Artificial intelligence is rapidly becoming embedded in the public systems that shape economic mobility—health, housing, workforce, benefits access, and financial stability. These new agentic tools can do more than provide advice; they can autonomously guide people through complex benefits applications and high-stakes decisionmaking processes and help them take action.

In these high-stakes contexts, errors, bias, or lack of transparency can directly affect people’s lives. The challenge is not just building these tools, it is ensuring they are safe, high quality, and trustworthy for both the end users and the organizations deploying them.

This playbook aims to help practitioners use agentic AI more rigorously and responsibly. It first explains how the Urban Institute defines responsible agentic AI, then offers guidance on developing, testing, and evaluating agentic AI tools for internal and external use. It also includes recommendations for creating governance processes to ensure AI tools meet an organization’s goals.

Who Should Use This Playbook

This playbook is aimed at agentic generative AI application builders who are creating responsible agentic AI agents (or tools) for non-experts to use in the field. Builders creating AI agents only for expert use may still wish to use these principles if they are concerned about reliably verifying the agent’s output.

The principles and processes in the playbook could help

• a state or local agency build a public-facing AI chatbot to help constituents apply for benefits;
• a programmer use agentic programming tools (like Claude Code or GitHub Copilot) to create new AI agents and subagents for non-experts to use; or
• a programmer or expert assemble existing third-party agents, connectors, skills, or tools into a new workflow for non-experts.

The playbook is not intended for

• an organization purchasing an enterprise generative AI tool (like Claude or ChatGPT’s web interface) for all staff to use; or
• a programmer or programming team using agentic programming tools (like Claude Code or GitHub Copilot) alongside agents, connectors, skills, or other tools built for their own use rather than for non-experts.

How to Use This Playbook

If you are new to this topic, go to Defining Responsible Agentic AI. This is also a useful introduction to the later sections.

If you are building an agentic tool, go to our practical guide to internal-facing agents or our practical guide to external-facing agents.

If you are overseeing a team building an agentic tool, go to Governing Responsible Agentic AI.

Keep in mind that the processes and recommendations in this playbook are goals to strive for. In many cases, fully meeting all goals may not be practical or feasible, particularly for early-stage projects or smaller teams. In those cases, builders should prioritize the elements most relevant to their AI agent's risk profile, document what they are and are not able to address and why, and develop a plan to close gaps as the agent matures.

The playbook is a living document. We will be updating it with lessons from Urban’s implementation of agentic AI tools and internal and external feedback. We hope it helps organizations like Urban learn as they grow.

Next section: Defining Responsible Agentic AI