In a previous blog post, I asked how public sector institutions can become more efficient by using AI and how they can redeploy those efficiencies to deliver services their constituents currently cannot access. My answer to both questions: it starts by rejecting the premise that AI’s contribution here is measured in hours saved.
This is what makes AI’s promise in the public sector different from the version told everywhere else. The instinct is to look for freed-up hours to redirect—a caseworker’s afternoon opening up once the paperwork is automated. But in the public sector, that afternoon was never truly empty. There’s often another step to take, whether that’s calling a parent about the other programs their children qualify for, getting to the license application that has been sitting in the queue for weeks, or pushing forward the eligibility determination that’s stuck in another round of review. That delayed or undone work falls hardest on the people least able to absorb it, such as the single parent working two jobs who needed the benefit approved weeks ago.
So the measure that matters in public institutions is reach: the capacity to get to the work that often goes undone. Reach comes from raising the ceiling on what workers are able to do. There are two ways to raise that ceiling, and a third move without which neither of the others holds.
Move 1: Redesign the work
Reaching that undone work starts with a choice about how the work itself is built. Hand a caseworker an AI tool and the paperwork gets faster, but the job is unchanged—the ceiling hasn’t moved, and nothing new gets reached. The gain comes only when the work is redesigned, with the routine, rules-bound parts handed to a reliable AI-driven system so the staff member can spend their time on what only a person can do.
Consider an eligibility worker. Much of the role today is verification: reading through the documents an applicant submits and checking them against the rules to determine whether they qualify. It’s necessary work, and it consumes the day. Push aspects of that verification to a reliable and responsibly designed AI system, with human review and approval, and shift the worker’s capacity to another part of the role. They could become more of a counselor—someone who sits with a family and asks the questions no one had time to ask before. You qualify for this benefit, but you have young children. Have you considered the childcare program you haven’t applied for? That counseling was always part of delivering on the public sector’s mission. There just wasn’t the time to do as much of it as the eligibility worker wished.
This is the difference between automating a task and redesigning the work, and it’s where the value lives. The McKinsey Global Institute research is specific on this point. Currently demonstrated technology could theoretically automate 57 percent of US work hours by 2030, but the midpoint scenario has only about 27 percent of hours actually automated once costs and practicality are accounted for. That’s why organizations that treat AI as an add-on to the old way of working tend to see gains so modest they never show up in the results—a point confirmed by McKinsey’s latest State of AI research that shows the companies seeing the biggest gains from AI are far more likely than others to redesign workflows to incorporate the technology. The goal is a different process, one that delivers a quick, accurate result with little to no rework required and puts people where their judgment matters.
Move 2: Accelerate workers’ expertise
Move 1 assumes the capacity of the worker freed up by redesigned work already has the knowledge to take the next step toward delivering on an agency’s mission. That assumption doesn’t always hold. Knowing how to verify a document against a rule doesn’t teach someone the landscape of adjacent programs a family might qualify for—that’s a different kind of knowledge and redesigning the job doesn’t automatically hand it to someone. AI can help close that gap.
In practice, that might mean an AI system flagging that a family with young children is likely eligible for a childcare program, or surfacing the exception that applies to a specific case type—the kind of thing a worker would otherwise only know from having seen it before. The worker still decides what to do with it. The system just puts the relevant knowledge in front of them at the point they need it, instead of requiring they already carry it.
Move 3: Make the transition explicit
Moves 1 and 2 only work if the people being asked to do them trust what’s behind them. Many leaders assume the workforce’s worry is that AI is a step toward needing fewer people rather than more capable ones. But that’s not exactly what the data shows.
A MissionSquare survey of 2,000 state and local employees found that only 20 percent were very or extremely concerned about AI replacing their job function. It also found that only 28 percent had received any AI training from their employer. Read together, those numbers show the issue is less about AI wariness and more about a lack of structural support to enable these shifts. Ongoing role-based training can help address both hesitancy and a lack of AI skills while improving outcomes for the institution and the public it serves. That requires something of the institution—funding and leadership accountability.
This does not mean overselling. Some roles really do shrink. But complete elimination is rare. Research by the McKinsey Global Institute finds that more than 70 percent of the skills employers ask for are used in both automatable and non-automatable work, which is why most roles are reconfigured rather than removed. Meanwhile, roles that barely exist in government today, from agent orchestration to AI audit and validation, will likely get created and filled. Leaders don’t need to promise that nothing will change. What they can credibly offer is that the institution will invest in training people and helping them move into the work that opens up.
What this support does is make worker elevation explicit. And it gives staff a clearer reason to engage with the tools in front of them: workers who believe the tool is there to lift them will use it that way, reaching further into the work the institution couldn’t get to before.
In conclusion
The measure of AI in a public institution is not hours saved or systems deployed. It’s whether the license application that sat untouched for six weeks gets picked up, whether the family gets the call about the program their children qualify for, and whether the eligibility determination holds up without a third pair of eyes checking it.
So here’s the question I keep asking public sector agency leaders: If AI handed your team back a fifth of its week tomorrow, which undone work should absorb that capacity first—and is your agency prepared to redirect it there?
This article was originally published on LinkedIn on September 21, 2026.
Hrishika Vuppala is a senior partner in McKinsey’s Southern California office.
The author would like to thank Megan McConnell and Bridget Sullivan for their contributions to this article.