Building agentic AI for broad use by non-experts in policy and research expands risk beyond text generation: it introduces the possibility of errors in tool use, workflow execution, quality control, and decision automation, among others.
In these situations, organizations need a practical standard that lays out exactly how to best develop agentic generative AI: what to measure, how to test, and what artifacts to retain so we can achieve our missions.
We consider agentic generative AI “responsible” when builders and organizations
- define a clear purpose and measure success,
- provide agent oversight and ownership,
- minimize known risks, and
create an accessible audit trail.
General operating principles: pilot, improve, release, improve again
How does an organization deploy an agentic generative AI tool when no previous evidence base exists, staff members have not fully wrapped their arms around risk mitigation, and internal decisionmakers want to get up and running quickly? It’s a great question.
In these cases, we recommend a standard technology practice focused on limited pilot programs. These pilots—thoughtfully constructed to minimize the risks and maximize the benefits outlined in this document—help organizations gather evidence for consideration for a full deployment, and they help accountable owners clearly assess and sign off on the return on investment relative to both cost and risk. Once sufficient progress is made, the organization should consider a broader release. This process follows best practices in software development, such as Lean product development and Agile management.
Next section: Principle 1: Define a Clear Purpose and Measure Success