When a general counsel tells the board that a claim has 65% prospects of success, the first question in the room will be: why? The same question should be asked of every AI tool that offers a view on the outcome of a dispute. A prediction on its own is of limited value. What earns trust is the reasoning behind it: what the evaluation looked at, what it assumed and where it might be wrong. pinqDR considers what an AI evaluation should show before anyone relies on it.
Why a score is not enough
Outcome prediction is one of the more familiar promises of legal AI. The Chartered Institute of Arbitrators' Guideline on the Use of AI in Arbitration (2025) recognises that AI tools may go beyond legal research and data analysis to provide predictions of potential case outcomes. The same Guideline identifies the "black box" problem: the difficulty of understanding how an AI tool arrives at a particular result. A percentage without reasons cannot be tested, challenged or explained, and a litigator cannot advise on a figure that cannot be interrogated.
The legal system learned this lesson long ago. Judgments give reasons, and, in England and Wales, section 52(4) of the Arbitration Act 1996 requires an award to contain reasons unless it is an agreed award or the parties have agreed to dispense with them. Reasons allow a decision to be understood, tested and, where the law permits, challenged. An AI evaluation that informs a decision to issue proceedings, settle or defend deserves the same discipline.
There is a further risk. A confident number invites confident decisions: a party told that it has a strong case may reject a reasonable settlement offer, while one told that its case is weak may abandon a good claim. Without reasons, neither can tell whether the number deserves that weight.
Six questions every evaluation should answer
In our view, any AI evaluation of a dispute should answer the following questions on its face, without the reader having to ask:
Issues: which issues did it identify? A good evaluation breaks the dispute into the questions a tribunal would actually have to decide.
Sources: which documents did it rely on, and what was missing? The reader needs to know, for example, whether the other side's key correspondence was in the file.
Assumptions: what did it assume? Assumptions about disputed facts, the governing law or the construction of a contractual term should be stated rather than buried.
Weaknesses: where is the evidence thin? The most useful part of an evaluation is often the list of points on which the case is weakest.
Reasoning: how does the reasoning run? The route from the documents to the conclusion should be traceable, issue by issue, so that a lawyer can check each step.
Sensitivity: what would change the view? A reasoned evaluation should identify the document, admission or finding that would shift the likely outcome.
What judges and arbitrators are being told
The judiciary's AI guidance, updated on 31 October 2025, reminds judicial office holders that information from AI tools may be inaccurate, incomplete or misleading, and that its accuracy must be checked before it is used or relied upon. The CIArb Guideline takes a similar approach: arbitrators should not relinquish their decision-making powers to AI, should always exercise independent judgment, and should avoid delegating tasks such as legal analysis, the interpretation of facts and law, or the application of the law to the facts, where that could influence their decisions. Where it is unclear how AI-generated content relates to the evidence, the Guideline also contemplates arbitrators seeking submissions from the parties on how inputs from the record are linked to the machine-generated output.
Neither document is directed at early neutral evaluation, but both rest on the same premise. AI output is only as valuable as the user's ability to check it, and checking requires the reasoning to be visible.
Where pinqPredict fits
pinqPredict produces structured, reasoned early neutral evaluations of potential, in-flight or retrospective disputes, in minutes, from the documents provided. Structure and reasons are the foundation of trust, and the six questions above are the right standard against which any evaluation tool, pinqPredict included, should be judged. We would encourage every user to apply them.
An evaluation informs a decision; it does not make one. Where a dispute cannot be resolved on the strength of an early view, it can move to pinqDR, where an experienced arbitrator decides it and issues a final and binding award, typically within 6 to 8 weeks. AI should augment, not replace, the judgment of those accountable for the outcome.
From answers to reasons
AI evaluation of disputes will no doubt become routine. As it does, the market will distinguish between tools that produce answers and tools that produce reasons. The first kind will be quick to use and easy to ignore. The second will change how disputes are run, leading to earlier decisions, better-targeted legal spend and settlements grounded in analysis that both sides can examine.
The right question to ask of an AI is not simply who will win, but how it reached its view.
To find out more about pinqPredict, get in touch at hello@pinqDR.com.
Useful links
Courts and Tribunals Judiciary, Artificial Intelligence (AI) Judicial Guidance (October 2025)
