Rehearsing the Tribunal Before the Tribunal Convenes

When a shipment of fertiliser fails to clear port, a long-term metals contract is repudiated, or a defective component triggers a downstream recall, the resulting dispute rarely ends in a courtroom drama. It ends in arbitration, often in another country, under rules the parties chose years earlier, before a tribunal that has read thousands of pages and will issue a binding award worth millions. The stakes are enormous, the dress rehearsal almost nonexistent. That asymmetry is exactly what a new category of agentic AI tools is now trying to close.

These systems assemble what practitioners increasingly call "panels": AI personas that play the arbitrator, the cross-examining counsel, the hostile fact witness, and the technical expert. Counsel runs a mock proceeding against them, pressure-testing a theory of breach, rehearsing a deposition, or stress-checking the argument that will eventually be put to an international tribunal. The premise is borrowed from aviation and surgery: you practise the high-consequence performance under realistic conditions until it stops surprising you.

The numbers explain the urgency. Legal fees, billed by the hour, and rising fast, make up roughly 83% of the total cost of an international arbitration, and a leading institution reported that cases concluding by final award in 2023 ran an average of 27 months. Meanwhile the value at stake keeps climbing: that same institution closed 2024 with US$354 billion in pending caseload value, the highest ever recorded. Every hour a team spends rehearsing cheaply against a machine is an hour it does not burn at tribunal rates.

The Old Way: Expensive Rehearsals for an Expensive Performance

Mock proceedings are not new. Trial consultants have run mock juries, mock arbitrations, and witness-preparation sessions for decades, and the discipline is genuinely effective, but it has always been gated by cost. A full-scale mock exercise has historically been quoted anywhere from US$10,000 to US$60,000 or higher, which confines it to the largest, highest-value disputes. For the typical mid-market supply-agreement breach, a category that commonly settles in the US$50,000 to US$300,000 range, a five-figure rehearsal simply does not pencil out.

So most teams improvised. A senior partner would impersonate the opposing advocate over a conference-room table; a junior associate would read the arbitrator's likely questions off a legal pad. Witnesses were prepped in a handful of sessions, often without the adversarial heat they would face under genuine cross-examination. The rehearsal that mattered most, testing whether an argument survives contact with a sceptical decision-maker, was the one most often skipped, precisely because the people qualified to deliver it were the most expensive in the building.

That left a structural gap in supply-chain litigation, where disputes are unusually fact-dense. Commodities, transport, and sale-of-goods contracts dominate cross-border caseloads: in one London institution's 2023 report, transport and commodities alone accounted for 36% of arbitrations, with sale-of-goods agreements the single most common contract type. These cases turn on quality specifications, delivery windows, force-majeure clauses, and chains of correspondence, the kind of granular factual record that rewards relentless rehearsal and punishes the unprepared witness.

The Shift: When the Opposing Counsel Is a Model

Two developments collided to make AI-persona panels viable. The first is professional appetite. Adoption of generative AI inside legal organisations roughly doubled in a year, from 14% actively using it to 26%, according to a 2025 survey of 1,702 professionals, with 78% of law-firm respondents expecting the technology to become central to their workflow within five years. The second is acceptance inside the arbitration community itself: in the 2025 Queen Mary University of London and White & Case survey of more than 2,400 respondents, 90% expect to use AI for research, analytics, and document review, with time savings (54%) and cost reduction (44%) the leading motivations.

The breakthrough is not that a model can write a brief. It is that a model can argue back, playing the arbitrator who interrupts, the witness who evades, the advocate who finds the weak seam in your timeline.

That adversarial capability rests on a research foundation that matured quickly. Academic work now treats role-playing as a first-class use of language models: a 2024 survey of the field formalised the distinction between assigning a persona to a model and the model adapting to a user's persona, mapping how the same system can credibly inhabit a judge, an examiner, or a witness role. The related "model-as-judge" literature is where the realism debate gets serious. Researchers have shown that evaluation frameworks can be engineered to guarantee over 80% agreement with human judges on selected instances, but only with deliberate confidence-gating, and only because raw model judgments are otherwise inconsistent.

The growth is real and it is broad. Across the leading institutions, 2025 brought near-record filings, one Singapore-based centre saw new cases jump roughly 42% year on year to its second-highest total ever, while a Hong Kong centre rose around 10%. Tariff turbulence and supply-chain shocks are a named driver: as contracts face delayed performance and questions of commercial impracticability, analysts expect tariff-related disputes to keep multiplying. More cases, larger sums, and tighter margins on legal spend together create exactly the conditions in which a cheap, repeatable rehearsal tool finds a market.

Those forum figures, drawn from each institution's own published statistics and a Hughes Hubbard comparative review, show why one rehearsal template never fits all. A documents-only expedited proceeding at the ICDR, where 72% of cases settle before an award, demands a very different prep posture from a US$100-million-plus contested hearing at the ICC, where median duration stretches to 22 months and witness testimony can decide the outcome.

What It Looks Like Now: Three Live Workflows

In practice, agentic panels are clustering around three supply-chain use cases that share one trait, high consequence, scarce rehearsal time.

1. Cross-border arbitration dry runs

A team facing a commodities tribunal can have the system role-play the chair's procedural questions, the opposing advocate's attack on causation, and the tribunal's likely scepticism toward a force-majeure defence. Because the model is conditioned on the actual contract and the chosen rules, the rehearsal surfaces the questions that genuinely sting, and lets counsel revise the argument before the only audience that matters hears it.

2. Deposition and witness prep for product-liability and breach disputes

Product-liability exposure is a defining anxiety for manufacturers, and the costs are sobering: U.S. product-liability MDLs held more than 197,000 pending cases as of early 2026, with the top ten class actions reaching US$17.9 billion in combined 2025 settlements. AI personas let a quality engineer or a procurement manager rehearse hostile cross-examination repeatedly, learning to handle the leading question, the document ambush, and the silence-after-the-answer, without exhausting a senior partner's billable hours.

3. Argument-testing before international tribunals

Before a position is committed to a brief, counsel can have a synthetic "tribunal" probe it for internal contradiction or precedent it ignores. This is the model-as-judge use case, and it is where discipline matters most: the value is in flagging weaknesses for a human to evaluate, not in trusting a synthetic verdict.

The Realism Limit: Where Over-Reliance Bites

The same survey that shows overwhelming appetite also shows where the community draws its line. Concern about AI errors and bias, the hallucination problem, was the leading deterrent at 51%, followed by confidentiality risk (47%), inexperience (44%), and regulatory gaps (38%). Practitioners broadly accept AI for procedural and administrative tasks, but only 23% approve of its use in drafting the legal-reasoning portions of awards. The message is consistent: simulate the rehearsal, never the verdict.

The consequences of ignoring that line are now documented in real courtrooms. Tracking projects have catalogued a fast-growing roster of cases in which lawyers filed AI-fabricated citations; in one 2026 matter a U.S. appeals court ordered a lawyer to pay US$2,500 over hallucinations in a brief. A mock panel that invents a helpful precedent, or a synthetic arbitrator that rewards a flawed argument, is dangerous precisely because it is persuasive. The realism that makes the rehearsal valuable is the same realism that can lull a team into mistaking a convincing simulation for a reliable prediction.

There is also a quieter risk specific to supply-chain matters: confidentiality. These disputes carry pricing formulas, supplier identities, and trade-secret specifications. Feeding that record into a persona system without proper safeguards can itself create exposure, which is why the same 47% confidentiality concern shows up across the profession, and why governance is becoming the gating factor for adoption rather than capability.

The Next Few Years: From Novelty to Standard of Care

The trajectory points toward normalisation. As caseloads and dispute values keep rising, and as expedited, documents-only procedures spread, compressing timelines further, the economic logic of cheap, repeatable rehearsal only strengthens. Expect AI-persona prep to migrate down-market, from the marquee US$100-million dispute to the mid-tier supply-agreement breach that could never justify a staffed mock exercise. The survey community already anticipates new roles dedicated to managing and implementing AI (cited by 40% of respondents), a sign that this is becoming infrastructure rather than experiment.

Three shifts look likely over the next three to seven years. First, disclosure norms will harden: institutional guidance is converging on the expectation that material AI use be disclosed, and rehearsal tools will need audit trails by default. Second, persona realism will improve to the point where the bottleneck is no longer believability but verification, keeping the human check fast enough to keep pace with the synthetic argument. Third, and most consequentially, declining to rehearse may start to look like a lapse: when a credible, low-cost dry run is available, choosing to send an unprepared witness into a US$50-million cross-examination becomes harder to defend.

The destination is not AI that decides cases. It is AI that makes sure no advocate, and no witness, ever walks into a hearing for the first time.

That is the honest promise of agentic panels in supply-chain disputes: not a shortcut around the tribunal, but a flight simulator for the people who must face it. The disputes are not getting cheaper, faster, or simpler, pending values are at record highs and average proceedings still run more than two years. What is changing is that the rehearsal, once the privilege of the best-funded cases, is becoming available to all of them. Practice before you perform, as the oldest professionals in high-stakes work have always known, is not a luxury. It is the standard of care catching up with the technology.

Sources

1. Burford Capital, Reducing time and costs in international arbitration (ICC duration and cost-composition data)
2. ICC, Dispute Resolution Statistics 2024 (US$354 billion pending caseload value)
3. ICC, 2024 Dispute Resolution Statistics (PDF; duration figures)
4. Oxford University Press, Empirical Findings on International Arbitration (cost composition, 221 ICC awards)
5. IMS Legal Strategies, What Does a Mock Trial Cost?
6. Yates Anderson, Average Supply Agreement Breach Settlement Amounts, 2025 to 2026
7. Reed Smith, LCIA Annual Casework Report 2023 (transport & commodities share)
8. LawSites / Thomson Reuters Institute, 2025 Generative AI in Professional Services Report
9. Herbert Smith Freehills Kramer, 2025 QMUL / White & Case International Arbitration Survey
10. King & Wood Mallesons, 2025 QMUL Survey: Lessons for the Asia-Pacific
11. ACL Anthology, Two Tales of Persona in LLMs: A Survey of Role-Playing and Personalization
12. arXiv, LLM Judges with Provable Guarantees for Human Agreement
13. Daily Jus, Where the Arbitration Market Is Heading: 2024 to 2025 Institutional Statistics
14. Hughes Hubbard & Reed, Comparing Timelines: Statistics on Arbitration Duration
15. Sam & Ash Law, Product Liability Lawsuits 2026: Key Cases & Stats (JPML / Forbes data)
16. Reuters, US appeals court orders lawyer to pay US$2,500 over AI hallucinations in brief