investment in models, and in computing capacity. But that advantage does not translate into operational use because the organizations that should be deploying this technology have not changed their internal architecture to receive it.
Cukor introduces here a distinction that is worth more than the majority of strategic frameworks circulating at C-suite conferences: the difference between an AI-augmented organization and an AI-native organization. This is not semantics. It is the difference between bolting a new engine onto an old structure and rebuilding the structure from the design stage with that engine in mind.
When the Pentagon launched Maven, it did not treat it as a technological experiment or as a limited pilot. Senior leaders took personal ownership of it and fought for it within the bureaucracy. Workflows were dismantled — software was not simply layered on top of them. The only metric that mattered was operational: what could soldiers do that they previously could not do. That discipline is what made it work. And that discipline is precisely what is missing from the vast majority of corporate AI programs that exist today.
China's "AI Plus" initiative operates with a logic structurally similar to that of Maven, although from the opposite end of the political spectrum. Beijing is not building models to compete in rankings. It is embedding AI into manufacturing, logistics, scientific research, healthcare, and education with specific industrial datasets and agents designed for concrete workflows. It does not debate control or containment. It deploys. That difference in speed between the United States' power of invention and China's capacity for integration is the gap that Cukor identifies as the central competitive risk of this decade.
Three Failures the Pentagon Barely Survived
Cukor identifies with precision the three mistakes that turned Maven into a possible transformation — because they nearly prevented it from happening at all.
The first is delegation without ownership. In too many corporations, AI strategy is delegated to a Chief AI Officer or an innovation lab. Those structures are designed — though no one admits it — to signal progress without threatening the existing distribution of power. The result is what Cukor calls "pilot purgatory": projects that never die because no one killed them, but that never scale either because no one truly pushed them forward. Maven worked because senior leaders did not delegate ownership of the problem. They assumed it themselves.
The second mistake is layering AI on top of legacy processes. There is a way of using AI that guarantees mediocre results: take the current workflow and add a model to it. Efficiency gains will be marginal because the structure generating the bottleneck remains intact. If, after launching an AI initiative, the organizational chart, the approval chains, and the operational rhythm of the company are the same as before, no transformation has occurred. What has occurred is decoration.
The third mistake is measuring activity instead of outcomes. Models trained, proofs of concept completed, partnerships announced: these are indicators of movement, not of impact. Maven was measured by what operatives could do that they previously could not do. That is the only question that matters in an AI program that claims to change something.
These three mistakes are not engineering accidents. They are governance accidents. They arise from organizations that want the image of transformation without assuming the internal political cost that dismantling what already exists entails.
What the May 13th Announcement Reveals About Real Execution
The agreement that the Department of Defense announced on May 13, 2026, with eight technology companies to operate across classified IL6 and IL7 networks is not simply news about contracts. It is a signal about what kind of institution can scale AI under conditions of maximum complexity and restriction.
The fact that the Pentagon — historically known for its software acquisition failures — managed to articulate a functional AI platform in maximum-classification environments is, in itself, a case study in structural maturity. Not in technology. The Department of Defense's CTO unified the process under an enterprise alignment structure, as analyses of the program describe. That means someone made the decision to break down the silos that historically prevented the Pentagon's technological systems from functioning coherently.
The lesson that movement offers for the private sector is an uncomfortable one: if a bureaucracy of that scale and that historical rigidity was capable of reorganizing itself to integrate AI operationally, then the most honest explanation for why private companies fail to do the same is not a lack of resources or a lack of talent. It is a lack of willingness to assume the internal political cost of changing who decides what, how, and at what speed.
Cukor's analysis of the difference between companies that run AI experiments and companies that run AI transformations points precisely to that issue. It is not a budget problem. It is a question of who at the top is willing to put their name on the intermediate failure that precedes any deep transformation.
Structural Maturity Is Not What the System Claims to Have
There is a specific fragility that appears before the problem becomes visible. Organizations that are going to fail at their AI transformation do not announce it. They have roadmaps, they have labs, they have Chief AI Officers with budgets. But if one looks closely at the decision-making chain, the pattern emerges: the difficult decisions — those that involve eliminating processes or changing power structures — never actually get made. They are postponed, piloted, studied. And the system keeps functioning, with an appearance of modernization, until the competitors who did make those decisions render the cost of inertia impossible to ignore.
Cukor describes this as a "white-collar reckoning" that could be worse than the wave of industrial offshoring in the 1970s, but faster and less forgivable. Not because AI is inevitably substitutive of employment, but because companies that do not reorganize their structures of decision-making, approval, and operation will find themselves in a position of cumulative disadvantage relative to competitors who did — whether those competitors are Western or come from Asian economies that have been treating AI as operational infrastructure from the outset.
The difference between the Pentagon before Maven and the Pentagon after it is not that it now has better models. It is that it learned to sustain a transformation with executive ownership, with the dismantling of what previously existed, and with metrics that measured results. That learning took years to consolidate and came close to collapsing several times. Private companies do not have the same tolerance for time, but neither do they have the same excuse for ignoring the pattern.
A system that appears strong because it possesses advanced technology, but that has not touched its decision-making structure or its approval chain, is not a transformed system. It is a system that bought itself time. How much time depends on how long it takes its competitors to do what it has still not done.
