Sunday, 6 September 2026
Artificial intelligence in government is often discussed in terms of surveillance, regulation and national security, but some of the most immediate applications are considerably less dramatic. Across US federal and state government, AI is increasingly being tested against a more familiar problem: ageing technology, administrative backlogs and public employees spending large amounts of time processing documents, reviewing forms and moving information between systems that were never designed to communicate with one another.
At AI4 2026, officials from the US Department of Homeland Security and the states of Utah and Indiana described how AI is beginning to move beyond isolated government experiments and into everyday public administration. Their examples ranged from contract analysis and software development to licensing, Freedom of Information Act requests and interactions between citizens and government agencies. The discussion, moderated by Reuters White House correspondent Jacob Bogage, also revealed a significant shift in how public-sector technology leaders are thinking about artificial intelligence. Rather than viewing AI primarily as a way of reducing government employment, the officials repeatedly described it as a mechanism for removing repetitive administrative work while keeping people responsible for consequential decisions.
That distinction could determine whether AI becomes genuinely useful across government. Public administrations face a combination of pressures that make automation attractive. Many operate legacy technology dating back decades, while the number and complexity of services they provide have expanded. Government employees must comply with extensive privacy, security, procurement and transparency requirements, yet citizens increasingly expect digital services comparable with those available from banks, retailers and technology platforms.
Indiana provides one example of how these pressures are changing technology investment. Rob Falk, chief information officer within the Indiana Secretary of State’s office, described an extensive modernisation programme covering business services, securities regulation, elections and the regulation of motor-vehicle dealers and manufacturers. One of the initial challenges was understanding legacy systems whose original documentation and institutional knowledge had disappeared over time. Rather than manually reconstructing every business rule, Indiana used AI as part of the process of analysing older systems, identifying their logic and helping developers rebuild applications on newer technology.
AI was subsequently introduced elsewhere in the development cycle, including software generation and testing. The state is also incorporating generative AI into public-facing services, including systems intended to guide users through licensing procedures and help review documentation. Behind those interfaces, AI can perform some of the preliminary administrative work previously handled manually. Falk said the changes have reduced workload in some back-office processes by approximately 60–70%. That figure represents Indiana’s reported experience with particular implementations rather than an independently established productivity benchmark, but it illustrates the scale of efficiency that government technology leaders believe may be available.
The more important point is what happens to the time that is released. Indiana’s strategy is not based on automatically reducing headcount as administrative work disappears. Instead, employees can spend more time dealing directly with complicated cases and citizens who need assistance. That could represent one of AI’s more important effects on public administration. Government digitisation has historically attempted to move people away from human interaction by placing services online. AI potentially creates the opposite possibility: automating enough repetitive administrative work that government employees have more time for situations where human interaction is actually valuable.
The US Department of Homeland Security faces the same challenge on a much larger scale. Roman Jankowski, DHS Chief Privacy Officer and Chief Freedom of Information Act Officer, described administrative processes in which information moves between systems that were never designed to work together. Freedom of Information Act processing is a particularly demanding example. DHS handles an exceptionally large volume of requests, including cases involving immigration records and interactions with federal agencies. Some requests require individuals to provide identifying information that must then be processed through separate systems before relevant records can be located.
Historically, parts of this process involved physical documents and repeated scanning. Even after information began arriving electronically, administrative procedures could still require employees to transfer it manually between systems. AI and automation provide an opportunity to remove some of these intermediate steps and accelerate the processing of requests. The objective is not to allow an algorithm to decide independently what sensitive government information should be released. Decisions involving privacy, national security, law-enforcement information and legally protected data remain subject to human review and statutory requirements. Instead, AI can reduce the mechanical work surrounding those decisions.
This division between processing and responsibility appeared repeatedly throughout the panel. AI can retrieve information, compare documents, identify discrepancies and recommend actions. Humans remain accountable for consequential decisions. Utah is pursuing a similar approach through its Division of Technology Services. The state has established a broader AI programme alongside an Office of Artificial Intelligence Policy and has adopted an explicitly people-centred strategy in which technology is intended to strengthen human capability rather than automatically replace workers.
One of Utah’s most interesting examples involves government contracts. State agencies process large numbers of agreements containing clauses that must comply with legislation, administrative rules and internal policies. Reviewing those documents manually can require significant amounts of specialist time. Officials initially experimented with a general-purpose AI assistant but found that it could not reliably handle the complexity of the task. The problem was not simply understanding the contract. The system needed to compare different sections against numerous statutes, policies and administrative requirements.
Utah subsequently developed a more specialised architecture involving several AI agents, each responsible for analysing different elements of a contract. According to Christian Napier, Director of AI at Utah’s Division of Technology Services, the resulting system can perform its analysis in around ten minutes for some documents and has reduced the time required from contract analysts by approximately 75%. Again, the percentage is a result reported by the state team rather than an independent assessment. Nevertheless, the example demonstrates an important lesson for organisations adopting AI: a general chatbot and a purpose-built AI workflow are not the same thing.
Early disappointment with artificial intelligence may sometimes reflect the way the technology has been deployed rather than a fundamental limitation of AI. Asking a general-purpose assistant to perform a highly specialised regulatory task without the necessary architecture, data and controls can produce unreliable results. Breaking the problem into defined components and designing the system around the actual workflow can generate very different outcomes.
This is likely to become increasingly important as governments move from experimentation towards production. The first phase of generative AI encouraged employees to use broad conversational tools for many different purposes. The next phase is likely to involve narrower systems designed around particular administrative functions.
Government technology departments are simultaneously reconsidering what they should build themselves and what they should purchase from technology companies. None of the officials argued that governments should develop their own frontier AI models. The enormous investment already being made by private technology companies makes that economically difficult to justify for most public authorities. Instead, governments can use commercially available cloud infrastructure and models while concentrating their internal resources on the business processes unique to government. That means understanding regulations, workflows, citizen requirements and the restrictions governing public data, then building applications around those requirements.
The result is likely to be a hybrid public-sector technology market. Large cloud and AI companies provide computing infrastructure and foundational technology, specialist vendors provide particular applications, systems integrators help connect them, and government technology teams develop or customise the workflows that are specific to their agencies.
For technology suppliers, this changes the nature of the public-sector opportunity. Government buyers on the panel showed little interest in generic claims that AI could solve any problem. They wanted suppliers that understood specific administrative challenges, could demonstrate functioning products and had a credible route through government security, procurement and implementation requirements.
Speed is becoming another consideration. Traditional public-sector technology programmes can take several years to procure and deploy. AI development cycles are considerably shorter, creating tension between rapidly changing technology and government processes designed around stability and long planning horizons. Indiana’s technology leadership described working around much shorter implementation periods, attempting to deliver projects within months rather than allowing modernisation programmes to extend indefinitely. That approach will not be appropriate for every government system, particularly those involving critical infrastructure or national security, but it demonstrates how expectations are changing.
Internal capacity remains a constraint. Utah, for example, has hundreds of software engineers distributed across state government but only a small central AI engineering team. That requires central specialists to work with individual agencies, transfer knowledge and help existing development teams adopt the technology. AI itself may make that decentralisation easier. Business users increasingly have tools that allow them to prototype interfaces, describe workflows and participate more directly in application design. Work that previously passed sequentially from a department to an IT team can increasingly be developed collaboratively.
This creates opportunities but also governance challenges. Government employees deal with some of the most sensitive information held by any organisation, including financial records, immigration information, law-enforcement material and personally identifiable data. AI systems cannot simply be connected across every database because doing so would undermine legal restrictions governing why information was collected and how it may be used.
Utah’s approach to conversational AI illustrates the issue. The state has developed a common architecture that agencies can use while keeping personal information within the systems responsible for it. Information can be accessed for an authorised interaction without creating a new central repository containing everything known about an individual. This principle becomes increasingly important as AI makes combining information technically easier. A system may be capable of connecting records from multiple agencies and producing a comprehensive picture of an individual, but technical capability does not automatically create legal authority to do so.
Public-sector AI therefore faces a constraint that many commercial deployments encounter to a lesser degree: information must remain connected to the purpose for which government was authorised to collect and use it. DHS faces the same issue at federal level. Privacy decisions can involve determining whether information may legally be shared between agencies or released publicly. Those decisions can depend on legislation, the circumstances of an individual case and potential consequences for national security, operational security or personal privacy.
AI may help organise the material required to make those decisions, but the officials argued that human responsibility cannot disappear. Jankowski described cases where lawyers review the circumstances and provide legal analysis before a final decision is made about whether information can be shared.
The distinction suggests that the familiar concept of keeping a person involved in an automated process may itself evolve. Rather than humans merely checking AI output at the final stage, public officials remain responsible for defining the boundaries within which automated systems operate. That is particularly relevant where AI systems make recommendations concerning licences, regulatory approvals or other decisions affecting individuals and businesses. Automation can identify whether documentation appears complete or whether an application meets defined conditions, but government remains accountable for the final outcome.
Privacy is only one of the risks. Cybersecurity is becoming equally important as malicious actors gain access to more capable AI tools. Utah’s technology leadership identified the need to strengthen thousands of existing applications against AI-assisted attacks as one of its most pressing concerns. DHS faces an additional problem through the transparency obligations of government itself. Information released individually through legitimate public-record requests can sometimes be combined to reveal patterns or operational details that were not obvious from any single document. AI could make that type of analysis substantially faster.
Government agencies will therefore find themselves using AI both to improve transparency and to defend systems against increasingly sophisticated attempts to exploit information. The economics of AI will also matter. Public-sector organisations operate under taxpayer scrutiny and cannot assume that every task should be sent to the most powerful and expensive model available. Utah’s strategy includes matching the technology to the complexity of the task, using smaller models where appropriate rather than automatically relying on frontier systems.
That principle could become increasingly relevant across both government and business. A narrowly defined administrative process may require classification or extraction rather than sophisticated reasoning. Using a smaller specialised model can reduce computing and operating costs while potentially producing more predictable results.
The panel also highlighted a structural feature of government that makes AI adoption particularly interesting. Utah officials said the state’s executive branch employs roughly the same number of people today as it did decades ago while serving a much larger population and providing substantially more services. Regardless of the exact historical comparison, the underlying challenge is familiar across public administrations: service demand can grow considerably faster than government staffing.
AI provides one potential way to increase administrative capacity without expanding employment at the same rate. That does not necessarily mean reducing the workforce. It can mean allowing the same number of employees to process more transactions while directing their attention towards cases that require judgement, communication and expertise.
The implications extend beyond government budgets. Public-sector productivity affects businesses directly. Licensing delays can postpone openings and investment. Slow regulatory reviews can hold up projects. Inefficient company-registration systems increase administrative costs. Delays in accessing government information create uncertainty for individuals and businesses. Better public-sector technology therefore functions partly as economic infrastructure. A government capable of processing routine administrative activity more quickly can reduce friction across the private economy.
This is where the concept of AI for the public good becomes commercially significant. The economic value of artificial intelligence will not come only from companies becoming more productive. It may also come from governments becoming easier to interact with.
The biggest obstacle may ultimately be institutional rather than technological. AI models and development tools are changing on cycles measured in months, while government procurement, regulation and technology governance have traditionally evolved much more slowly. Public authorities must therefore find ways to move faster without abandoning the privacy, security and accountability requirements that distinguish government from ordinary commercial technology deployment.
That balance will define the next stage of public-sector AI. Moving too slowly risks leaving employees trapped in inefficient systems while citizens receive services that increasingly fall behind private-sector expectations. Moving too quickly could introduce security vulnerabilities, privacy violations or automated decisions that governments cannot adequately explain.
The examples from Utah, Indiana and DHS suggest a more pragmatic route. Start with repetitive administrative work, build specialised systems around clearly defined processes, maintain human responsibility for consequential decisions and measure whether the technology genuinely reduces cost or processing time.
If that model succeeds, some of the most consequential applications of artificial intelligence may prove surprisingly ordinary. They may not involve autonomous governments or algorithms making public policy. They may involve contracts being reviewed faster, licences being processed more efficiently, information requests taking less time and public employees spending fewer hours moving data between ageing computer systems.
For citizens and businesses interacting with government, those seemingly mundane improvements could ultimately be among AI’s most tangible benefits.
Source: CIJ.World Research & Analysis Team
