The Probabilistic State: When Government Lawyers Start Rolling the Dice on Purpose

For most of the modern administrative era, a government lawyer deciding whether to settle or fight a case relied on a tool no spreadsheet could capture: judgment. A seasoned litigator in an agency general counsel's office or a state attorney general's bureau would read the file, weigh the judge, recall the last dozen matters like it, and offer a verbal estimate, "we'll probably win, but it could get expensive." That estimate, however experienced, was a single point on a curve that nobody ever drew. Today a quietly powerful idea is reshaping that calculus: instead of guessing the most likely outcome, public legal offices are beginning to simulate the entire range of outcomes, assign probabilities to each, and let the numbers inform strategy.

The technique is not new to finance or engineering, where Monte Carlo simulation, running thousands of randomized scenarios to map a distribution of possible results, has been standard practice for decades. The U.S. Government Accountability Office has long recommended such probabilistic methods for federal cost estimating, treating uncertainty as something to be modeled rather than ignored (GAO Cost Estimating and Assessment Guide). What is new is the migration of that mindset into litigation strategy, and, increasingly, into the public sector, where the stakes are measured not in shareholder value but in taxpayer dollars and public trust.

The Old Way: Judgment Without a Distribution

The legacy of government litigation strategy is a story of incomplete information. Agencies historically did not even track what their cases cost. When the GAO examined wetlands "takings" claims in the early 1990s, it found that the Department of Justice and the Army Corps of Engineers simply did not keep data on the government's litigation costs, leaving the government's potential liability in pending cases impossible to determine (GAO, Clean Water Act takings review). A legal office cannot model what it does not measure.

The pattern persisted for decades. A landmark GAO review of environmental litigation found that across fiscal years 1995 through 2010, the Department of Justice defended EPA in roughly 2,500 cases, an average of about 155 a year, at a total cost of about $43 million, or roughly $3.3 million annually, with the Treasury Judgment Fund paying prevailing plaintiffs about $14.2 million more (GAO-11-650, Environmental Litigation). Those figures were only assembled after the fact, by auditors reconstructing the past, not by strategists forecasting the future.

Three structural features defined the old way. First, decisions were point estimates: a lawyer's "we'll likely win" hid enormous variance. Second, settlement-versus-trial choices were made case by case, with little portfolio view across an agency's docket. Third, the public sector's distinctive incentive, the institutional pull toward making a problem disappear, went unmeasured. A 2026 congressional inquiry found that when federal adverse-action cases were not dismissed at the Merit Systems Protection Board, agencies chose to pay and settle 68% of the time, even though when agencies actually proceeded to dispute claims they won more than 80% of decisions (U.S. House Committee on Oversight).

The Shift: Analytics Becomes Table Stakes

Outside government, data-driven litigation moved from novelty to norm in under a decade. Industry surveys show roughly seven in ten legal professionals now use some form of legal analytics, 68% in 2024, up sharply from a small fraction in 2018, with 100% of users calling the insights valuable and 80% reporting that clients now require or expect them (Lex Machina 2024 Legal Analytics Survey). The same survey found that successful litigation outcomes (70%) and improved efficiency (69%) were the top drivers of adoption.

The intellectual engine beneath these tools is decades old. Litigation risk analysis builds a decision tree of every contested issue, assigns each a probability, and computes an expected value by weighting outcomes by their compound likelihood (Litigation Risk Analysis, ACC chapter). When uncertainties are continuous rather than binary, analysts layer Monte Carlo simulation on top, running thousands of randomized trials to produce a full distribution of recoveries and exposures rather than a single deterministic number (Cornerstone Research). In one widely cited valuation, decision-tree-plus-Monte-Carlo analysis pegged the expected value of a single lawsuit at $457.7 million (Valuation of Litigation, Cogent Valuation).

The shift is not from intuition to certainty. It is from a single guess to a labeled distribution, from "we'll probably win" to "we win 78% of the time, and the bad 22% costs us this much."

Government is now riding the same wave that drove broader AI adoption across agencies. Federal AI use cases reported across agencies grew to 3,611 across 56 agencies in 2025, up from roughly 2,133 across 41 agencies a year earlier (ExecutiveGov, on OMB's inventory). The GAO documented the acceleration in detail: across 11 selected agencies, total reported AI use cases nearly doubled from 571 in 2023 to 1,110 in 2024, while generative-AI use cases jumped about nine-fold, from 32 to 282 (GAO-25-107653). Analytical and predictive tooling is no longer exotic inside the public sector, it is a budget line.

What It Looks Like Now: Four Public-Sector Use Cases

In practice, outcome-simulation engines are being pointed at four distinct government problems. None requires a courtroom crystal ball; each turns scattered docket history and case facts into a probability distribution a decision-maker can actually use.

1. Litigation outcome prediction for agency cases

Predictive models trained on docket data estimate the likelihood of procedural events, a motion to dismiss being granted, summary judgment, time-to-disposition. Independent reviewers note these tools are strongest on procedural outcomes, where narrow questions like "will this motion be granted?" reach the 80 to 90% accuracy range, and weakest on ultimate jury verdicts, which most systems decline to predict (Accumulated, Litigation Forecasting). For an agency defending a high-volume docket, even reliable procedural forecasting reshapes triage.

2. Settlement-versus-trial modeling

This is where simulation earns its keep. A decision tree converts each contested issue into a probability and a dollar consequence; Monte Carlo runs roll those uncertainties thousands of times to produce an expected value, a most-probable result, and the full range of gain and loss (TreeAge, on decision-tree modeling). Set against a settlement demand, that distribution tells a government lawyer whether the offer sits above or below the case's modeled value.

3. Enforcement scenario comparison

Regulators choosing among enforcement paths, administrative action, civil suit, negotiated consent, can model each as a branch with its own probability of success, cost, and timeline. GAO's environmental-litigation data shows why this matters: cases against EPA clustered heavily under the Clean Air Act (59%) and Clean Water Act (20%), and lead plaintiffs ranged from trade associations (25%) to national environmental groups (14%), each pattern carrying a different expected cost (GAO-11-650).

4. Public-resource allocation

Aggregated across a docket, simulated case values let an office allocate scarce attorney hours toward the matters with the highest expected return, a portfolio view rather than a case-by-case scramble. It is the litigation analogue of the probabilistic budgeting GAO already endorses for federal programs (GAO-20-195G).

The volumes involved make the case for systematization. Federal district courts saw 271,802 civil cases filed in the year ending in 2025 and terminated 507,326 (U.S. Courts, Judicial Caseload 2025). At the state level, a study of 925,344 disposed civil cases, about 5% of the national caseload, found three-quarters of judgments were $5,200 or less and that at least one party was self-represented in more than three-quarters of cases (State Justice Institute / NCSC). High-volume, low-value, lopsided dockets are precisely the environment where modeling beats intuition.

The Next Few Years: Distributions, Accountability, and the Limits of Prediction

Over the next three to seven years, expect outcome simulation to become a routine input, not the decider, in public legal strategy. The trajectory of federal AI adoption suggests the infrastructure is arriving fast: with mission-enabling functions accounting for 61% of reported generative-AI use cases and government services another 15% (GAO-25-107653), legal offices will inherit analytical tooling already normalized elsewhere in their agencies.

But the public sector cannot simply import the corporate playbook. A government that quantifies case value invites a hard question: if a model says you will probably win, is settling a prudent risk decision or an abdication? The same dynamic cuts the other way, pursuing a case the data says you will likely lose wastes public money. Simulation makes both judgments visible, and visibility is double-edged. Scholars warn that when public bodies outsource judgment to algorithms without transparency, the result can be bias, opacity, and accountability gaps that erode public trust (AI Now Institute, Algorithmic Accountability Policy Toolkit).

Three guardrails will define responsible use. First, prediction uncertainty must be reported, not hidden. A point estimate of "72% win probability" is less honest than a confidence interval; models trained on thin or skewed docket history can be confidently wrong. Second, human-in-the-loop accountability must remain, explainability, auditing, and a named official who owns the decision are emerging as the consensus safeguards for algorithmic governance in administrative law. Third, the model's incentive must be examined. A simulation tuned to minimize short-term cost could quietly entrench the very over-settlement reflex the Oversight Committee flagged.

Conclusion: The Honest Number

The promise of outcome simulation in government is not that machines will out-lawyer attorneys. It is that the distribution, the labeled, defensible range of what could happen, replaces the silent point estimate that has governed public litigation for a century. Done well, it tells an attorney general's office when an 80% win rate makes a settlement look like a giveaway, and when a confident-sounding case is actually a coin flip. Done badly, it dresses guesswork in the authority of math. The difference will come down to whether public legal offices treat these engines as oracles or as instruments, and whether they have the discipline to publish the uncertainty along with the prediction. The probabilistic state is arriving either way. The only choice is whether it arrives accountable.

Sources

1. U.S. Government Accountability Office, GAO-25-107653, "Artificial Intelligence: Generative AI Use and Management at Federal Agencies" (July 2025).
2. Akin Gump, summary of GAO report on Generative AI Use and Management at Federal Agencies (2025).
3. ExecutiveGov, "Federal AI Use Cases Surge Past 3,600 as Agencies Scale Adoption" (April 2026), on the OMB Federal Agency AI Use Case Inventory.
4. U.S. House Committee on Oversight and Government Reform, settlement-rate inquiry (May 2026).
5. U.S. Government Accountability Office, GAO-11-650, "Environmental Litigation: Cases against EPA and Associated Costs over Time."
6. U.S. Government Accountability Office, "Clean Water Act" takings-claims review, RCED-93-176FS (1993).
7. U.S. Government Accountability Office, GAO-20-195G, "Cost Estimating and Assessment Guide."
8. Lex Machina, "2024 Legal Analytics Survey."
9. Accumulated, "Litigation Forecasting: Can Analytics Predict Case Outcomes?" (2023).
10. Marc B. Victor, "Evaluating Legal Risks and Costs with Decision Tree Analysis," Association of Corporate Counsel.
11. Cornerstone Research, "Applying Monte Carlo Simulations in Litigation."
12. Kam, Reinemann, Puiggali & Ruiz, "Valuation of Litigation," Cogent Valuation reprint.
13. TreeAge Software, decision-tree and litigation-risk modeling overview.
14. Administrative Office of the U.S. Courts, "Federal Judicial Caseload Statistics 2025."
15. State Justice Institute / National Center for State Courts, "The Landscape of Civil Litigation in State Courts."
16. AI Now Institute, "Algorithmic Accountability Policy Toolkit."