Why ‘mundane’ uses are most effective for AI

Until recently, one analyst supporting San Jose, California’s 311 system spent nearly a quarter of the workday reading requests filed under “Other Issues.” Residents use this catch-all field to report everything from potholes to broken streetlights. Each submission had to be read, interpreted and sorted so the city could understand what people needed.

So the city gave the sorting to a large language model. The analyst now spends that recovered time identifying patterns, improving service insights and helping departments make better decisions about future services. The backlog shrank and the work got better, and a public servant got back a meaningful part of the day, focusing their judgement where it was most useful. 

This is what artificial intelligence in government should mostly look like right now: not the chatbot that finally fixes the Department of Motor Vehicles, not the algorithm that ends homelessness, but the unglamorous tool that gives a caseworker their afternoon back to get real work done.

There’s reason for caution, certainly. Governments have a painful history of using automation for the wrong jobs. Michigan’s MiDAS system, a rules-based fraud-detection program, falsely accused tens of thousands of residents of unemployment fraud, then garnished wages and seized tax refunds. In Arkansas an algorithm used to allocate Medicaid home-care hours left some people with severe disabilities facing sudden cuts they couldn’t meaningfully understand or appeal.  

Adjudicating who deserves benefits, who counts as a fraudster, who gets care requires human judgment. 

The mundane use cases are different. Case managers, housing staff, workforce navigators, benefits administrators, police officers and analysts spend huge portions of their day on the administrative scaffolding: copying information between systems, filling forms with data they’ve already collected, tracking down missing documentation, wading through eligibility guidelines written in legalese. 

About 69% of social workers cite administrative burden as a key driver of burnout, and a majority of police officers report spending more than three hours of every shift on paperwork. The work that actually requires their humanity — building trust with clients, understanding complex circumstances, exercising judgment — gets squeezed into whatever time remains. 

There are clear opportunities. Natural language processing can pull structured information out of unstructured documents; smart forms can walk users through complex applications and catch errors before they cascade; chatbots can answer the routine staff questions that might eat up a senior colleague’s afternoon. 

None of this requires a technological breakthrough. And critically, none of them touches the question of who is eligible, who is committing fraud, or who deserves care. They sit underneath those decisions, doing work that ideally should never have been a person’s job to begin with. 

Our partners have gravitated toward these kinds of projects: a chatbot to help Alexandria, Virginia staff navigate building codes; a real-time document checker for benefits eligibility in New Mexico; a data cleaning and automation tool in San Antonio. None of these will make a magazine cover, they’re not moonshots, but they will give thousands of public servants their time back to focus on what’s most important.

There are hundreds of organic examples like these. Since 2023, San Jose has helped organize a national group of public servants to share what’s working, and the GovAI Coalition now includes more than 900 jurisdictions helping tune the right level of scrutiny to the right problem and train governments to do the work well.

To accelerate these efforts, we have to change the systems around them. Governments should adopt risk-tiered AI governance that moves quickly on low-stakes administrative automation, requires human review wherever errors really matter, and reserves the heaviest privacy, bias and security controls for systems that touch rights, eligibility, or enforcement. In other words, automate the paperwork, not the judgment: use AI to summarize, sort, translate, flag, draft and check; require people to decide, approve, deny, sanction and enforce.

Six months from now, AI will be more capable than it is today. Governments that wait for the technology to settle will be waiting forever, and the people they serve will keep paying the cost of paperwork that should have been automated a decade ago.

The question for public-sector leaders isn’t whether to bet on the frontier of what AI can do, it’s whether to use the tools that already work to give their people back the parts of the job that actually require a person. The real promise of AI in government isn’t replacing public servants, it’s giving them more time to be public servants.

Jake Segal is managing director for the public sector practice of Social Finance, a nonprofit that mobilizes capital toward measurable impact by building and enabling outcomes-focused approaches.