Don’t just pick the low-hanging fruit — harvest the whole orchard

The General Services Administration published its Elimination, Optimization and Automation Handbook to guide federal agencies in using AI to eliminate unnecessary tasks and automate repetitive ones. However, the government's adoption of AI is hindered by a lack of explainability, as black-box AI cannot defend its decisions to regulators or the public. The article argues that agencies should build transparent, explainable AI systems to tackle high-stakes work like export control, rather than limiting AI to low-risk tasks.

Don’t just pick the low-hanging fruit — harvest the whole orchard The instinct is to point AI at low-stakes busywork. The bigger payoff is the high-stakes work everyone’s avoiding. Why don’t we have both? The private sector isn’t just diving into the deep end of the artificial intelligence pool; it’s swimming with the sharks. The promise of dramatic efficiency improvements, accelerated production schedules and perhaps most important, vastly lowered costs, is fueling AI’s meteoric expansion across every industry. But these businesses are also exchanging AI’s potentially exponential rewards for increased tolerance to risk. Government agencies don’t have that luxury. Early last month, the General Services Administration published its Elimination, Optimization and Automation EOA Handbook https://www.gsa.gov/system/files/Federal%20EOA%20Playbook%20-%20v1%20-%206.3.2026 0.pdf , a guide to help agencies eliminate unnecessary tasks, optimize processes worth keeping and automate repetitive ones. The obvious tool to help achieve these goals is AI, and the instinct is to aim AI at easy targets like low-stakes work where mistakes aren’t as costly or as risky. That’s not entirely wrong. Automating the easy, mundane tasks that are less risky can save tremendous amounts of time and money, but it also keeps more consequential work stuck. And it’s stuck for good reason because AI can’t be explained or defended to regulators, congressional oversight committees, or already AI-weary taxpayers. The bottleneck in using AI isn’t missing capabilities or the inability to handle volume, but explainability. When AI makes a decision about a benefit, a claim, a case or a security call, someone will eventually need to defend that decision to an Inspector General, a court, or the public. That’s something that black-box AI just can’t do. In fact, an April 2026 report https://bit.ly/4vGm6Km from the Brookings Institution found that more than 85% of the government’s high-impact AI deployments in 2025 were missing some of the risk information agencies are required to publish. The most consequential systems already used in the field today can’t account for how they reached their results. For government agencies, results that can’t be explained are results they can’t use. The answer isn’t to aim low, but to build AI that can answer for itself no matter where you aim. That requires more than a human in the loop; it means workflows must be transparent and explainable using words we can all understand. Every output should point back to specifics: a rule, a record or a verified source that drove it. A human should have the real authority to review, correct, and overrule. And rules should only update when a change is proven, not silently in the background without human intervention. When AI can account for itself, you can aim at more consequential, higher-stakes work. Export control, for example, is a decision that carries criminal liability if it's wrong. That’s the kind of work you’d never want to hand off to some black box, but a system that shows its work can now take the first pass. This is precisely the kind of rule-bound work the GSA’s EOA handbook says to automate. There are cautionary tales all around us about why explainability is critical in high-stakes work. Recently, UnitedHealth allegedly used an algorithm https://www.cbsnews.com/news/unitedhealth-lawsuit-ai-deny-claims-medicare-advantage-health-insurance-denials/ to terminate post-acute care for Medicare Advantage patients. But patients who appealed won more than 90% of their cases. The company’s rationale for continuing to use the algorithm? Only 0.2% of patients actually appealed the terminations. But because, as the lawsuit alleges, the humans in the loop couldn’t understand how the decisions were made, and since they didn’t have authority to overrule the algorithm, it was allowed to continue running. It’s not all bad news, though. There are plenty of examples of how tackling consequential work has real benefits. Take, for example, how the Transportation Security Administration TSA implemented new AI-powered scanning technology https://www.dhs.gov/ai/use-case-inventory/tsa with the goal of reducing busy work for TSA agents. If agents have more time to focus on complex issues, they can move people through security lines faster and provide help to travelers who need it most. So, when used well, automation doesn’t replace judgment, it frees up more time for it. Using AI to automate repetitive, low-stakes tasks isn’t wrong, but it does limit the value AI can provide. AI deployments stuck across government agencies shouldn’t be steered around high-stakes, critical, meaningful work. For example, the GSA’s Federal EOA Playbook identifies how AI can leverage Natural Language Processing NLP and Document Understanding DU to address challenges associated with unstructured data, which has long been a perennial hurdle in providing efficient service delivery. What is most needed here is for AI outputs to be made defensible and auditable. If agencies get that part right, the consequential work they’ve been avoiding becomes the work they can finally do. Derek is Senior Vice President of Government at Seekr. He previously served as the Principal Investigator and Research Scholar at the National Security Agency’s Laboratory for Analytic Sciences. Derek earned his B.S. in Operations Research from the U.S. Air Force Academy and a Master’s degree in Operations Research from Southern Methodist University.