The Empty Chair Rehearsal

A bank executive walks into a regulatory examination having read every brief, memorized every figure, and rehearsed every answer in their own head. Then the lead examiner asks a question nobody anticipated, sharp, slightly hostile, circling back to a contradiction buried three answers earlier, and the careful preparation dissolves. This is the oldest problem in high-stakes testimony: you can study the material endlessly, but you cannot study the room. The pressure, the interruption, the follow-up that twists your own words against you, those only exist live, and live is precisely where you cannot afford to be learning.

That gap between studying and rehearsing is now being closed by a category of tools that legal teams are beginning to call agentic panels, AI systems that adopt the personas of the people a banker or in-house lawyer will actually face. A synthetic examiner who probes a capital-adequacy answer. A simulated opposing counsel who runs an aggressive cross-examination. An artificial bench of jurors who deliberate over a theory of the case and tell you, coldly, why they did not believe it. The premise is borrowed from aviation and surgery, two professions that long ago accepted a simple truth: you practice before you perform, against a realistic adversary, until the unexpected stops being unexpected.

The stakes for banking are unusually concrete. U.S. banking regulators issued 277 enforcement actions in 2024, a 34.5 percent jump from the 206 the year before, according to the Committee on Capital Markets Regulation's annual tally (Committee on Capital Markets Regulation). When a single examination or deposition can anchor a multibillion-dollar resolution, the marginal value of one more realistic rehearsal is enormous, and that is exactly the case agentic panels are built to make.

The Old Way: Rehearsing Without a Room

Traditional preparation for a regulatory hearing, a deposition, or trial testimony was constrained by three things in short supply: people, money, and candor. The most rigorous method, a full mock trial or mock examination staffed by trial consultants, surrogate jurors, and lawyers playing opposing counsel, works, but it is expensive and slow. Litigation consultants put the price of a well-designed mock trial anywhere from roughly $10,000 to $60,000 or higher, with some complex exercises running past $100,000 (IMS Consulting & Expert Services). Because of that cost, most matters never got one. A witness might instead get a single conference-room session with a colleague gamely impersonating a hostile examiner, useful, but limited by one human's imagination and stamina.

The forensic-psychology literature is blunt about how thin the evidence base for traditional witness preparation actually is: for years, only a handful of published studies, most conducted in laboratory settings, directly examined whether witness-preparation procedures worked, even as the practice consumed enormous billable hours (Boccaccini, in Behavioral Sciences & the Law). What that small literature did find was encouraging in one specific way: witnesses who participated in simulated questioning emerged measurably more confident, composed, and less anxious. The act of rehearsing the room, not just reviewing the facts, was the part that worked. The problem was that realistic rehearsal was the scarcest and costliest resource of all.

This scarcity collided with a structural shift in how disputes actually resolve. The American "vanishing trial" means almost everything is now settled or negotiated in the shadow of a hearing room rather than decided in open court. Federal civil cases resolved by trial fell from roughly 20 percent in 1938 to under 2 percent by 2002 and approximately 1 percent today, with the jury-trial disposition rate near 0.7 percent (Duke Judicature). Paradoxically, that makes preparation more valuable, not less: the deposition, the examination, the settlement posture forged under questioning is now where the matter is effectively won or lost, because the trial that would have tested it never comes.

The Shift: Borrowing the Practice-Partner Playbook

What changed is that realistic adversaries stopped being scarce. Generative systems can now sustain a coherent persona across a long, adaptive exchange, adjusting tone, pressing on weak answers, and improvising follow-ups that a static script never could. The training-science case for practicing against such partners is, by this point, well established. A 2025 meta-analysis spanning 12 studies and 907 participants found that role-play-based training produced an effect size of 0.82, a magnitude statisticians classify as "large", with the strongest gains in practical skill (Fu & Li, International Journal of Instruction, as summarized in industry analysis). Immersive simulation research has long pointed the same direction, with one widely cited study finding simulation-trained learners up to 275 percent more confident applying skills and able to train roughly four times faster than classroom peers (PwC analysis, via Training Journal).

The same forces are reshaping the legal profession's appetite for AI generally. In Thomson Reuters Institute's 2025 survey, active generative-AI use among legal organizations nearly doubled year over year to 26 percent, while 78 percent of law-firm respondents expected the technology to become central to their workflow within five years (Thomson Reuters Institute). Preparation and rehearsal, long the most stubbornly manual corner of litigation, is a natural place for that adoption to land, because it is where realistic practice was historically rationed by cost.

Banking is fertile ground for a specific reason: the industry already lives inside structured adversarial proceedings. Supervisory examinations, consent-order negotiations, enforcement interviews, and depositions in enforcement-driven litigation are recurring, high-frequency events. The Federal Reserve alone conducted 316 examinations of state member banks and nearly 2,900 inspections of bank holding companies in 2024, and completed 58 formal enforcement actions carrying $372 million in civil money penalties (Federal Reserve). Each of those interactions is, in effect, a performance, and each is a candidate for rehearsal.

Aviation does not let a pilot meet an engine failure for the first time at altitude. Banking should not let a general counsel meet a hostile examiner for the first time under oath.

What It Looks Like Now

Inside a bank legal or compliance team, an agentic-panel rehearsal today typically runs in four stages. It begins with scenario construction: the team loads the matter, the examination scope, the enforcement theory, the relevant filings, and configures the personas. A synthetic examiner might be instructed to probe anti-money-laundering controls the way regulators did in the landmark 2024 Bank Secrecy Act matters; a simulated opposing counsel might be tuned to an aggressive, repetitive deposition style. Next comes the live exchange, where the witness answers in real time and the AI persona adapts, pressing harder when an answer wavers, circling back to expose an inconsistency, shifting tone from solicitous to skeptical exactly as human adversaries do.

The third stage is where the technology earns its place: structured feedback. Because the system records and analyzes every exchange, it can surface patterns a human coach might miss across a long session, the witness who answers a yes/no question with a three-sentence speculation, the "yes train" of agreeable answers that opposing counsel exploits, the filler words and hesitations that read as evasion. Forensic-trial consultants describe exactly these failure modes, over-explaining, succumbing to flattery, failing to tolerate silence, as the behaviors that systematic, repeated mock-deposition drills are designed to extinguish (Courtroom Sciences). The fourth stage is iteration: because each run costs a fraction of a staffed mock exercise, the witness can repeat the proceeding ten times across two weeks rather than once, and spaced, repeated practice is precisely what the witness-preparation literature identifies as most effective (Law Journal Newsletters).

The synthetic-juror use case is no longer purely theoretical. In November 2025, a University of North Carolina School of Law experiment staged a mock trial in which student advocates argued before a "jury" of AI systems, which processed the arguments, applied the jury instructions, deliberated, changed their positions, and returned a verdict, an early but striking demonstration that AI personas can act as a deliberating audience for testing a theory of the case (UNC School of Law). For a bank weighing whether to contest an enforcement action or settle, the ability to run its narrative past a synthetic panel, cheaply, repeatedly, before committing, is the kind of low-stakes rehearsal that was previously available only to the best-funded matters.

The economics explain the pull. A single staffed mock proceeding sits in the five-figure range; an AI-persona rehearsal can be run for a marginal cost closer to that of software, which collapses the price of an additional repetition toward zero. That is the same logic immersive-training research has documented in other professions, where simulation became dramatically more cost-efficient than classroom or live methods as the number of practice sessions scaled (TechClass, summarizing PwC cost analysis).

The Next Few Years: Rehearsal as Standard Operating Procedure

Over the next three to seven years, the trajectory points toward AI-persona rehearsal becoming a routine step in matter preparation rather than a novelty reserved for bet-the-bank cases. Expect persona libraries calibrated to specific supervisory styles and enforcement postures, so a bank can rehearse against a panel tuned to the questioning patterns of the regulator it actually faces. Expect rehearsal to extend upstream into continuous readiness, compliance officers and front-line bankers practicing examination scenarios well before any matter is open, the way pilots log recurrent simulator hours. And expect the synthetic-juror panel to mature into a standard tool for stress-testing whether an enforcement theory or litigation narrative will survive contact with a lay audience.

But the most important development will be discipline, because the same realism that makes these systems useful makes them dangerous when over-trusted. The first risk is a realism ceiling. A synthetic examiner is a model of an examiner, not the real one; it can miss the idiosyncratic question, the institutional memory, or the human chemistry that decides a hearing. A synthetic juror has no lived experience to draw on and may reason in ways that look plausible but diverge from how twelve real people in a box actually behave. Treating either as a faithful oracle, rather than a credible sparring partner, invites a false sense of readiness.

The deeper risk is automation bias, the well-documented human tendency to over-rely on a machine's output simply because it arrived with apparent authority. Research on automated decision aids has found not only that people defer to flawed system recommendations, but that adding explanations sometimes increased that misplaced trust rather than tempering it (Vered et al., University of Melbourne). A witness who rehearses only against a predictable AI persona may be lulled into a script, fluent against the machine, brittle against the surprise the machine never generated.

Banking is, fortunately, structurally suited to impose that discipline. The industry already subjects its quantitative models to formal validation, independent review, and ongoing monitoring under long-standing supervisory expectations for model risk management. A culture that demands a bank demonstrate why it trusts a credit model will, almost reflexively, ask the same of a system that rehearses its witnesses, scoping where the persona is realistic, where it is not, and where human judgment must take over.

Conclusion: Practice Before You Perform

The deeper change is cultural. For most of its history, banking's legal function prepared for its most consequential moments by studying rather than rehearsing, reading the briefs, memorizing the answers, and hoping the room behaved. Agentic panels invert that, importing into law the practice-partner discipline that aviation, surgery, and the military adopted long ago: you do not meet the worst-case scenario for the first time when it is real. You meet it a dozen times in a simulator first, against an adversary good enough to make the rehearsal honest.

None of this removes the lawyer, the witness, or the judgment from the loop, and it should not. The synthetic examiner proposes the hard question; the human decides how to answer it, and whether the rehearsal taught the right lesson. But for an industry that faces a rising tide of examinations and enforcement actions, where a single deposition can shape a billion-dollar resolution and where the trial that would once have tested everything almost never comes, the ability to fill the empty chairs and practice the room, cheaply, repeatedly, before it counts, may prove to be one of the quietest and most consequential upgrades to banking's legal craft. The rehearsal has begun. The institutions that take it seriously will walk into the real room having already been there.

Sources

1. Committee on Capital Markets Regulation, Annual Enforcement Data Report 2024 (banking-regulator actions, mean sanctions). capmktsreg.org
2. IMS Consulting & Expert Services, "What Does a Mock Trial Cost?" imslegal.com
3. Duke Judicature, "Going, Going, But Not Quite Gone: Trials Continue to Decline." judicature.duke.edu
4. M. Boccaccini, "What do we really know about witness preparation?" Behavioral Sciences & the Law (PubMed). pubmed.ncbi.nlm.nih.gov
5. Zenobits, "AI Roleplay Training: The Evidence" (Fu & Li 2025 meta-analysis, effect size 0.82). zenobits.co.uk
6. Training Journal, "Practice makes profitable: how immersive AI roleplay drives productivity" (PwC immersive-learning findings). trainingjournal.com
7. Thomson Reuters Institute, 2025 Report on the State of the US Legal Market (generative-AI adoption). thomsonreuters.com
8. Federal Reserve, 2024 Supervision and Regulation Report (examinations, enforcement actions, civil money penalties). federalreserve.gov
9. Courtroom Sciences, "Science-Based Witness Preparation Strategies That Actually Work." courtroomsciences.com
10. Law Journal Newsletters, "Preparing a Medical Witness for Deposition and Trial: A Different Approach" (spaced practice). lawjournalnewsletters.com
11. UNC School of Law, "AI Jury Finds Teen Not Guilty in Mock Trial." law.unc.edu
12. TechClass, "Transform Corporate Training: Role-Play Simulations" (PwC cost-efficiency analysis). techclass.com
13. Vered et al., "The effects of explanations on automation bias," University of Melbourne. psychologicalsciences.unimelb.edu.au
14. Fenergo, "Regulatory penalties in North America account for 95% of global financial penalties in 2024." fenergo.com
15. SEAK, Inc., Expert Witness Fee Study. seak.com