Arizona’s policy search bot that helps child safety workers find the right procedures and guidance

Arizona’s approach began with a simple question: How can AI make it easier for workers to find the information they need to do their jobs? Steven Hintze, Chief Data and Product Officer, Information Technology Data and Product, Arizona Department of Child Safety explained that they wanted to demystify AI and start with “boring” straightforward use cases. Every new case manager must navigate an enormous policy manual—the backbone of agency operations and compliance. Knowing where to find the right procedure or clarification can take valuable time away from families.

The Arizona team designed a policy bot, a smart search tool embedded within their case management system, capable of interpreting natural-language questions and returning precise policy guidance. The bot does not summarize policy or paraphrase—it surfaces the exact section to answer a query, verbatim, from the agency’s official manual. A worker might ask: How do I drug test a parent? ... and the system will respond by confirming the intent of the question—Do you mean a parent, a child, or an employee?—before retrieving the relevant policy.

It is designed with a human-in-the-loop structure: the worker validates the question, confirms the context, and decides whether to act on the guidance. Hintze described it simply: “It’s not going to make a safety decision. It’s not going to decide permanency.” It will help a worker do their job with less friction. The underlying model runs inside Arizona’s own secure cloud computing platform, meaning no data leaves their environment.

The state has built in disclaimers against entering personally identifiable information, and has embedded controls to prevent the model from “learning” on sensitive data. The design philosophy was to build internal capability, not dependency on an external source. Greater capability drives greater effectiveness. Put another way: “You can do cloud poorly or locally poorly,” Hintze said, “or you can do both well.”

Lessons from the Arizona Team

The Arizona demonstration generated high engagement—not because it was flashy, but because it was practical. Participants were struck by how much intentional design had gone into applying AI responsibly. The Arizona agency began by defining what success would look like:

• Not replacing people, but replacing the parts of the job that drain people
• Not automating judgment, but automating access to information
• Not experimenting in the dark, but documenting everything—what works, what doesn’t, and why

One attendee asked whether Arizona’s ethical guidelines were made available for peer review, as a model for states drafting ethical guidelines and policies for AI. Hintze explained that Arizona established a governance council reporting directly to the agency director, with representatives from across divisions. This group meets monthly to review new ideas and reports to the director–to audit unstructured data use, and ensure every pilot has clear ethical oversight. Arizona also built an internal peer-review process modeled loosely on IBM’s AI Ethics Board, creating a homegrown system of accountability that can evolve with experience.

Other lessons learned were candid and valuable. When the agency upgraded to a new Gen AI model, accuracy initially dropped. The team had to conduct extensive user acceptance testing (UAT) to recalibrate prompts and restore trust in the tool. They discovered that even small updates could alter tone, mannerisms, and confidence levels—proof that these systems need continuous evaluation. They learned that tracking metrics mattered: which questions workers were asking, how often the bot was right, and when it failed. The team tracked “rage clicks,” a small but telling metric showing when workers clicked repeatedly out of frustration—an indicator of where to improve. By building feedback loops into the pilot itself, Arizona’s DCS didn’t just test a product; it tested an approach to cultural adoption. It built trust by treating caseworkers as co-designers. An IT administrative leader from another state shared that his state was currently “crawling and learning to walk” when it comes to AI, and appreciated the opportunity to learn from the group how this could be applied to child welfare successfully.

Attribution: Case study provided to the AI Navigator courtesy of the IBM Centre for The Business of Government, Apolitical's content partner. “AI in State Government: Balancing Innovation, Efficiency, and Risk”, https://www.businessofgovernment.org/reports/ai-in-state-government, IBM Centre for The Business of Government.