Orange County Employees Retirement System

Darren Dang

Institutional Technology Steward

Imagine an AI agent processing a retirement application. It retrieves a policy, checks a member’s record, drafts a response and queues an update. Every technical step succeeds. But the policy was superseded and the reasoning behind the change was never captured in the documents the agent searched.

The workflow moved faster. The member received the wrong answer.

This hypothetical illustrates a leadership challenge facing government: our ability to deploy agents may be advancing faster than our ability to direct them responsibly.

AI can help public agencies analyze information, improve service and preserve expertise. Its greatest contribution should be stronger people and institutions. Achieving that requires building agency before agents.

Human agency is our capacity to understand, question, choose, intervene and remain accountable. Institutional agency is the collective capacity to sustain purpose, trustworthy knowledge, legitimate decision rights and learning across personnel and technology changes.

A person clicking “approve” does not guarantee either. Review becomes meaningful when the reviewer has the evidence, expertise, time and authority to challenge the recommendation. Citizens also need understandable explanations and practical ways to correct or contest consequential decisions.

I call the mismatch between technical capability and institutional readiness the Agency Gap. An agency may have sophisticated tools while lacking clarity about which policy governs, who may authorize an exception or how an error becomes a correction.

“Technology can amplify purpose and expertise; it can also amplify confused authority and fragmented knowledge.”

To close that gap, I propose five readiness gates.

Purpose. Start with a clear North Star: the public outcome the work should improve. Define whose interests are affected and what must be protected. Faster processing matters only when it advances accurate, fair and understandable service.

Knowledge. Establish which sources are authoritative, current and appropriate for the decision. Preserve ownership, effective dates, permissions and correction histories. Search relevance cannot establish authority. An outdated training presentation and an approved policy may both look convincing to a retrieval system.

Judgment. Make decision criteria, tradeoffs, uncertainty and exceptions visible. Preserve why consequential choices were made so successors can assess whether the reasoning still applies. Identify where professional discretion is required and where an agent must stop and escalate.

Governance. Translate policy into enforceable permissions, monitoring and stopping power. Name the accountable owner. Separate access to information from permission to act. Test whether the institution can detect unauthorized behavior, interrupt execution and repair an error.

Bounded delegation. Grant only the autonomy justified by evidence from the specific workflow. An assistant drafting general guidance and an agent changing an official record carry different consequences. Expand authority as performance earns confidence, with explicit conditions for review or withdrawal.

These gates support experimentation. A bounded pilot can reveal missing knowledge, contradictory procedures and weak controls. The threshold should rise with the consequence, uncertainty and difficulty of reversing an action. Learning can begin early while consequential delegation remains carefully constrained.

A practical starting point is one workflow over 90 days. First, map its purpose, sources, decisions, owners and exceptions. Next, build a limited prototype and test difficult cases, including conflicting guidance and missing information. Then compare results with the existing process and decide what authority has been earned.

Measure more than time saved. Examine accuracy, service outcomes, source traceability, correction time, retained employee expertise and whether human intervention actually changes outcomes. A dashboard showing tasks completed cannot tell leaders whether people received better service.

Keep experienced staff involved in testing and correction. Their knowledge helps expose exceptions and their continued practice preserves the expertise needed when automation encounters an unfamiliar case.

Institutional learning must continue after deployment. Capture what went wrong, correct the governing knowledge, revise permissions when necessary and preserve the reasoning for those who follow. Otherwise, automation can repeat yesterday’s mistakes at tomorrow’s scale.

AI is extraordinary rocket fuel for human intelligence. Leaders still have to establish direction, build judgment and remain responsible for the consequences. Technology can amplify purpose and expertise; it can also amplify confused authority and fragmented knowledge.

For government CIOs, the next step is concrete: choose a workflow and ask whether the institution can explain, supervise, correct and own what an agent does. Those capabilities create the conditions for durable innovation and earned public trust.

Build agency before agents. Let AI help people see more clearly, decide more thoughtfully and serve more faithfully.

The views expressed are the author’s own.