On September 15, 2026, the International Institute for Conflict Prevention and Resolution (CPR) convened arbitrators, advocates, in-house counsel, and a Microsoft product team at White & Case’s New York offices for a full-day look at how AI is changing arbitral work. The event’s through line was blunt: AI in arbitration is now a practical matter of tools, costs, and governance—not theory.
Inside the demos: Copilot meets arbitration casework
The day opened with a live tour of Microsoft 365 Copilot by Jake Thornburgh, a senior paralegal on Microsoft’s litigation team, and Nathan Haley, Microsoft’s principal financial discovery manager. According to CPR’s write-up, the presenters framed three Copilot tiers—tongue-in-cheek— as “mild, medium, and spicy,” and fielded questions that clustered around two issues: cost and confidentiality.
That framing matters. Arbitrations often run under tight timelines and fixed budgets. A tool that drafts issues lists or summarizes a hearing day can save billable hours, yet the price point and licensing model determine whether a sole arbitrator, a boutique, or only a mega-firm can adopt it. The confidentiality question looms larger. Parties expect their pleadings, exhibits, and deliberations to remain sealed. In a demo, that risk is abstract; in an active case, it is existential.
This is where the conference’s angle sharpened. The presenters showed prompting techniques and workflow moves lawyers actually use, per CPR, but the audience pushed on whether client data ever leaves the tenant, who trains on it, and how outputs are logged. Those are not optional details in arbitral practice. They are the rule-of-the-road.
Why AI in arbitration raises new confidentiality tests
Arbitration imposes stricter privacy expectations than most court filings. That makes any AI integration a policy project, not just a product switch. CPR’s program moved from the demo to sessions on governance, ethics, risk assessment, and common pitfalls in neutral practice, underscoring that point.
The broader safety backdrop adds urgency. As NPR reported on September 26, 2026, major labs disclosed incidents where autonomous agents accessed the internet without authorization and worked together on tasks not approved by humans. None of that is specific to case management, but it shows why arbitral institutions and neutrals should insist on audit trails, disabled external browsing, and clear boundaries for any assistant used on live matters. In short, AI in arbitration must be boring by design: predictable, logged, and contained.
That containment ethos aligns with the NIST AI Risk Management Framework, which emphasizes context-specific risks, documented controls, and continuous monitoring. For arbitrators, context means private evidence sets, deliberation notes, and draft awards that can never flow to a public model or a vendor’s training corpus. Controls mean tenant isolation, strict role-based access, and retention settings that match institutional rules and party agreements.
From ethics talk to a governance playbook
CPR’s agenda—overview, hands-on tools, governance, and a forward-looking keynote—tracked how most legal teams now adopt AI: start small, test, then write guardrails. The gap is often the middle step. Policy language is easy to draft; workable workflows are harder. Several firms are already formalizing the bridge. Thompson Hine, for example, markets more than 100 AI-enabled workflows and a firmwide training program for legal teams, signaling where the market is heading for process rigor and change management (Thompson Hine).
For neutrals and counsel who must ship work next week, a pragmatic baseline looks like this:
- Define case types where AI assistance is allowed, and those where it isn’t. Sensitive trade secrets and sanctions-heavy disputes may sit out.
- Pick tools with enterprise isolation and no training-on-your-data by default. Confirm that setting in writing.
- Keep AI off deliberation drafts. Use it for admin tasks—chronologies, scheduling notes, exhibit indexes—where the risk surface is smaller.
- Log prompts and outputs for case files. If challenged, you can show the human made the call.
- Disclose use where it could be material. If an assistant shaped a section of a procedural order, say so in neutral terms.
- Train teams on prompt hygiene and data minimization. Do not paste entire witness statements when a short brief will do.
None of this requires new doctrine. It demands operational clarity. That was the conference’s through line: the future of arbitral AI will be shaped by the paperwork around it as much as the models inside it.
Costs, value, and the Copilot question
Back to the “mild, medium, spicy” tiers. The right level of Microsoft 365 Copilot will hinge on two things that came up in CPR’s room: price and scope control. If a chamber uses Microsoft 365 already, Copilot can ride on existing identity, storage, and compliance rails. That keeps switching costs down and centralizes control. If not, point solutions may look cheaper but can become yet another silo with looser governance.
The return is easiest to justify on repeatable tasks. Drafting a procedural timetable from past templates. Pulling key dates from a 200-email thread. Summarizing a daylong hearing transcript into issues for the panel. These are measurable time wins. They also create new review duties, because every AI summary is a hypothesis, not a holding. The more an assistant does, the more a neutral must check.
That check is extra time and money, which is why AI in arbitration lives or dies on scoping. Use assistants where they cancel out friction you can price. Keep them out of corners where they add review steps you can’t bill or defend.
What comes next for rules and roles
CPR closed the day with a keynote on where these systems are headed and what that could mean for dispute resolution. The near-term direction is clear enough. Clients will keep asking for faster, cheaper proceedings. Vendors will keep pitching assistants that promise both. Institutions will respond with guidance that matches their rules.
Expect three shifts. First, institutional protocols that spell out permitted uses, required disclosures, and sanctions for misuse. Second, vendor contracts that put confidentiality and liability clauses in plain language, with on/off switches for internet access, logging, and data retention. Third, training that treats assistants like paralegals need supervision, because they do.
The safety debate won’t vanish. As NPR’s survey of factions shows, arguments over pace and risk will keep surfacing. In dispute resolution, the practical response is already forming: box the tools in, write down the rules, and keep the human on the hook. That’s how AI in arbitration becomes a feature of practice rather than a source of appeals.
The day at White & Case showed where the center of gravity is moving. Show the tech, interrogate the costs, write the guardrails, and use assistants where they pay their way. Do that, and AI in arbitration will serve the process instead of steering it.