The Triage State: How Machines Are Learning to Sort the Public’s Legal Business

Every government legal office begins with the same deceptively simple act: deciding what matters. A records request lands in an inbox, a citizen complaint arrives by mail, a new statute triggers a wave of agency litigation. For most of the past half-century, a human being read each one, guessed at its urgency, and dropped it into a queue. That guess, made thousands of times a year, under deadline pressure, was the hidden engine of public-sector legal work. It is now being rebuilt in software.

The shift is being driven less by enthusiasm for technology than by arithmetic. In fiscal year 2024 the U.S. federal government received a record 1,501,432 Freedom of Information Act requests, a 25% jump over the prior year, while the backlog of unanswered requests grew 33% to 267,056, according to the Department of Justice Office of Information Policy. State attorneys general carry litigation loads measured in the tens of thousands of active matters. The work is growing faster than the workforce, and the only lever left to pull is the front door, how requests are received, classified, and prioritized.

The Old Way: A Filing Cabinet and a Hunch

Intake in a public legal office was historically an analog ritual. A paralegal or line attorney opened the mail, decided whether a matter was routine or urgent, ran a conflict check against memory and a card index, and logged it into a docket that might be a ledger, a spreadsheet, or, at best, an aging case-management database. Risk was assessed the same way it had been for generations: by experience and instinct. The senior litigator “knew” which cases would settle and which would metastasize.

That model worked when volumes were modest and stable. They have been neither for decades. Federal records litigation alone has been accumulating since the 1970s; a 1979 Government Accountability Office review found that information-access cases were arriving at the Justice Department’s Civil Division roughly twice as fast as they were being closed, tripling the pending inventory over the period studied. Decades later the same structural mismatch shows up in employment disputes: GAO documented that federal-sector discrimination complaint inventories more than doubled in the 1990s, with the average case taking well over a year, and in some pipelines more than three years, to resolve.

The cost of manual triage was not only delay; it was inconsistency. When prioritization lives in an individual’s head, it is invisible, unauditable, and unevenly applied. A complex request and a trivial one could sit in the same queue in order of arrival, because, as the federal FOIA program itself notes, agencies typically process requests in the order received. The public could not see how decisions were made, and neither, often, could the agency.

The Shift: From Order-of-Arrival to Order-of-Risk

The pressure is now impossible to ignore. Federal FOIA volume has climbed relentlessly, and the backlog has reaccelerated after a brief plateau. The trajectory is the clearest argument for automation in the entire public-sector legal world.

Agencies have responded by processing more, 1,499,265 requests in FY2024, up nearly 34%, yet still fell behind, with the backlog rising to 267,056 and the average response time for even simple requests lengthening from 39 to 44 days, according to the University of Florida’s Brechner Center. Running faster on the same track is not a strategy. Re-engineering the track is.

That re-engineering centers on three capabilities that were once separate clerical chores and are now collapsing into a single automated layer: intake (capturing and structuring an incoming matter), conflict checking (screening for relationships that disqualify the office or an attorney), and risk scoring (predicting which matters are urgent, costly, or precedent-setting). Surveys of the broader legal market show how fast the underlying appetite is moving: the Corporate Legal Operations Consortium found 30% of legal departments already using AI and another 54% planning to adopt within two years, with 83% expecting demand to keep rising.

The economics are stark. A landmark Deloitte Center for Government Insights analysis estimated that automating routine federal tasks could free between 96.7 million labor hours (roughly $3.3 billion) at low investment and 1.2 billion hours ($41.1 billion) at high investment, potentially up to 30% of government workforce time within five to seven years. Crucially for legal offices, Deloitte found that simply documenting and processing information consumed nearly 800 million federal staff hours a year, the exact clerical substrate that automated intake attacks.

What It Looks Like Now: Triage as a Pipeline

In a modernized public-sector legal office today, a matter no longer waits for a human to notice it. An incoming FOIA request or complaint is parsed on arrival: the system extracts the requester, the subject, the responsive record types, and statutory deadlines, then routes it to the correct component automatically. A central repository with automated sorting, relevance prediction, and redaction reduces the bottlenecks that the Thomson Reuters Institute identifies as the root cause of records-request delays and downstream litigation.

Conflict screening, once a manual archaeology of old files, is being rebuilt around entity mapping. Market and practitioner studies report that AI-driven conflict systems cut preliminary screening time by roughly 70%, and that natural-language matching catches the corporate-subsidiary and alias relationships that keyword systems miss. The stakes are not theoretical: the American Bar Association’s 2025 technology survey found that 23% of firms reported at least one conflict “near miss” in the prior year, while offices using AI-powered entity mapping reported near-miss rates below 3%.

Risk scoring is the most consequential, and most contested, piece. Machine-learning models trained on historical dockets now generate probability estimates for litigation outcomes. Published accuracy figures cluster in the 78 to 92% range for data-rich case types such as contract and employment disputes, comfortably above the 60 to 75% range typically attributed to experienced human forecasts, according to a survey of legal-prediction research compiled by Clio and an academic study summarized by Harvard Law School’s governance forum. For an agency deciding which of 12,000 active suits deserves senior attention, that gap is operationally meaningful.

What does this buy a public office? Time, mostly, and the redirection of it. Government lawyers report that 47% spend too much time on administrative tasks and 54% lack time to research complex issues, per the Thomson Reuters Institute. Where automated workflow tools have been deployed, internal ROI surveys cited by the same institute found 83% of users agreed the technology reduced administrative time and 88% said it freed them for higher-value work. A UK government trial of more than 20,000 civil servants found AI assistance saved the equivalent of nearly two working weeks per person per year.

The savings concentrate exactly where the drudgery lives. Deloitte’s task-level work estimates that smart technologies can absorb 75% to 95% of the effort on tasks such as drafting routine reports and routing documents to the right reviewer, precisely the labor that clogs an intake desk.

The maturity ladder

Most government legal offices are not at the frontier; they are climbing toward it. A rough maturity model, synthesized from public-sector legal-tech assessments, places the median office in early transition: solid document management, nascent automated intake, and risk scoring still largely aspirational.

Illustrative maturity estimates synthesized from Thomson Reuters Institute and CLOC adoption data; bars indicate relative penetration across government legal departments, not a single audited survey.

The Triage Math for One Office

Consider a single state attorney general’s office. Michigan’s attorney general has reported carrying a litigation caseload of roughly 29,000 cases plus about 4,000 administrative matters at a given time, per the Michigan Bar Journal. No human can rank that volume by risk in real time. This is the use case that automated scoring was built for: not replacing judgment, but ordering the queue so that scarce senior judgment lands on the matters that move the needle.

Procurement disputes illustrate the same dynamic from another angle. The Government Accountability Office received 1,803 bid-protest cases in FY2024 and reported a 52% effectiveness rate, more than half of protests yielding some relief to the protester, according to its annual report to Congress. An office that can score which protests are likely to succeed can settle early, conserve litigation resources, and reduce the corrective-action churn that drives the effectiveness rate in the first place.

The Next Few Years: Promise, and the Bias Problem

Three trajectories look durable. First, intake and scoring will fuse into a single “triage layer” that sits in front of every queue, records, litigation, advisory, rather than living as separate tools. Second, public-interest risk scoring will expand beyond cost-and-likelihood into impact: which matters affect the most people, set precedent, or carry reputational exposure for the agency. Third, the regulatory scaffolding will harden. Federal guidance such as the White House’s M-25-21 memorandum on federal AI use already pushes agencies toward governed, documented adoption.

But the public sector cannot import private-sector scoring uncritically. Risk models encode the patterns in their training data, and where that data reflects skewed enforcement or representation, the model reproduces it. As an academic review of AI legal triage argues, the legitimacy of automated triage in public institutions depends on embedded transparency, explainability, and sustained human oversight, not accuracy alone. Calibration and error-rate fairness can be mathematically irreconcilable, which means an agency must choose, deliberately and on the record, what “fair” means for its mission.

The investment case, meanwhile, only sharpens with delay. Microsoft-commissioned research on the UK public sector estimated £17 billion in potential annual savings from generative AI by 2035, but warned that a five-year delay could forgo £150 billion in cumulative economic benefit. For legal offices specifically, the Thomson Reuters Institute projects AI could save professionals up to 12 hours per week within five years. The constraint is rarely the technology; it is the 53% of government legal teams that cite bureaucratic approvals and the procurement frictions that slow public buyers.

Conclusion: The Queue Is the Policy

For decades, how a government legal office sorted its work was treated as plumbing, invisible, unexamined, beneath strategy. Automated intake and risk scoring make that plumbing visible and, for the first time, designable. The queue is no longer an accident of who opened the mail; it is a deliberate expression of what the office values, encoded in rules that can be audited and revised. That is the real transformation. The machines are not deciding the public’s legal business, they are forcing public institutions to decide, explicitly, how that business should be decided. In an era of record demand and flat resources, that clarity may be the most valuable output of all.

Sources

1. U.S. Department of Justice, Office of Information Policy, Summary of FY2024 Annual FOIA Reports
2. University of Florida Brechner Center, FOIA requests, denials, backlogs surge in FY2024
3. Government Executive, Federal government likely to receive record FOIA requests
4. U.S. Government Accountability Office (LCD-80-8), Federal records-access litigation backlog
5. U.S. Government Accountability Office, Rising trends in EEO complaint caseloads (GGD-98-157BR)
6. Thomson Reuters Institute, Technology: the missing ingredient for government legal work
7. Thomson Reuters Institute, Legal technology transformation in government agencies
8. Corporate Legal Operations Consortium, 2025 State of the Industry Report
9. Deloitte Center for Government Insights, AI-augmented Government
10. Deloitte, AI: Can smart technologies drive government efficiency?
11. Clio, Legal AI case outcome prediction (compilation of published accuracy data)
12. Harvard Law School Forum on Corporate Governance, Predicting litigation risk via machine learning
13. PW Consulting, Worldwide conflict-checking software market research
14. Journal of Law & Emerging Technologies, AI applied to legal triage: efficiency, ethics, oversight
15. Michigan Bar Journal, Plain English in the Department of Attorney General (caseload data)
16. Holland & Knight, GAO Annual Bid-Protest Report, FY2024
17. Microsoft Source EMEA, Public-sector AI time-savings and UK civil-servant trial
18. Thomson Reuters, AI set to save professionals 12 hours per week by 2029
19. The White House, Memorandum M-25-21 on accelerating federal use of AI