The best AI models in the world can pass a bar exam, draft a merger memo, and solve math problems that have stumped leading mathematicians for years. So it’s reasonable to assume they’re ready to take administrative work off of the plates of overstretched federal employees.
Yet somehow, government AI adoption remains remarkably low. Nearly 60% of reported federal AI use cases with available deployment information are still in pilot or pre-deployment, according to an April 2026 Brookings analysis. At the IRS, 61% of the agency’s 126 active AI use cases remained in development, according to a March 2026 GAO review. Enthusiasm is not lacking: Everyone from the White House to the General Services Administration has called for more AI in repetitive government workflows, and the desire to save taxpayer money is real.
Why is this a square peg, round hole problem?
In developing AI, the frontier labs have primarily tested for isolated intelligence. Today’s best models are extraordinarily strong at generating and checking patterns in text, code, and formal reasoning. But they haven’t optimized for the real workflows in government, including the mundane ones such as long chains of small, high-consequence actions where one missed field, wrong attachment, or bad click can invalidate the whole result. Things like a procurement to award, a budget reprogramming, or a contract closeout. These processes run 40, 50, or 60 dependent steps, each feeding the next, tapping multiple offices, and requiring oversight at every gate.
If you talk to most federal employees who have actually piloted AI, they’ll tell you it was a lot of noise and hard to trust. That’s because they’ve probably tried a version of ChatGPT or Claude, lightly optimized for their workflows. AI pilots fail when these models collide with the government’s real operating environment: fragmented data, legacy systems, procurement and security constraints, high legal stakes, and no clear path from experiment to accountable production service. As a former member of the federal workforce, I’ve seen enough of this world to know the pattern. Federal employees are willing to try new systems and have good ideas for improvement. They are short on tools that actually work. We cannot afford to have the government be five-to-10 years behind the state of the art.
This discipline already exists in the vertical AI movement, which has become prolific in the private sector. Consider Harvey, a domain-specific AI platform for legal and professional-services teams that supports work such as legal research, contract analysis, due diligence, and compliance on a secure platform; Candidly, an AI-native wealth management platform accessed via employers; or Abridge, a healthcare AI platform that converts patient-clinician conversations into evidence-linked clinical documentation integrated with electronic health-record workflows and subject to clinician review. Each of these selected a domain and went deep. None tried to be general-purpose AI.
Rather than generating plausible answers to open-ended prompts, these vertical AI solutions execute defined, end-to-end workflows with embedded domain expertise, guardrails, and accountability. They pair probabilistic models (which generate the most likely output from patterns in data) with deterministic workflows (which follow predefined rules to produce consistent, verifiable results). The result is a reliable system that ensures each result can be checked and reproduced.
The challenge, of course, is that “government” isn’t a single vertical — certainly not in the way that HR, legal, or tax are. The government operates all of these functions and hundreds more, each with its own rules, data, and stakes. Building across that many verticals is, by definition, a horizontal problem, one where the requisite breadth tends to come at the cost of depth, guardrails, and accountability.
The opportunity here is refusing the tradeoff. Government agencies must build with the discipline of vertical AI, but at the horizontal scale government actually requires.
And they must do so with full awareness of the cultural blockers at stake. Government is both the ecosystem best positioned to benefit from AI, and the one least conditioned to demand its efficient adoption. Its technology ecosystem is a patchwork of decades-old legacy systems, disconnected databases, paper-based processes, and vendor-built tools that do not share data or workflow standards. Its employees are fully accustomed to re-entering information, reconciling conflicting records, navigating duplicative approval chains, and relying on manual workarounds simply to complete routine tasks.
The supply side has historically taken advantage of this tolerance for inefficiency. For the past two decades, commoditized commercial solutions wrapped in contract vehicles have made millionaires around the Beltway. In the AI era, we risk the same. Companies selling ChatGPT wrappers are everywhere. Too many will promise transformation while adding another disconnected interface atop the same broken processes — one more dashboard, license, and layer of abstraction for already overburdened public servants to manage.
The job of our federal executives is to demand a better solution: Secure, mission-specific systems that integrate with legacy infrastructure, turn policy into auditable workflows, and demonstrably reduce the burden on the people our government serves. This is all possible today, and whether it comes to fruition is simply a matter of our willingness to admit we deserve better.
We’re standing at the precipice of a generational opportunity to get the next generation of technology-enabled government right. We can’t afford to waste it.
Alex Cohen is a serial govtech founder who previously served as an acquisition program manager for the Bureau of Indian Affairs.