By Vicki Qingning Zhang, in conversation with Rashid Muhamedrahimov, Director of the Applied Research Center, Cornerstone Research
Author’s Note: Mr. Muhamedrahimov spoke in his personal capacity; his views are his own and not those of Cornerstone Research. The author is a Steering Member of the SVAMC Young Professionals Committee; she works in AI and studies English law. Sections marked “My View” reflect the author’s views alone, written in her personal capacity, and should not be attributed to Mr. Muhamedrahimov, Cornerstone Research, SVAMC, or the author’s employer. Nothing in this post is legal advice.
Expert Evidence in Arbitration
How much would a company have earned if its trade secrets had not been taken? Does a platform’s algorithm disadvantage the rivals that sell through it, and by how much?
Expert evidence assists tribunals with questions like these, especially in matters that are particularly technically complex. And because arbitration can keep source code, trade secrets, and commercial data out of the public record, it is well suited to technology disputes. In May 2026, the English Court of Appeal accepted that a global patent licensing dispute concerning Nokia could be resolved through arbitration (Acer v Nokia [2026] EWCA Civ 564 at [91]–[92]). An expert is often supported by a team of consultants who, under the expert’s instruction, carry out the data analysis, economic modeling, and research that feed into the expert’s report. I spoke with Rashid Muhamedrahimov, who directs the Applied Research Center at Cornerstone Research, one such firm, about AI’s role in expert testimony work.
Judgment: Now and Future
Q: Where is AI actually being used in expert work today?
Rashid: “To simplify greatly, you can think of there as being two ends to the spectrum of work. One is high-judgment, high-expertise work: which arguments are most compelling, whether a statistical model matches the economic theory and the facts of the market. The other is implementation and execution, such as document formatting. AI is an incredible set of technologies for doing that execution work much more quickly. Instead of writing the code that cleans a dataset, you spend the time on which hypotheses to test and what is interesting in the data.” Still, even the execution end is full of judgment calls: “When you clean a client’s data and a field has a missing observation, do you drop it, or assume it is zero? AI will happily make that call for you. Whether it is the right call, for the question you are answering, is a human judgment.”
Q: Will the increasing competency of AI “encroach” on the number of judgment calls available for humans?
Rashid: “It would be an incorrect framing to think there is a fixed amount of work, where AI takes over more of the small judgment calls, and we are left with less. The number of judgment calls goes up, because we can execute on so much more: analyze more data, ask more questions. Everyone will be able to produce more output. The skill of understanding all those analyses, and how they fit the question and the industry, only matters more.”
Quality Control and Verification
Q: Expert reports are scrutinized down to the tiniest detail. What measures can safeguard against AI-driven errors?
Rashid: “Quality control should be built into any AI tool use. We never just let the AI do a task and run with the output. In document review, say, sample from the set, have a human label the sample, and check the AI output against it, or have agents cross-check other AI tools. But the most important part is the judgment of practitioners: knowing to what extent to trust different levels of AI output, and how to verify them. That is an organizational skill, and I think it will be a big differentiator between firms. With responsible use, with people who own the output and are qualified to own it, under the scrutiny of the other side, we have an opportunity to level up the work we do. The downsides are manageable. They should not stop us doing something good.”
Impact for Young Professionals
Q: Might frictionless AI stunt the development of junior analysts?
Rashid: “Whether removing friction is good depends on whether what you would have learned from it is still relevant in a world of agentic AI. The syntax of code matters much less now. What matters as much as it ever did is the logical steps you take to transform or interpret a dataset. So you learn those foundations, you eat your vegetables, and then you work in a hybrid way, pair programming with AI rather than pure vibe coding.”
The Cost of Opting Out
Q: Is it ever irresponsible for a firm not to use AI?
Rashid: “I won’t comment on irresponsibility. That is a value judgment. But suppose a firm says, let’s not use AI, because even used carefully it mislabels documents, say, half a percent of the time. You miss out on all the ways AI expands what you can do: the analyses you run, the data you can handle, the hypotheses you can test. We will always use tools in a way where we are confident the outputs are careful and reliable. But by using AI we can provide better work more quickly, and that is what clients want to pay for. To leave that on the table because of a very mitigable risk is probably not the right business decision.”
My View
The “Legal Statement on Liability for AI Harms” issued by the UK Jurisdiction Taskforce (UKJT) in 2026 explores the allocation of responsibility for AI-assisted work under existing English legal principles. The UKJT statement indicates that a professional who fails to exercise reasonable oversight of AI outputs is likely acting negligently (¶ 64(f)). SVAMC’s own Guidelines on the Use of AI in Arbitration (2024) contemplate a tribunal asking an expert about “the review process applied to ensure the accuracy of the output” (Commentary to Guideline 4, Scope section). To inspect AI-assisted work product, Rashid mentioned quality control mechanisms such as labeled samples and cross-checking agents. In my view, such a paper trail can double as the auditable record of responsible use. While the public discourse concentrates on negligent use of AI, the UKJT statement also sheds light on the inverse: a professional could be liable for “failing to use AI for a task when a professional exercising reasonable care and skill would have done so” (¶ 67). From cockpit automation to drug-interaction alerts, the advent of new tools is often met with initial resistance before an industry-wide recognition of their value and adoption of a balanced approach. AI may well be on the same path.