90% of Government AI Projects Never Reach Production
Nine out of 10 proposed public-sector AI projects never make it into production, according to new research from Adobe—a finding that points to an operational problem small businesses may recognize in their own organizations: promising pilots often stall because nobody has clear authority to move them forward.
Adobe’s study of 5,633 executives across 12 industries and 11 markets included 502 public-sector respondents. It found that government organizations are not primarily struggling to choose an AI model or vendor. Instead, many lack a clear process for deciding who owns an initiative after the pilot stage, how success will be measured and which teams must approve expansion.
“Out of every 100 proposed AI use cases in the public sector, 90 never reach production,” Adobe reported.
The finding has implications beyond government. Small businesses developing AI tools for public agencies—or adopting AI internally—could face similar delays if leadership, information technology, security, compliance and frontline teams operate separately. A pilot may demonstrate technical potential, but scaling it requires decisions about budgets, risk, data, training and measurable business results.
Key takeaways
- Only 7% of public-sector respondents said their organization has a formal AI owner with a defined title and budget authority.
- Ninety-two percent described the relationship between business functions and IT as siloed, transactional or consultative.
- Just 4% said their organizations comprehensively measure AI’s impact.
- Sixty-two percent of active public-sector respondents remain in the piloting or deploying stages.
- Adobe identified five practices among higher-performing organizations: assigning ownership, measuring outcomes early, involving IT from the start, learning from failure and empowering internal champions.
The concentration of organizations in the pilot and deployment stages represents what Adobe calls the “Frozen Middle.” These organizations have moved beyond experimentation but have not reached the expansion or transformation phase. A pilot may work in one department, office or program, yet fail to spread because the organization has not established a repeatable process for approving and funding the next step.
For a small company, the lesson is practical. An owner or manager does not necessarily need to establish a large AI department. Instead, the business can assign one person responsibility for a specific project, define the expected result and decide in advance what evidence would justify continuing.
For example, a small government contractor testing an AI system to summarize case files could measure processing time, error rates, staff review requirements and customer satisfaction before deployment. If the tool reduces administrative work but introduces unacceptable accuracy or privacy concerns, the company can rework or stop the project based on agreed criteria rather than momentum.
Adobe’s research suggests that successful organizations set those criteria before building. Among the high-performing “Leaders” identified in the study, 44% reported comprehensive measurement, compared with 4% of public-sector respondents. The Leaders were selected based on operating-model maturity and deployment depth, not because they belonged to a particular industry.
The research also found that Leaders were 13.8 times more likely than public-sector respondents to have a formal AI owner with a title and budget authority. That role matters because committees can identify risks but may lack the power to resolve disputes over funding, privacy, security or operational priorities.
Small businesses may not have the resources to create a dedicated executive position. They can still name an accountable project owner with authority to coordinate decisions and escalate issues. Without that person, AI initiatives can become everyone’s responsibility—and therefore no one’s priority.
The timing of collaboration also matters. Adobe found that 81% of Leaders described the relationship between business and IT as collaborative or integrated, compared with 8% of public-sector respondents. Security, privacy and accessibility teams participate early in the Leaders’ process, while siloed organizations often consult those groups only after a system has been designed.
That late involvement can create expensive setbacks for small businesses. A company may spend weeks configuring a tool before discovering that customer data cannot be used as planned, that records need additional safeguards or that the system does not meet accessibility requirements for a public-sector client. Bringing those stakeholders into the initial design discussion can slow the beginning of a project while shortening the approval process later.
The study also raises a cultural issue. Less than 1% of Leaders reported destructive failure patterns such as blame, cover-ups or repeating the same mistake. Among public-sector respondents, 56% reported those patterns. For smaller companies, a pilot that does not deliver the expected result can still provide valuable information—if employees can report problems without fearing punishment.
Internal champions can help. Adobe said 42% of Leaders identified empowering employees who have already made an AI project work as their most impactful cultural move. Those employees can show colleagues how a tool fits into everyday work more effectively than a broad policy memo or mandatory training session.
There are limits to applying the research directly to small businesses. The survey focused on executives, and its public-sector findings may not represent every agency, local government or company serving the public sector. Small firms also face different constraints, including limited budgets, fewer technical staff and less leverage when vendors control data or product road maps.
Still, the recommended starting point is modest. Before taking one AI use case into production, a business can identify one accountable person, document the intended outcome and baseline, involve technical and oversight partners while the design can still change, and establish evidence for expanding, revising or ending the project.
For companies selling to government, those practices may also strengthen proposals and implementation plans. Agencies increasingly need vendors that can explain not only what an AI system does, but how its performance will be monitored, how risks will be handled and who will make decisions when conditions change.
Adobe’s full research is available in its State of AI Readiness in the Public Sector report. For small businesses, the central message is straightforward: AI adoption may depend less on finding the perfect tool than on deciding who owns the work, how results will be proven and what happens after the pilot ends.
Image via Google Gemini
This article, “90% of Government AI Projects Never Reach Production” was first published on Small Business Trends
