Most conversations about government AI adoption still default to a few technology questions — which model, which vendor, and which use case. New research from Adobe found that this is the wrong layer to focus on. Among 502 public sector respondents in a global study of 5,633 executives across 12 industries and 11 markets, the organizations struggling to scale AI are rarely held back by the technology itself. Instead, they face a more basic challenge: determining who has the authority to decide what happens to an idea after the pilot ends.

Run the numbers, and the pattern is stark. Out of every 100 proposed AI use cases in the public sector, 90 never reach production. They fall away one review at a time — IT evaluation, security, integration, budget — until what started as a promising idea quietly stops being anyone's priority. That's not a story about caution, red tape, or an unready workforce. It's what happens when an organization's operating model hasn't kept pace with the demands of AI — determining who owns a decision, who's in the room when it's made, and what counts as proof of success.

The middle is where government AI actually lives.

Our study finds that 62% of active public sector respondents are in the piloting or deploying stages — further along than exploring, but nowhere near expanding or transforming. That's the largest single cluster in the study, but it doesn’t signal a slowdown. Instead, it shows where the friction actually occurs. Getting a use case into the pilot stage is comparatively easy. Turning that pilot into something repeatable — a second office, a second program, or a second jurisdiction that can be adopted without starting from scratch — requires a different kind of infrastructure than the pilot needed. Not more computing power, but greater clarity about who decides what happens next.

3 gaps do most of the damage.

Underneath this “Frozen Middle” sit three specific deficits that compound each other.

1. Nobody is actually in charge. Just 7% of public sector respondents report having a formal AI owner with a defined title and budget authority. Everyone else operates under a part-time lead, a committee, a project-by-project arrangement, or no accountable person at all. A committee can flag a risk, but it usually can't resolve a funding dispute between the privacy office and the program team or decide whether a use case is worth the political capital required to scale it. That requires a name, a mandate, and a budget line.
2. IT and business functions communicate via tickets. They do not operate as partners. Among public sector respondents, 92% describe the relationship between business functions and IT as siloed, transactional, or consultative — one side hands off a request, the other side reviews it. Governance, security, and accessibility only get a real seat at the table once the design is essentially locked in. This sequencing is exactly why the approval funnel narrows so hard — the checks that could have shaped a proposal early end up being the checks that kill it late.
3. Almost no one can prove what AI actually delivered. Only 4% of public sector respondents report that they comprehensively measure the impact of AI. Usage statistics and processing time gains are easy to produce and easy to misread. They show that a tool is being used, not that a service has become better, faster, or more trustworthy for the people who rely on it. Once a workflow has shifted, reconstructing a reliable baseline against which to measure improvement is nearly impossible. Institutions spending public money need to be able to explain what that investment delivered, but most currently can't.

What the highest-performing organizations do instead.

The study also identified 342 Leaders, accounting for 6% of active respondents across 12 industries, based solely on operating model maturity and deployment depth, not by sector. They function as a useful yardstick precisely because they weren’t selected as a government benchmark. Five things distinguish them, and none of the five is a technology choice:

1. Leaders name a real owner. They are 13.8 times more likely than public sector respondents to have a formal AI owner with a title and budget authority — a single accountable point of contact for decisions that cut across functions, rather than a coordinator who can only escalate and wait.
2. Leaders measure before they build, not after. Among Leaders, 44% report comprehensive measurement, against 4% of public sector respondents. The baseline is defined at the same time as the use case — not reverse-engineered from whatever data survives deployment.
3. Leaders treat IT as a partner from day one. Among Leaders, 81% describe their business-IT relationship as collaborative or integrated, against 8% of public sector respondents. Security, privacy, and accessibility requirements get worked out while the use case is still a proposal, not after it's already been built.
4. Leaders let failure teach rather than punish. Under 1% of Leaders report destructive failure patterns — blame, cover-up, and repetition of the same mistake — compared with 56% of public sector respondents. That's a striking gap, and it's the one most directly tied to public trust. Institutions that can talk openly about what didn't work are the ones that stop making the same mistake twice.
5. Leaders put champions front and center, not on the sidelines. Among Leaders, 42% name empowering internal champions as their single most impactful culture move — ahead of formal training, executive memos, or new mandates. People who've already made something work in their own corner of the organization turn out to be more persuasive than any policy document.

Line these five behaviors up, and a pattern emerges that goes beyond any one survey. Organizations pulling ahead have stopped treating governance as a checkpoint near the end of a project. They've rebuilt it as a set of standards that travel with the work from the beginning — decision rights assigned early, evidence requirements agreed on early, and oversight partners looped in early. This shift shows up wherever institutions are being asked to move at AI's pace: rules rewritten so they can be applied at the point of decision rather than reviewed later, teams built around a mission and measurable outcomes rather than a fixed organizational chart, and feedback treated as a permanent feature of how the work runs rather than a lessons-learned memo filed away once a year.

A smaller starting point than it sounds.

None of this requires a reorganization or waiting for a bigger budget. It requires picking one use case that matters and doing four things before it goes anywhere near production: name the one person accountable for it, write down the outcome it's supposed to produce and the baseline it's starting from, get the delivery and oversight partners into the room while the design is still open to change, and agree in advance on what evidence will decide whether it expands, gets reworked, or gets shut down.

Although a small list, it’s the difference between a pilot that stays a pilot indefinitely and one that becomes the case a public institution can point to — with evidence — when it decides to do more of what worked.

Explore the full report to see where public sector organizations stand and what Leaders do differently.

https://business.adobe.com/fragments/resources/cards/thank-you-collections/government