The replacement story is sabotaging government's AI rollout

Federal agencies reported more than 3,600 AI use cases last year, a roughly 70% jump over the year before and more than six times the count from 2023. Billions are committed across civilian and defense budgets, with the Pentagon alone requesting north of $13 billion for AI and autonomous systems as a standalone category. By every conventional measure of adoption, government is moving fast.

I've written before that these numbers measure throughput, not value, and that a Government Accountability Office review of four major agencies found a consistent pattern: tools scaling faster than the human readiness to use them well, with lessons learned in isolation and never shared. That argument stands.

But I want to make a different one here. The problem is not only that we are measuring the wrong things. It's that the story we're telling about why we're adopting AI is quietly guaranteeing that adoption fails.

That story is the replacement story: the premise, imported wholesale from Silicon Valley and now wearing a federal efficiency badge, that AI's purpose is to let the government do its work with fewer people. It is not a fringe view.

In early 2025, a U.S. official monitoring the government's cost-cutting effort told The Washington Post that the goal was to replace "the human workforce with machines," that "everything that can be machine-automated will be," and that "the technocrats will replace the bureaucrats." Whatever you think of that as policy, as a way to actually get good outcomes out of AI, it's self-defeating.

Here's the problem. GenAI models render output that sits in the statistical center of everything they were trained on. Used well, that consistency helps build processes and synthesize knowledge. But government work often requires differentiation. Circumstances change, facts vary, and public servants must exercise judgment to produce original solutions. When GenAI substitutes for that judgment, it pulls memos, briefings, and options papers toward the same defensible middle.

The government earns its legitimacy precisely in the cases that don't fit the model: the pandemic with no precedent, the cyber crisis at two in the morning, the unexpected emergency. Those are exactly the moments where statistical averages fail.

Now layer the replacement story on top.

An employee who believes the institution is looking for reasons to need fewer people faces a quiet calculation every day. The AI surfaces a confident, plausible answer. Overriding it means more work, more visible risk, and asserting judgment the organization has implied it no longer values. Deferring to the machine is faster and safer. The fear doesn't make people quit. It makes them stop arguing with the model.

That deference is the failure. GAO's own science and technology arm has noted that even the best-performing AI agents complete only about 30% of complex tasks autonomously without error. The other 70% still depends on a human catching mistakes before they affect someone's benefits, security clearance, or enforcement action. The replacement story is, functionally, an instruction to stop catching them.

This is the part many leaders miss. They treat workforce anxiety as a morale issue when it's really a performance issue. Every responsible AI framework, including the federal government's own guidance on high-impact uses, depends on humans remaining meaningfully in the loop. You cannot tell people their judgment is a cost to eliminate and then expect them to exercise that judgment rigorously against the very tool replacing them.

The same dynamic is reshaping government contracting.

For decades, contractors competed on reliable execution at scale. AI is collapsing the price of exactly that. When agencies can generate competent first-pass analyses from their own subscriptions, contractors whose work product is indistinguishable from that output have a problem they can't price their way out of.

We're already seeing that pressure arrive. The General Services Administration has been auditing agreements with major firms and pressing them to defend their value beyond speed and cost. Read those audits correctly and they ask one question: what are we paying you for that we cannot now get cheaper somewhere else?

A contractor that has internalized the replacement story has no good answer. If your pitch is doing the same work with fewer people, you've conceded that your value is execution - and execution is exactly what AI is commoditizing. The firms that will win the next decade won't be those that automate the most bodies off contracts. They'll be the ones that can demonstrate what I describe as Original Intelligence: human judgment that departs from what the model would have generated. AI plus that human signal is a defensible value proposition. AI as a headcount-reduction strategy is a race to the bottom.

So what should leaders do?

First, change AI's stated purpose. The goal isn't fewer people producing the same average work. It's the same people producing work the machine cannot while AI carries the genuine drudgery.

Second, measure what is actually at risk. Most AI training teaches people to operate the tools. That's necessary but insufficient. Leaders should measure whether AI strengthens or erodes people's ability to contribute originality, synthesis, and judgment over time.

Third, treat public trust as the real product. Citizens accept government decisions because they believe competent, accountable humans stand behind them. AI that produces faster decisions no one can explain or appeal does not build trust - it spends it.

Ultimately, government's AI transformation won't be judged by the technology itself. It will be judged by which agencies and contractors preserved human originality while using AI as a tool rather than a substitute.

The story we tell today about AI adoption will largely decide that outcome.