The Civil Servant That Never Sleeps

Every government legal office runs on the same scarce currency: attorney and analyst time. A statutory deadline does not care that a records team is short-staffed, that a docket drew a million comments, or that a single benefits appeal touches six legacy systems. For decades the public sector's answer was to add bodies, add overtime, or quietly let the queue grow. That equation is now being rewritten by a class of software that does not merely assist a human at a keyboard but acts on its own, reading an incoming request, deciding what to do next, taking the step, and stopping to ask for help when it is unsure. The shift from passive tools to agentic workflows is the most consequential change to government legal operations since the move from paper to PDF.

The premise is deceptively simple. An agentic system is triggered by an event, a new Freedom of Information Act (FOIA) request lands, a regulation enters a comment period, a case file is opened, and then executes a multi-step process within defined guardrails: search, classify, draft, route, redact, summarize. When confidence drops below a threshold, or a decision carries legal weight, it escalates to a person. Crucially, every action is logged. That last property is what makes the category plausible for government at all, where the question is never just "can it work?" but "can you prove what it did, and why?"

The Old Way: Drowning in Paper Deadlines

To understand why agentic workflows landed in government legal offices, start with the arithmetic of the inbox. Federal agencies received a record 1,501,432 FOIA requests in fiscal year 2024, a 25% jump in a single year, and although they processed nearly as many, the backlog of unanswered requests still climbed 33% to 267,056, according to the Department of Justice Office of Information Policy. Processing that mountain cost roughly $601 million, with another $55 million spent litigating disputes, the FOIA Advisor reported from the underlying data.

The work itself was punishingly manual. A records officer would open a request, hunt across email archives and shared drives, eyeball thousands of pages for responsive material, apply exemptions line by line, and redact by hand. The average response time for even a "simple" request stretched from 39 to 44 days, while the share of requests granted in full has slid from a high of 38% in 2010 to just 16%, per analysis from the University of Florida's Brechner Center. Agencies overwhelmingly blamed rising request volume, growing complexity, and the loss of experienced staff.

Rulemaking carried its own scale problem. Notice-and-comment dockets routinely attracted enormous public response, the Environmental Protection Agency's 2015 Clean Water Rule alone drew more than one million submissions, a volume no team can read line by line, as documented in the agency's own docket and summarized by a U.S. Senate review of the process. Casework, benefits adjudication, immigration matters, compliance reviews, multiplied the same pattern across every agency. The cost was not only money. An influential Deloitte analysis estimated that simply documenting and recording information consumes roughly half a billion federal staff hours each year, worth more than $16 billion in wages.

The Shift: From Tools to Agents

The early wave of public-sector automation was robotic process automation and templated search, rigid scripts that broke the moment a form changed. What is different now is autonomy paired with judgment. The pivot is visible in the adoption data. The Government Accountability Office found that documented AI use cases across 11 surveyed federal agencies nearly doubled from 571 to 1,110 in a single year, while use cases involving generative AI specifically jumped ninefold, from 32 to 282, according to reporting on the watchdog's findings by Nextgov/FCW.

The broader inventory tells the same story of acceleration. The government-wide tally of AI use cases grew from 710 in 2023 to more than 3,600 in 2025, a five-fold expansion in two years, per the Brookings Institution. The intermediate 2024 inventory listed 1,757 uses, a figure later revised upward, and the 2025 count represented a 105% jump over the prior year, Nextgov/FCW reported. Mission-enabling and internal-support functions, exactly the back-office legal and administrative work agentic systems target, accounted for the majority of those cases.

Yet adoption ambition still outruns delivery. A government survey by EY found that while 64% of public-sector respondents recognized AI's potential to cut costs, only 26% had integrated AI across their organization and just 12% had deployed generative AI. Nearly two-thirds (62%) named data privacy and security as the chief barrier. A 2026 EY survey of federal leaders reinforced the gap: 92% see AI as critical to efficiency, yet 86% report barriers to scaling it, and only 38% have a unified AI governance strategy, the firm reported.

The DOJ line is the telling one: it reduced its FOIA backlog by more than half even while receiving nearly 20% more requests, crediting "acquiring new technologies to improve FOIA processing," per its 2025 Chief FOIA Officer Report. That is the agentic thesis in miniature, capacity decoupled from headcount.

What It Looks Like Now

Strip away the branding and present-day agentic legal workflows in government share a recognizable anatomy. An intake agent classifies an incoming request by type, urgency, and the office responsible, then opens a matter and starts the statutory clock. A retrieval agent searches across records systems, assembles candidate documents, and proposes which are responsive. A review-and-redaction agent flags likely exemptions and personally identifiable information for a human to confirm. A drafting agent composes the response letter or the summary of public comments. Throughout, a confidence threshold governs when the chain pauses and hands off to a person, the principle that 89% of leaders in an EY survey insist must remain non-negotiable: built-in human intervention.

The audit trail is no longer paperwork after the fact. In a public-sector agent, the log of every decision is the work product.

This is where guardrails stop being a slogan. Practitioners in regulated settings now treat each agent run as producing a single, queryable decision record, capturing the prompt, the documents retrieved, the model version used, the tool calls made, and every point at which a guardrail fired or the system escalated, as described in guidance on audit trails for regulated industries. Many agencies anchor their controls in the voluntary NIST AI Risk Management Framework, whose Govern function requires that "legal and regulatory requirements involving AI are understood, managed, and documented." The result is a tiered model: low-impact actions run autonomously, medium-impact actions require human-in-the-loop sign-off, and high-impact legal determinations stay human-authorized.

The payoff, where the workflow is matched to the right task, is measurable. Deloitte's modeling suggests AI could free up roughly 30% of government workers' time within five to seven years under strong adoption, freeing between 96.7 million and 1.2 billion federal staff hours annually, worth an estimated $3.3 billion to $41.1 billion, per its analysis of cognitive technologies in government. Early state deployments hint at the trajectory: Pennsylvania reported average savings of more than an hour per day among users of a generative AI tool, and in Colorado more than a third of users saved at least six hours per week, according to Deloitte's state and local government research.

The Next Few Years: Autonomy Meets Accountability

The trajectory is steep. Gartner projects that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI, up from effectively zero in 2024, and that 33% of enterprise software applications will embed agentic capabilities, up from less than 1%, as reported by Reuters. For government legal offices, the most plausible near-term frontier is the bounded, repeatable, reversible task: triaging FOIA queues, clustering and summarizing public comments, drafting routine correspondence, and pre-screening compliance filings, while final legal judgment stays human.

The cautionary signal is just as loud. Gartner expects more than 40% of agentic AI projects to be scrapped by 2027, undone by escalating costs, unclear value, and inadequate risk controls, per Reuters. In government the failure modes carry a sharper edge: algorithmic bias in benefits decisions, exposure of citizen data, and the prospect of an automated denial that no one can fully explain. Surveys already show oversight lagging deployment, in one EY poll, 52% of department-level AI initiatives operated without formal approval and 78% of leaders admitted adoption was outpacing their ability to manage the risks, the firm reported.

This is why the public-sector adoption curve will look different from the private sector's. The institutions that move fastest will be those that solve oversight first, that can show a court, an inspector general, or a citizen exactly which records an agent read, which exemptions it applied, and where a human stepped in. Public trust is the binding constraint, not raw capability. A denial that a person cannot defend, or a redaction an agent cannot justify, is a legal and political liability regardless of how fast it was produced.

Conclusion: Augmentation, Not Replacement

The arc is clear. Government legal work began the decade buried under record request volumes and shrinking staff; it now has, for the first time, a way to scale capacity without simply scaling headcount. Agentic workflows will not, and should not, make the legal calls that define public accountability. Their value is in clearing the path to those calls: triaging, retrieving, drafting, and summarizing so that scarce human judgment lands where it matters most. The technology that wins in the public sector will be the one that can prove its work as reliably as it does its work. In government, the receipt is the product, and the civil servant that never sleeps is only trustworthy if it never stops writing things down.

Sources

1. U.S. Department of Justice, Office of Information Policy, Summary of FY2024 Annual FOIA Reports
2. FOIA Advisor, Final FY2024 Data Indicate More Requests and Delays
3. University of Florida Brechner Center, FOIA Requests, Denials, Backlogs Surge in FY2024
4. U.S. Department of Justice, 2025 Chief FOIA Officer Report
5. Nextgov/FCW, Agency AI Use Doubled in 2024, GAO Finds
6. FedScoop, Generative AI Use 'Escalating Rapidly' in Federal Agencies, GAO Finds
7. Brookings Institution, Assessing the State of AI Adoption Across the Federal Government
8. Nextgov/FCW, Agencies Report Over 3,000 AI Use Cases in 2025
9. Reuters, Over 40% of Agentic AI Projects Will Be Scrapped by 2027, Gartner Says
10. Analysis of Gartner Forecast, 15% of Work Decisions Autonomous by 2028
11. EY, Survey Reveals Gap Between Government AI Ambitions and Reality
12. EY, Federal Government Agencies' Efficiency Efforts Face Significant Barriers
13. EY, AI Investments Surge, but Agentic AI Understanding and Adoption Lag
14. EY (via PR Newswire), Autonomous AI Adoption Surges as Oversight Falls Behind
15. Deloitte, AI-Augmented Government: Cognitive Technologies and Federal Time Savings
16. Deloitte, Scaling Generative AI in US State and Local Governments
17. NIST, AI Risk Management Framework (AI RMF 1.0)
18. Guidance, AI Agent Audit Trails for Regulated Industries
19. U.S. Senate, Staff Report on the Federal Notice-and-Comment Rulemaking Process