Agentic AI: what it is, and why it matters

• Inheriting it through vendors. Because roughly a third of enterprise applications will embed agentic capabilities, agencies will get these features through their existing software supply chains, and they still own the business risk for what a vendor's agent does.
• The regulatory update. In April 2025, Office of Management and Budget (OMB) memos M-25-21 and M-25-22 replaced the more cautious M-24-10 with a pro-innovation stance, directed agencies to retain or name Chief AI Officers and publish AI strategies, and pushed agencies to upskill staff with AI training. In February 2026 the Department of Labor followed with a national AI Literacy Framework aimed at state and local workforce and education systems (TEN 07-25). In the absence of one federal law, states are passing their own (a patchwork local staff must navigate).

How this settles is not yet known, and the early evidence points in more than one direction. One economic argument, the Jevons paradox, holds that when a service gets cheaper and faster, demand for it tends to grow, which would shift work toward higher-judgment tasks (framing problems, supervising AI, reviewing output) rather than remove it. The counterpoint: automation displaces workers from tasks and can reduce demand for labor even while it raises productivity, and the offset depends on new tasks appearing for people; in United States data, economists Daron Acemoglu and Pascual Restrepo found that offsetting effect weakened over the last three decades. The early data on AI carries its own caution: Stanford's analysis of payroll records finds hiring pressure on entry-level workers in the most AI-exposed fields, while experienced workers in the same fields have held steady so far. The finding is debated: in a 2026 response to critics, the authors report that under stricter statistical controls the decline starts later, in 2024, that other factors likely contributed early on, and that the pattern has continued to grow. Deskilling is a caution here too: over-reliance can erode the skills people stop practicing, an irony Lisanne Bainbridge documented for automation in 1983: skills deteriorate when they are not used, and the person monitoring an automated process needs them when it fails. The concern is current: a June 2026 MIT Sloan Management Review discussion takes up the same problem for technology leaders. The federal government moved its AI hiring toward demonstrated skills in 2024 (the OPM competency model). Questions for your office rather than predictions: which roles would you redesign around judgment and review, and if AI drafts the routine work, where do junior staff learn?

The concern here is scale, not a single use. Early computer-science work flagged the training cost of large models (Strubell and colleagues, 2019), and by 2023 the issue had moved from training to everyday use: writing in the journal Joule, Alex de Vries warned that AI’s growing electricity demand could come to rival that of a small country. The aggregate figures are the ones to watch. The Department of Energy’s Berkeley Lab reports that United States data centers used about 4.4 percent of national electricity in 2023 and projects 6.7 to 12 percent by 2028; the International Energy Agency (IEA) expects global data-center electricity to roughly double by 2030, with the AI share tripling. Water is part of the footprint too: a peer-reviewed study estimates that a short AI session can consume around half a liter once cooling and the electricity supply are counted (Li and colleagues, 2025). Per single query, though, the cost is small and still disputed: an early figure of about three watt-hours per query is now considered high, with an independent estimate (Epoch AI, 2025) and a vendor’s estimate for its own product (Google, 2025, which leaves out indirect water and uses market-based carbon accounting) closer to a quarter to a third of a watt-hour, and researchers have measured a wide spread across models (Luccioni and colleagues, 2024). For comparison, the demo in this session keeps its own footprint small: its one AI step runs on a local model on a laptop, and a full run costs about two dollars, nearly all of it web-search credits. Questions for your office rather than predictions: does your procurement ask where a vendor’s data centers draw their power and water, and is that cost weighed the way you would weigh any other utility a new system adds?