Rehearsing the Verdict Before the Jury Is Sworn

For a hospital sued over a missed diagnosis, the trial is the part nobody gets to rehearse with real stakes. The physician on the stand has testified perhaps once before. The defense theory has been tested only against the lawyers who wrote it. And the first time anyone hears how a skeptical juror reacts to a complex causation argument is often the moment a verdict form is being filled out. Healthcare litigation has always been high-consequence improvisation. A new class of tools, agentic AI personas that can play the judge, the plaintiff's expert, the cross-examiner, and the deliberating juror, is quietly turning that improvisation into rehearsal.

The stakes explain the appetite. The average U.S. medical malpractice settlement in 2024 sat near $425,000 per the National Practitioner Data Bank, while cases that reach a jury verdict average roughly $1.1 million in compensation. Yet only about 21% of malpractice cases that reach a jury end in a plaintiff verdict, and the path to that courtroom can take two to five years from injury to resolution. The thin margin between winning and losing, multiplied by million-dollar exposure, is precisely the environment where better rehearsal pays for itself.

The Old Way: Mock Trials Only the Wealthy Could Afford

For decades, the gold standard for testing a case before trial was the live mock trial or focus group, a logistically heavy production that priced out all but the largest matters. Jury consultants recruit 25 to 35 mock jurors to form three deliberating panels, paying each participant $150 to $300 a day, with juror compensation alone running upward of $10,000 for a two-day exercise. A full mock trial unfolds over two or three days and demands weeks of attorney preparation.

The all-in price reflected that intensity. Trial-consulting practices commonly quote ranges from a few thousand dollars to $60,000 or $100,000, with most exercises landing between $10,000 and $25,000. Sophisticated focus groups and mock juries routinely cost $8,000 to $30,000 or more, a sum that, as practitioners openly admit, is "not economically feasible" for the modest-value cases that make up most dockets.

The consequence was structural inequity in preparation. A health system facing a catastrophic birth-injury claim could afford to test its narrative against real human reactions; the same system facing a $300,000 surgical-error claim almost never could. And even when the budget existed, a mock trial offered exactly one rehearsal. You could not iterate. You could not ask the same juror panel to react to three different versions of your causation argument. You ran the exercise once, absorbed the feedback, and walked into the real courtroom hoping the lesson transferred.

The Shift: Adoption Outruns the Skepticism

Generative AI moved from novelty to working tool inside the legal profession in roughly two years. According to the Thomson Reuters Institute's 2025 survey, 26% of legal organizations were actively using gen AI in 2025, nearly double the 14% recorded in 2024. While only 15% of law-firm respondents said the technology was central to their workflow today, 78% expected it to become central within five years. Among small and solo firms, adoption reportedly jumped from 27% in 2023 to 53% in 2024. By the following year, one industry report found nearly seven in ten legal professionals using gen AI tools for work.

That broad momentum has spilled directly into trial preparation. The same large language models that summarize documents and draft briefs can be instructed to inhabit a persona, and to hold it under pressure. A growing body of computer-science research, including a 2024 survey of more than a hundred studies, traces how role-playing with language models evolved from simple persona consistency to complex character-driven simulation involving behavioral alignment and sustained characterization. The leap that matters for litigators is that a model can now be told to be a hostile plaintiff's expert with thirty years in obstetrics and a grudge, then made to answer the very questions your physician will face.

Early efficacy signals from structured deposition-rehearsal pilots are striking. In one five-week beta of an AI deposition simulator, participants logged more than 160 hours of testing, with 97% strongly agreeing the tool was valuable for litigation training and 94% saying they would use it again. Crucially, structured feedback, scored against rubrics covering questioning technique, exhibit handling, and witness management, began streaming back within minutes of a session concluding, a turnaround impossible with human panels.

What It Looks Like Now: A Day of Synthetic Practice

In practice, an agentic-panel workflow assembles a roster of AI personas around a single case file. A defense team preparing a healthcare matter might convene four distinct synthetic roles, each grounded in the record and tuned to behave like its real-world counterpart.

The defendant physician can sit for a simulated deposition at 6 a.m. and a second one at noon, facing a witness persona that adapts its line of questioning each time. The team can run an opening statement past a synthetic juror panel, rewrite it, and run it again, something the economics of live focus groups never permitted. Because the marginal cost of an additional rehearsal approaches zero, the constraint shifts from money to judgment about how many iterations are genuinely useful.

The timeline data underscores why teams want to rehearse early and often. Across large carrier datasets, roughly 60 to 70% of claims close within two to three years, only about 5 to 10% reach trial, but the cases that do go the distance frequently last four to seven years. A long-running matter accumulates expert depositions, evolving theories, and witness fatigue, exactly the conditions under which low-cost, repeatable rehearsal compounds in value.

The Limits: Realism, Bias, and the Over-Reliance Trap

The technology's most seductive claim, that a synthetic persona behaves like a real human, is also its weakest empirical link. A peer-reviewed study quantifying the "persona effect" found that across most subjective datasets, persona variables explained less than 10% of the variance in human responses, with a large share of behavior driven by factors a persona prompt cannot capture. A synthetic juror is a plausible juror, not a representative one. Teams that mistake a fluent performance for a calibrated prediction of how twelve real strangers will deliberate are testing against a flattering mirror.

Bias is the second hazard. Research on persona-assigned models shows they can embed and amplify stereotypes when generating human-like characters, meaning a synthetic juror panel may systematically misjudge how real demographic groups would react, a serious risk in cases where jury composition is decisive. Practitioners are experimenting with mitigation, such as multi-persona frameworks designed to reduce social bias, but no technique fully neutralizes the problem.

The third and most immediate danger is over-reliance. The legal profession has already learned this lesson painfully through AI "hallucinations." Courts have sanctioned attorneys repeatedly for filings containing fabricated citations, with one federal court in Oregon fining an attorney $15,500 in December 2025 and a federal appeals court ordering another to pay $2,500 over hallucinated authorities in early 2026. The same credulity that pastes a fake case into a brief can lead a team to trust a synthetic expert's confident-but-wrong rehearsal of the science. Regulators are responding: a proposed Federal Rule of Evidence 707 would subject machine-generated evidence to the same reliability standards as expert testimony.

The Next Few Years: From Practice Room to Standard of Care

The trajectory points toward agentic rehearsal becoming a default rather than a luxury. With 78% of firms expecting gen AI to be central within five years, the question for healthcare defense practices is shifting from whether to rehearse synthetically to how to do it defensibly. Three developments look likely.

Democratization of jury research. The same exercise that once cost $25,000 and was reserved for catastrophic claims becomes feasible for the routine $300,000 matter, narrowing the preparation gap between well-resourced systems and everyone else. Expect synthetic panels to handle first-pass case assessment, with expensive human focus groups reserved for the highest-stakes trials as validation.

Governance hardens. As courts formalize standards through measures like proposed Rule 707 and as sanctions for AI misuse climb, firms will adopt internal protocols requiring human verification of any synthetic output that touches the record. Roughly 41% of firms already report having gen AI policies; rehearsal-specific governance will follow.

Validation against ground truth. The most credible providers will be those who can show their synthetic juries track real verdict patterns. Research already demonstrates that, in favorable settings, a well-prompted model can capture around 81% of the annotation variance achievable by a model trained on real human data, evidence that calibration is possible, but only where the underlying behavior is predictable and the prompts are rich.

Conclusion: Practice Before You Perform

Healthcare litigation will remain a high-stakes performance staged in front of strangers who decide a doctor's reputation and a hospital's balance sheet. What has changed is that the performance no longer has to be the first run-through. Agentic panels let defense teams rehearse the cross-examination, stress-test the causation story, and watch a synthetic jury deliberate, repeatedly, cheaply, and early enough to matter. The technology cannot tell you what real jurors will do, and the teams that forget that will be punished for it. But used as rehearsal rather than prophecy, it converts the most expensive improvisation in the legal world into something a careful litigator can finally practice. In a domain where the plaintiff win rate at trial hovers near 21% and the average verdict tops a million dollars, the value of a second take is hard to overstate.

Sources

1. CostSignals, Medical Malpractice Claims Guide 2026: Process, Costs & Timeline (NPDB-based plaintiff win rate, verdict averages, costs)
2. Brown & Barron, Average malpractice settlement per NPDB ($425,000 in 2024)
3. Opveon, Selecting the Best Form of Jury Research for Your Case & Budget (mock juror pay, panel sizes)
4. Trial Dynamics, FAQs on focus group and mock trial costs
5. Aitken Law, Focus Groups on a Shoestring Budget (cost ranges)
6. LawSites, Thomson Reuters Institute 2025 Generative AI in Professional Services Report (adoption rates, policies)
7. LawSites, 2026 Legal Industry Report: AI adoption more than doubled
8. arXiv, "The Oscars of AI Theater: A Survey on Role-Playing with Language Models"
9. LawSites, AI deposition simulator beta pilot results (160+ hours, 97% / 94% feedback)
10. ACL 2024, "Quantifying the Persona Effect in LLM Simulations"
11. arXiv, "The Personality Trap: How LLMs Embed Bias When Generating Human-Like Personas"
12. arXiv, "Multi-Persona Thinking for Bias Mitigation in Large Language Models"
13. Bloomberg Law, AI-Faked Cases Become Core Issue Irritating Overworked Judges (sanctions data)
14. Reuters, US appeals court orders lawyer to pay $2,500 over AI hallucinations
15. Sterne Kessler, AI Hallucinations review and proposed Federal Rule of Evidence 707
16. Residency Advisor, Malpractice litigation duration and timelines (CRICO/PIAA-based)