Earlier this year, I spoke with a candidate running to lead one of America’s largest cities. The candidate shared that workers were being pulled off the front lines to do clerical and administrative work because the city couldn’t afford to hire backfill staff. The ask, like that of nearly every government leader I’ve met this year, was simple: Can we get more money?
The problem is that there is no more money. Many public sector institutions around the world are being asked to do more for constituents with budgets that can’t stretch. In the United States, for example, state general fund spending was nearly flat in enacted fiscal 2026 budgets, with states increasingly turning to targeted spending cuts and hiring freezes.
Efficiencies and capabilities supported by AI can become part of the answer, but the AI discourse, as it stands is dominated by two conversations: a use case conversation, focused on what AI can do, and a high-level responsible-AI conversation, focused on establishing principles and managing risks. Both matter, particularly the responsible AI conversation, as these systems touch more individuals’ lives. But neither directly answers the question these leaders actually face: how to create value under the operational constraints that have always defined these institutions. The questions I hear most often—from agency and ministry heads, university and health-system leaders, and the officials running public services—are not in either conversation. They are:
- How do we make the case for investments in AI when budgets are constrained?
- How do we get past vendor demos and free licenses to actually create value?
- How do we get more efficient when reducing the frontline workforce isn’t the answer?
- How do we redeploy efficiencies into services our constituents currently can’t access?
These are the questions that determine whether AI helps public sector institutions. And almost no one is answering them.
The AI value gap is wide at public sector institutions
Across the broader economy, AI adoption is soaring, but measurable results are not. McKinsey’s State of AI research shows the vast majority of organizations now use AI in at least one business function, yet only about a third have begun to scale it across the organization, and just 6 percent qualify as “high performers” capturing meaningful bottom-line impact.
While the sample of governments and public sector institutions in the research was small, it confirms what my conversations with leaders of these organizations suggest: The paradox is just as stark. And the obvious question, as the next wave of agentic AI arrives, is: Why would the result be any different?
This isn’t because public sector leaders are short on AI activity. NASCIO has named “pilot paralysis” as the central challenge for state chief information officers (CIOs) trying to move generative AI from controlled experiments into broader service delivery. And the pressure is visible in their own priorities: In 2026, for the first time in the list’s 20-year history, US state CIOs ranked AI as their number-one priority, ending cybersecurity’s 12-year hold on the top spot. Budget and cost control, meanwhile, jumped from sixth to third. The ambition and the constraint are rising together and colliding.
But that’s the symptom rather than the story. When IT priorities reshuffle this dramatically, it’s because the underlying problem has changed. AI in public sector institutions has risen from a technology decision to one that an agency leader must make. And the playbook on their desk is not written for their reality. It is written for institutions with adaptable budgets, flexible workforces, and customers they can choose to serve. Public sector institutions have none of those.
Four operating constraints that define the work
The following four operating constraints define the public sector.
- No new money. Operating budgets are flat or shrinking, and there is no separate “AI transformation budget” arriving from the sky. Any meaningful investment in AI capability will typically need to be found from what is already there. So, in a public sector institution, the case for it clears a higher bar than it does in a growing enterprise.
-
Workforce capacity is constrained. This is where the conversation has gone a bit off track. In the private sector, using AI sometimes translates into “doing more with fewer people,” and a number of companies have attributed layoffs to AI.
For most public agencies, health systems, and universities, reducing frontline staff isn’t on the table. Many already struggle to recruit the talent they desperately need. In US healthcare, the federal government projects a shortage of roughly 141,000 physicians by 2038, along with nursing shortages concentrated in rural and underserved areas that can least absorb them. The labor question in these institutions is not fewer people doing the same work. It is the same people freed to serve more people by streamlining their work and expanding their capabilities.
- Public scrutiny is constant. Every decision affecting a citizen, patient, or student is one news cycle, one records request, or one legislative hearing away from public examination. The standards for explainability, auditability, and equity can’t be compliance afterthoughts. They must be built into what counts as a working system.
- The mission is nonnegotiable. A private firm can deprioritize a customer segment that costs too much to serve. The safety net health service cannot turn away the patients who arrive at its doors. A public benefits agency cannot decline the costly-to-serve. AI that performs well on average but fails the people the institution exists to serve is a failure.
What a different operating model looks like
The good news is that there are already examples of what works. And the most instructive are those that go beyond simply handing employees a chatbot.
Singapore’s Government Technology Agency (GovTech) offers a useful baseline. It built Pair, a secure generative AI copilot for public officers, alongside AIBots, a low-code platform that lets officers create their own task-specific assistants drawing on their agency’s knowledge bases. Within two months of rollout, more than 11,000 officers across more than 100 agencies were using it. Today, roughly 80 percent of Singapore’s 150,000 public officers have used Pair, and thousands of custom assistants automate specific internal workflows. That is real, and at this scale impressive, but it is step one: Give capable people capable tools, and they get time back.
A harder and more valuable step two embeds AI in a workflow that delivers a service. Consider professional licensing: a high-volume, high-stakes process where applicants often wait weeks for a decision. Working with a healthcare regulator in Asia, we built an agent-assisted licensing workflow for physician applications. Licensing rules were codified into a structured rulebook, the agent validated applications and flagged issues up front, and human reviewers retained control over every consequential decision, supported by explainable recommendations rather than a black box.
The core philosophy was not to replace expert judgment but to let the system handle the roughly 80 percent of the work governed by deterministic rules so reviewers could concentrate on the 20 percent that genuinely requires it. In proof-of-concept testing, the redesign indicated a 40 to 45 percent reduction in applications sent back for correction, about half as many review rounds, and roughly 12 fewer days per review cycle. The work is early, but the difference in kind is the point. Singapore put tools in officers’ hands at scale. The redesign of the licensing process changed what an applicant actually experiences.
That logic—AI embedded in a workflow, measured by what changes for the citizen—is the operating model. It rests on three principles I’d offer as a starting point:
- Redeploy, don’t reduce. The goal of AI in public-serving institutions is not to shrink the workforce, but to free an already-stretched workforce from the work that drains it and redirect them toward what they were hired to do. Officers back on patrol. Nurses back at the bedside. Caseworkers back with families. Faculty back in the classroom. The licensing redesign did exactly this. Rather than cut reviewers, it let the system clear the routine, rules-based 80 percent of the work so expert reviewers could focus on the complex cases that genuinely need human judgment, improving both throughput and decision quality. In institutions that are short-staffed to begin with, the dividend from AI is more from the people you already have rather than fewer people.
- Invest creatively. When the operating budget is flat, capital for AI transformation has to be assembled from elsewhere: reallocation within existing spend, modernization and innovation funds, shared-services partnerships, and financing structures that pay the investment back through the value it creates. The shift that unlocks this is treating AI as a capital strategy question—owned at the top, funded like an investment with a return—rather than an IT line item. Getting that funding model right is its own discipline, and one worth a fuller treatment that I’ll address in a separate article.
- Measure citizen outcomes rather than deployment milestones. A model in production is an engineering achievement. A model that has positively changed what a citizen, patient, or student actually experiences is an operational one. The institutions that get this right define success at the citizen level—applications resolved without rework, review rounds cut, wait times shortened, access widened. In the licensing pilot, success looked like 40 to 45 percent fewer send-backs and a faster decision for the applicant, not the number of seat licenses issued or tokens consumed, which says nothing about AI’s impact on the people public institutions serve.
The art of the possible
AI in public sector institutions is not just about efficiency. It’s a story about ambition.
Every leader I talk to has a list of things their institution cannot currently afford to do. Childcare investments a government wants but can’t fund. Preventive care that a health system wants to deliver but lacks the staff to do so. Student services at the university that were promised but had to be scaled back. AI’s promise, properly captured, is the capacity to do the things these institutions exist to do and currently cannot.
That is the conversation I want us to be having. The one that starts not with “what AI can do for us” but with “what could we do for the people we serve if AI freed us to do it?” The specifics differ by country and level of government—a national ministry, a state agency, a city, a public university—but the shape of the problem and the opportunity is strikingly consistent.
Public sector institutions are operating under the hardest version of the AI problem. They also have the most to gain by solving it. The playbook that emerges from that solving will be one of the more consequential management contributions of this decade. And the leaders writing it now are the ones who decide to ask the harder questions.
This article was originally published on LinkedIn on June 30, 2026.