The Future of Government Services in an AI-Powered World

Imagine a system where public services are simple, responsive, and built around people’s lives. Where government isn’t just easy to navigate, but proactive and helpful to every person it serves.

AI is expanding what government can do, and our vision is expanding with it. Fifteen years of adapting to changing policy landscapes and new technologies means Code for America is well-positioned to help build a bolder vision: We believe AI can fundamentally change people’s experience with government for the better. 

Imagine an applicant applying for benefits the way they’d explain their situation to a friend. They share that their hours just got cut, provide a paystub and a housing lease, and answer a few plain-language questions. The AI-enabled system identifies which programs they qualify for across agencies, pulls forward documents they’ve already submitted, and walks them through what’s left—with communications in their language needs and tailored for their specific situation. On the agency side, caseworkers receive a complete, verified case file ready for review instead of a stack of raw documents to sort through.

The opportunity we see in AI is advancing a more efficient, effective, and human-centered government: simplifying complex processes, making benefits easier to access, and reducing friction for clients who need support. In the future we’re helping to build, government leverages AI to work with the speed and care people expect, building trust through systems that are truly human-centered. But how do we get there?

Are you a government leader working to implement AI in benefits delivery, caseworker operations, or some other application of emerging technologies? Get in touch.

Choosing the right applications of AI

We prioritize AI solutions that strengthen programs and reduce barriers for the people who need them most, and deploy AI when it unlocks capabilities that weren’t previously feasible at the scale, speed, or cost that public services demand. To decide when we use AI, we ask two questions: 

• Does AI unlock something new? Traditional software is great when the logic is well-defined and the data are structured. But the reality of government service delivery is that so much of the work isn’t clean and structured. We start by asking: does AI make something newly possible, or does it just add complexity to something we  could handle with simpler approaches? If a simpler tool is available and can do the job, that’s typically the better choice.

• Can we manage the risks that come with it? No technology is risk-free—including the systems governments already rely on. But these risks can be managed with deliberate design choices, and every new implementation warrants a thoughtful review to determine how to get those choices right. 

These questions are the foundation for considering AI implementation opportunities, and how we also design approaches that mitigate risk. Let’s dive a bit deeper into each of these topics to paint the picture of what’s possible with AI, including how we leverage human-centered design thinking in implementation approaches.  

We look for two important criteria when evaluating use cases: that they elevate government staff impact and effectiveness, and that they improve client access to services and information.

Using AI to unlock new possibilities 

AI has potential to improve government—but what does that look like in practice? We look for two important criteria when evaluating use cases: that they elevate government staff impact and effectiveness, and that they improve client access to services and information. Given these criteria, here are some examples of what is possible if government makes the right strategic investments in capacity and infrastructure:

Document processing: Many benefits applications require a stack of documents (pay stubs, IDs, tax forms, lease agreements, utility bills and so on) that caseworkers must manually review, classify, and verify. Multimodal AI models can classify uploaded documents by type, extract key data fields, and check for completeness to improve efficiency of existing processes. 

But what about the new experiences that AI document processing capabilities enable? Rather than asking people to fill out forms and then submit documents to verify what they entered, what if the documents are the application? Generative AI could process uploaded documents in real time, extract key eligibility criteria, and ask the applicant targeted follow-up questions (or additional documentation requests) to address what’s missing or ambiguous. Instead of a resident navigating the system, the system navigates the person’s information to provide the most relevant information and accelerate benefits delivery. 

Caseworker support: Caseworkers making eligibility decisions must synthesize information scattered across uploaded documents, prior case history, and cross-program records. This is a time-consuming process that is often made difficult by the fragmented systems caseworkers are navigating. AI can sort through case characteristics and identify cases that require attention—essentially creating a “morning briefing” for caseworkers with prioritized work item suggestions. 

But AI capabilities, like generative AI, enable new possibilities beyond summaries. AI can read entire case histories and synthesize all the signals to highlight key factors, flag potential issues like missing verifications or data mismatches, and identify which cases are most likely to result in errors or churn. AI can triage cases or trigger actions such as routing flagged cases to supervisors or pre-populating renewal forms.

Language access: Government communications are often delivered in dense English. For the millions of people navigating benefits systems in other languages or struggling with jargon-filled notices in English, this is a barrier to getting the assistance they are entitled to. We can use generative AI to translate and rewrite government notices, forms, and instructions into plain language across multiple languages—producing versions that are culturally appropriate and program-specific. 

If done well, this would meaningfully improve the experience for residents navigating benefits while reducing the downstream issues agencies have to resolve when notices are misunderstood. In the future, AI systems could respond to questions in real time and deliver personalized case updates by text, voice, or chat so residents get clear answers when they need them. For agencies administering these benefits, every question resolved accurately in the moment is one that doesn’t cause a missed recertification or an avoidable error.

Strengthening solutions by managing risks

Every new project brings distinct concerns that will require a thoughtful review. As we integrate AI into our solutions, we’re proactively monitoring several key risk areas—accuracy and reliability, privacy and security, fairness, transparency, and operational impact—to ensure our implementations are responsible and effective. Rather than treating these risk areas as barriers to implementation, we see them as design considerations that strengthen solutions. We choose mitigations that are proportional to their project’s impact to ensure a responsible deployment.

Accuracy and reliability: Generative AI models are probabilistic and sometimes generate plausible-sounding information that isn’t accurate. In benefits delivery systems, unreliable responses can lead to inaccurate guidance, missed benefits, or decisions based on false information. We manage these risks by building test suites of inputs with known correct answers and regularly running evals. And we ground AI solutions with curated sources of information—using techniques like retrieval over document stores and knowledge graphs, direct database queries, and tool use—so responses are informed by trusted sources. 

Privacy and security: We protect the sensitive information people share—such as income or family details—because that trust is foundational to effective public services. We make deliberate choices to configure AI solutions to protect privacy. For example, we implement zero-day data retention policies and opt out of model training so sensitive inputs aren’t stored or used to improve commercial models. And we mask or tokenize personal identifying information before it reaches AI systems whenever the task doesn’t require it.

Fairness: We build AI that performs consistently across the full range of people and cases a government service encounters. Models trained on historical data can inherit errors and patterns that produce worse results for some people than others. We anticipate these differences and build performance metrics into our evaluations that test across different user segments and case types, not just aggregate accuracy. And we validate solutions with people whose circumstances reflect that full range of lived experience.

Observability and transparency: We build AI systems we can inspect, audit, and improve over time. Modern AI models are often described as a “black box,” but the visibility we have is largely a design choice. For example, we enable audit trails by structuring complex AI tasks into smaller, distinct stages, which lets us trace the decision path and assess the logic of each stage. This is one way we keep a human-in-the-loop and gives us the visibility to catch problems early, improve the system over time, and build shared trust in our AI solutions.

Operational impact: Thoughtful implementation choices let us deliver AI capabilities that are efficient, responsive, and cost-effective. ​​Large language models can be slow and resource-intensive when deployed naively, so we design for performance from the start. For example, we right-size model choice by evaluating smaller, efficient models that often match frontier models on focused tasks. And we are deliberate about when to use AI at all—choosing it where it unlocks capabilities that are not feasible otherwise, and calling on the most powerful models only when lighter ones won’t do.

We want to be shoulder to shoulder with government at every stage of their journey—thinking through where AI might be useful, carefully considering the risks, deploying an implementation strategy, and tracking measurable progress towards goals.

Working together to realize the potential of government

Implementing AI can unlock incredible potential in government when it’s deployed responsibly. At Code for America, we’re working to make that happen. We want to be shoulder to shoulder with government at every stage of their journey—thinking through where AI might be useful, carefully considering the risks, deploying an implementation strategy, and tracking measurable progress towards goals. Together, we can build a future where AI powers a proactive government, improving the experience for clients, caseworkers, and public servants alike. A government that works well for everyone is possible. AI can help get us there.