The Clerk That Never Sleeps

A court case is, at its administrative core, a sequence of small decisions repeated millions of times a year. Does this filing meet the rules? Which judge gets it? When can the parties be heard? Who must be notified, and by when? For most of the modern era those steps were performed by human clerks moving paper, or, more recently, moving PDFs. In 2024 alone, state trial courts in the United States received an estimated 70 million new case filings, a four percent rise over the prior year even as filings sat 27 percent below their 2012 peak (National Center for State Courts). Behind that single number sits an ocean of repetitive procedural work that has historically swallowed clerk capacity whole.

What is changing now is the unit of automation. Earlier systems digitized documents; the emerging generation of agentic workflows automates the decisions about those documents. An agentic system is triggered by an event, a filing arrives, a deadline lapses, a hearing is set, then executes a multi-step task on its own, within fixed guardrails, escalating to a human when it is uncertain and writing every action to an audit trail. It is the difference between a faster typewriter and a tireless junior clerk who never loses a file but always knows when to knock on the judge's door.

The Old Way: Paper, Queues and the Tyranny of the Inbox

For decades, court administration ran on human throughput. A filing landed in a clerk's queue; a person checked it against procedural rules, keyed data into a case management system, assigned it to a calendar, and mailed notices. The system worked until volume outran headcount, and volume usually won. The strain is not abstract. In England and Wales, the Crown Court's open caseload climbed to a series peak of 79,619 cases by September 2025, up nine percent year-on-year, with more than 20,000 cases open for a year or longer for the first time on record and a median case age of 175 days (UK Ministry of Justice). The magistrates' courts hit their own peak of roughly 373,000 open cases, a 17 percent annual rise.

The U.S. federal picture is more uneven but no less revealing of where the pressure points sit. In the twelve months ending March 2025, bankruptcy filings rose 13 percent to 529,080 and criminal filings climbed nearly 12 percent to 73,644, even as civil filings fell sharply (Administrative Office of the U.S. Courts). Whatever the trend line, the clerical burden each case generates is stubbornly fixed, and when one category surges, the queue behind it lengthens.

The deeper problem was never just speed; it was consistency. A human reviewer on a Friday afternoon does not apply the rules identically to one on a Monday morning. Notices were missed. Filings sat unaccepted for days. And critically, none of it left a structured, queryable record of why a given step was taken, the institutional memory lived in the heads of veteran clerks, and walked out the door when they retired.

The Shift: From Digitizing Paper to Automating Decisions

Courts have layered technology for half a century, typewriters, fax machines, electronic case management, e-filing, document storage, then dashboards and remote hearings (South Dakota Unified Judicial System). Optical character recognition and robotic process automation have been quietly in service for years. What is genuinely new is the agentic layer that sits on top: software that interprets intent, reasons over case data, and takes multi-step action with minimal human input.

The trajectory mirrors the broader enterprise software market. Research firm Gartner forecasts that 33 percent of enterprise software applications will embed agentic AI by 2028, up from less than one percent in 2024, enabling roughly 15 percent of day-to-day work decisions to be made autonomously (Process Excellence Network, citing Gartner). The same firm expects 70 percent of AI applications to rely on multi-agent systems by 2028 (Rocking Robots, citing Gartner). Independent market trackers put the agentic AI market on a path from roughly $7.8 billion in 2026 to about $52.6 billion by 2030, naming legal services among the fastest-growing end-user segments (Agent Market Cap).

Courts are not waiting for the broader market to mature. In Texas, a state court facing e-filing volumes its clerks could not clear by hand deployed automation to verify whether filings met minimum standards for acceptance. The model began at roughly 60 percent accuracy out of the box and, after staff fed it corrections, climbed to almost 95 percent, described as better than human performance (National Center for State Courts). Tellingly, the court eventually throttled the system to business hours only: it was so productive overnight that staff arrived to a 2.5-day backlog of approved filings awaiting their secondary review.

The constraint is no longer how fast the machine works. It is how fast humans can meaningfully supervise what the machine has already done.

What It Looks Like Now: Triggers, Guardrails and Trails

Strip away the branding and present-day court automation reduces to a recognizable pattern. An event triggers an agent; the agent runs a defined procedure; it acts only within explicit guardrails; it escalates anything ambiguous to a human; and it logs the whole sequence. The use cases already in production span the full clerical lifecycle.

The efficiency gains reported by individual jurisdictions are substantial. A finance bot in one court system cut invoice-processing labor from 16 hours a day across four staff to four hours for one. A small-claims review process dropped from 15 hours a week to two or three. An onboarding chatbot loaded with 150 civil procedures compressed staff training from two-to-three months to two-to-three weeks (South Dakota Unified Judicial System). In one Florida county's e-filing operation, automation reached the point where roughly 42 percent of filings are never reviewed by a human, with nine bots performing the work of about 45 full-time staff and processing four times the volume of a human clerk.

That separation of duties is exactly what the audit trail enforces. Where a retiring clerk once carried away the reasoning behind a routing decision, an agentic system records each step as structured data, what triggered it, what rule it applied, what confidence it held, and when it handed control back to a person. Oversight bodies increasingly treat this auditability as non-negotiable, alongside transparency about a system's limitations and margin of error (Guidelines for the Use of AI Systems in Courts and Tribunals, 2025).

The Next Few Years: Multi-Agent Dockets and the Guardian Layer

The near-term direction is toward orchestration. Instead of one bot per task, courts will run chains of cooperating agents, an intake agent handing to a docketing agent handing to a scheduling agent, coordinated end to end. Gartner's expectation that 70 percent of AI applications will use multi-agent systems by 2028 implies a court docket managed less like a filing cabinet and more like an air-traffic control system, with humans cleared to intervene at any point (Rocking Robots, citing Gartner).

Crucially, a second class of software is emerging to watch the first. Gartner forecasts that "guardian agents", systems built to review, monitor and constrain other agents, will capture 10 to 15 percent of the agentic AI market by 2030 (Rocking Robots, citing Gartner). For a judiciary, that maps neatly onto an institutional instinct: every actor in a courtroom is checked by another. An oversight agent that flags anomalies, halts drift and verifies outputs is the digital analogue of appellate review applied to administrative automation.

The optimism is bounded by a sobering forecast. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear value and inadequate risk controls, often because organizations automate workflows that were already broken (Reuters, citing Gartner). For courts, the lesson is pointed: agentic automation amplifies whatever process it inherits. Pointed at a clean, well-governed workflow, it multiplies capacity. Pointed at a tangled one, it multiplies error at machine speed.

Judicial bodies are codifying these guardrails faster than vendors can ship features. Guidance for judicial office holders warns that anything entered into a public AI tool should be treated as published to the world, that accuracy must be independently verified, and that judges remain personally responsible for any material produced with AI assistance (Courts and Tribunals Judiciary, UK). National court bodies are pairing such rules with practical readiness frameworks, guiding principles, internal use policies, data-governance assessments and literacy programs, so that automation arrives as governed infrastructure rather than improvised experiment (National Center for State Courts, AI Readiness for the State Courts).

The Bottom Line

Agentic workflows are not coming for the judge's gavel; they are coming for the clerk's queue. The trajectory is clear, from paper, to PDFs, to autonomous agents that triage filings, build dockets and dispatch notices around the clock while escalating the close calls and logging everything they touch. Done well, this is the most consequential operational upgrade courts have seen since electronic filing, with reported accuracy already edging past human review and labor freed for the work that genuinely requires a person (National Center for State Courts). Done badly, it scales dysfunction and erodes trust. The deciding variable is not the model. It is the discipline of the guardrails, the integrity of the audit trail, and the unwavering rule that humans, not agents, render judgment.

Sources

1. National Center for State Courts, Court caseload data (2024 state court filings)
2. Administrative Office of the U.S. Courts, Federal Judicial Caseload Statistics 2025 (Judicial Caseload Indicators)
3. UK Ministry of Justice, Criminal Court Statistics Quarterly, July to September 2025
4. National Center for State Courts, Improving document processing with automation
5. South Dakota Unified Judicial System, Artificial Intelligence in State Courts: Promise & Peril (2025)
6. Process Excellence Network, Gartner forecast: 33% of enterprise software with agentic AI by 2028
7. Rocking Robots, Gartner: guardian agents to capture 10 to 15% of agentic AI market by 2030; 70% multi-agent by 2028
8. Reuters, Over 40% of agentic AI projects will be scrapped by 2027, Gartner says
9. Agent Market Cap, Agentic AI Market Trajectory 2026 to 2030
10. Guidelines for the Use of AI Systems in Courts and Tribunals (2025)
11. Courts and Tribunals Judiciary (UK), Artificial Intelligence Guidance for Judicial Office Holders
12. National Center for State Courts, AI Readiness for the State Courts (2025)
13. UN OHCHR, Navigating AI in the Judiciary: Guidelines for Judges
14. National Center for State Courts, Judicial use of generative AI: Lessons learned