GSA EOA push tests explainability in high-stakes AI
The operational barrier is not whether models can process the work, but whether an agency can defend the result.
TL;DR
A NextGov commentary ties GSA’s Elimination, Optimization and Automation Handbook to federal AI explainability, arguing agencies should not stop at low-risk automation. The affected audience is the program office, counsel and engineer asked to use AI for benefits, claims, security calls or export-control reviews. The useful warning is basic: a human reviewer without visibility or override authority is decoration, not control.
GSA’s Elimination, Optimization and Automation Handbook is framed as process improvement, not an AI rule. NextGov’s commentary turns it into the more important implementation question: if agencies automate only the safe and boring work, they may save time while leaving the consequential backlog untouched. If they automate benefits, claims, case work, security decisions or export-control triage, someone eventually has to explain the result to an inspector general, a court, Congress or the public.
That is the right fault line. The piece cites a Brookings Institution report finding that more than 85% of the government’s high-impact AI deployments in 2025 were missing some risk information agencies were required to publish. Treat that as a compliance signal, not a technology complaint. An agency can tolerate an imperfect first-pass tool more easily than it can tolerate a decision path it cannot reconstruct.
The practical standard is not “human in the loop.” It is whether the human can see the rule, record or verified source that drove the output, correct it, and overrule it. If the workflow cannot do that, the agency has not automated a decision. It has outsourced the explanation and kept the liability.
Published ·Deep Fathom