In 2022, the President and Treasury Secretary asked me to lead the IRS through “the largest technology-enabled transformation of the agency” in its history. The agency had significant new funding and practical applications of gen AI had debuted, opening up new possibilities for modernization.
As a result, the IRS received numerous proposals from emerging AI solutions vendors, but questions remained about how AI would fit into our reforms. Our funding came directly from taxpayer dollars, so every AI project had to make a measurable difference for the people we served. Although much of this work was still in its infancy during my tenure, it revealed durable lessons about what effective AI implementation requires.
Company leaders now face similar implementation challenges. By the end of 2025, only 15 percent of executives reported that AI had improved profit margins, even as 68 percent of CEOs plan to invest even more in AI this year.
As organizations close Q1, they must develop practical AI strategies with integrations that yield effective outcomes in the quarters to come. While private businesses usually operate in a wholly separate world than the public sector, every so often they can learn an important lesson from government work.
Here’s one: For effective AI implementation, enterprises can follow the IRS blueprint, identifying a sweet spot of risk-appropriate initial fixes to operational problems instead of abstract transformations.
Establish a ‘Risk Window’
While exciting, the bold, all-or-nothing solutions we received from the initial set of AI vendors revealed potentially extensive risks we needed to mitigate. This led to initial paralysis: How could we guarantee the safety of taxpayer data when AI would be working in close proximity to it?
We tackled this challenge by establishing a workable permission structure to ease us into risk without letting the perception of danger stop us. We established an acceptable tolerance window in which we could bundle smaller-scale solutions, test and refine them, and be ready to scale them to an enterprise level. Starting with controlled, micro-implementations helped us find those that balanced acceptable risk with high reward, illuminated AI’s limits, and showed how it would augment employees’ roles.
Company leaders can establish a similar “risk window” to avoid defaulting to safe but ineffective AI implementations. Like the IRS, executives may find solutions that address customer pain points hold the most promise.
Quickly Solve Frontline Challenges
Instead of looking for the easiest application of AI tools, C-Suite leaders can focus on how the technology could solve specific, persisting problems that plague their customers or end users, without sacrificing employee know-how.
At the IRS, this meant focusing on challenges that were already causing friction, like the taxpayer hotline. The IRS phone lines faced persistent backlogs, long wait times and inconsistent responses. Instead of waiting for large telecommunications upgrades, we deployed AI tools to help route calls more effectively and surface accurate answers faster.
Our primary KPI on phone service went from a historic low of less than 30 percent of calls getting through to a historic high of more than 85 percent. We followed this win up with similar quick AI improvements for taxpayer pain points, reducing the backlog of unprocessed paper returns and correspondence from a historic high of 23-24 million in late 2022 to historic norms of under 1 million by 2024. We also used backend AI for more functional improvements to our website and app platform in two years than in the previous 20.
Importantly, all our early AI projects were deliberately scoped to move quickly. They did not require rewriting legacy systems, migrating massive data sets or waiting for a multi-year modernization program to finish. This quick pace, however, did not mean replacing our human employees. These folks were already deeply embedded in processes we were looking to streamline with AI, so we chose to partner them with the technology rather than lose their specific expertise.
To identify the best AI solutions to enhance existing capabilities, leaders should explore across the spectrum of customization.
Choose When to Customize
Formulating a winning AI strategy can no longer rely on a single AI tool. At the IRS, we recognized that employing three forms of AI was necessary for success.
Off-the shelf tools helped executives streamline non-tax-related administrative tasks but lacked the focused training data set to help with tougher inquiries. Domain-specific AI solutions helped our phone line improvement program properly route niche tax inquiries but were no match for the most sensitive of determinations.
For these instances, we chose custom solutions. We deployed bespoke case management AI to isolate high-risk cases from millions of transactions, ultimately preventing and recovering billions in fraud and improper payments in FY2024. Yet tailored AI solutions can carry high upfront costs and lack adaptability, meaning enterprise leaders should look to implement all three levels of customization to succeed across varied applications.
Build on Initial Wins
The initial success of quick-turn, isolated AI use cases does not mean they’re the right approach in perpetuity.
Rather, solving bounded, functional problems will yield ROI learnings that will inform leaders for the next, more complex wave of AI adoption. As these insights compound, executives will develop a greater knowledge of their organization’s tolerance for risk and find new ways AI can build on existing employee knowledge.
Though their AI tools may operate automatically, executives can’t go on autopilot. Instead, they must work within their company’s risk window to identify a mix of solutions that quickly make a visible difference.
