September 14, 2026
Artificial intelligence offers government organizations major opportunities to improve productivity, but former IRS Commissioner Danny Werfel says agencies should resist the temptation to move too quickly.
Werfel, now Executive in Residence at the Johns Hopkins School of Government and Policy, says organizations need to balance the potential benefits of AI with the new risks the technology introduces.
His advice can be summed up in three words: “slow your roll.”
Werfel explains that the basic discipline of risk management does not change with AI. Organizations still need to identify risks, prioritize them, establish controls and monitor whether those controls are working. What changes are the risks themselves.
AI introduces challenges such as hallucinations, drift from an original assignment and the potential for systems to behave autonomously outside a programmer’s instructions. Werfel says those characteristics mean organizations are navigating new territory.
Tax, in particular, could be well positioned to benefit from AI. Werfel describes tax as a rules-based and transaction-based field with large amounts of data that can be analyzed. The potential business case is significant because AI could perform some work much faster while operating at the scale required to serve households and businesses.
But Werfel says much of the conversation focuses on those productivity benefits without giving enough attention to risk.
One tool he recommends is a risk register. Werfel stresses that a risk register should not become a check-the-box exercise. Building one requires organizations to examine the scope of an AI product, determine how different risks could emerge and consider how those risks will be controlled.
He also calls for greater collaboration across professional communities. Werfel argues organizations should share lessons when AI deployments go wrong so others can avoid making the same mistakes. He points to the aviation industry as an analogy, where information about failures is shared to help prevent similar incidents.
That approach becomes especially important in government, where Werfel says the stakes of AI errors can be high.
Organizations need to determine what level of precision is acceptable for each use case. A tolerable error rate in one business process could be completely unacceptable in another. Tax, medicine, law and accounting may therefore require their own conversations about risk tolerance and responsible AI use.
For federal agencies, Werfel’s message is not to reject AI or its productivity gains. Instead, he recommends adopting the technology incrementally.
Rather than using AI to immediately turn a four-week process into a four-minute process, an organization could first reduce it to three weeks. During that stage, leaders can study which risks emerge, determine what training employees need and understand how the workforce may need to be reskilled.
Once those lessons are understood, the organization can move further.
Werfel says that approach allows organizations to capture efficiency while simultaneously building the foundation for responsible AI adoption.
“There's no reason to race ahead to try to grab all that productivity,” Werfel says. “Grab a nice portion of the productivity.”
The ultimate goal is still transformation. Werfel’s argument is that government should get there through a deliberate journey that protects taxpayers, clients and the people relying on AI-enabled services.