AI in the Employment Tribunal

Over the last few years, AI has become a common feature of employment disputes. More and more people use cloud-based services such as ChatGPT, Gemini and Claude to draft grievances, claims and submissions. For many employees, the ability to open a web page, ask a legal question and receive an immediate answer that sounds correct, at little or no cost, is understandably attractive, though what is produced sometimes causes more problems than it solves.

Practitioners will be familiar with the risks, whether from clients who now arrive at a first meeting with pages of inaccurate AI-generated advice, or from the cautionary tales of AI-generated submissions or claims that have not survived judicial scrutiny. One of the best known is R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin), in which the Divisional Court considered grounds that cited authorities which, on closer inspection, did not exist.

At the same time, AI opens up new ways of accessing and delivering legal advice. It can help people locate relevant case law and show how tribunals have approached similar issues, and when used well it can support the advice and advocacy that people rely on rather than replacing it. The purpose of this project is to explore whether the well-known problems with cloud-based AI could be remedied with a more grounded approach.

The problems

The first concerns sources. One of the challenges in advising clients who rely on their own AI-generated advice is unpicking what is real and what is not. Most cloud-based AI services answer legal questions from patterns learned in training, drawn from case law, statute and commentary but also from forums, blogs and much else besides. Many can now search the web as well, but the pages they find are as likely to be law firm articles or discussion threads as the judgments themselves. Where a secondary source states the law inaccurately, or attributes a proposition to a judgment that does not contain it, the model can reproduce the error with complete confidence. Where it has no relevant source at all, it may produce something that sounds plausible but is wrong, or get the surface detail correct but provide a wholly invented source for it.

The second concerns the questions asked. A cloud-based AI tool works only from the question as it is framed. Consider an employee with ADHD whose manager proposes moving them from an open-plan office to a separate room, remarking that they would be less distracted there. Unhappy at the prospect of being separated from colleagues, the employee asks an AI tool whether this is disability discrimination. The tool may well describe the move as punitive and identify claims of harassment and discrimination arising from disability.

A human adviser would first ask about the background. Had the employee raised difficulties concentrating? Had occupational health recommended a quieter workspace? Was the move proposed or imposed, and what reason was given for it? The answers may reveal an employer trying, if clumsily, to make a reasonable adjustment, which changes the analysis considerably.

The approach

Since February 2017, Employment Tribunal judgments have been published on GOV.UK. The register now holds more than 130,000 decisions, available to anyone under the Open Government Licence. It is one of the largest freely available bodies of first-instance decisions in the UK, and it is rarely examined as a whole.

I have taken the judgments themselves as the starting point, developing a series of tools to search, extract and analyse information from the complete register. Anything the tools produce can be traced back to a specific decision and checked against it. This series describes what the tools do, how they were built and what they have found.

How far can a modest, local set-up go? The largest commercial AI systems run on infrastructure that no individual practitioner or small organisation could replicate. At the top end of what is realistic locally, a well-specified desktop computer with a graphics card carrying 24 GB to 32 GB of memory can be built for roughly £1,800 to £2,600. A single data-centre processor of the kind behind the major AI services carries 180 GB of memory and costs an estimated £22,000 to £37,000, and those processors are operated in their thousands. I wanted to see whether a single desktop computer could nonetheless produce something of practical value, and to establish what is lost in exchange.

How reliable can the output be made? I have treated verification as central to the project. Any figure a model extracts from a judgment must appear in the text, and any passage it quotes must be quoted verbatim. Results that fail these checks are set aside.

The series is therefore exploratory. Some of the tools have produced results that I think are useful in practice; others mainly illustrate the limitations of the approach.

The questions this series will look at

I have organised the project around three groups of questions.

For an individual seeking advice

• Are there potential claims here, and what are they?
• Are they good claims, and are they worth pursuing?
• What might they be worth, and what have tribunals awarded in similar cases?

This is probably the most common use of AI tools in legal work, and the one for which a desktop model is least well suited. A model small enough to run on a single computer cannot match the large commercial systems on breadth of knowledge or reasoning, and I have not tried to make it do so. The question is whether a different approach can produce a comparable result.

Instead of asking the model to assess a claim in one step, the tool uses a separate list of issues for each claim type, drawn from the statutory provisions and checked against the lists of issues set out in judgments on the register. Time limits and calculations are handled by fixed rules. The model is asked narrower questions, and each answer must be supported by a quotation from the employee’s description. Where a potential claim is identified without sufficient detail to complete the list of issue template, the tool will say so and request further information.

For a practitioner seeking more technical guidance to give a client

• In a discrimination claim, given the effect on the client and the seriousness of the treatment, what would be a typical award for injury to feelings, and are there decided cases broadly similar on the facts?
• Where compensation may be reduced for Polkey or contributory fault, or adjusted for failure to follow the Acas Code, what percentages have tribunals applied on similar facts, and what is likely in this case?
• Where a claim is out of time, which arguments have succeeded in extending time on “just and equitable” or “reasonably practicable” grounds, and how often do such applications succeed?

The most developed part of the project at this stage is a database of more than 2,500 injury to feelings awards extracted from the register, each recorded with its Vento band, the protected characteristic concerned and, where given, the tribunal’s reasons. It underpins the first of these questions and several in the next group. The same method of extraction can be applied to reductions, uplifts and extensions of time, which practitioners usually have to estimate from experience.

For anyone interested in how the tribunal works

• Have injury to feelings awards kept pace with the uprating of the Vento bands?
• Is there a discernible difference between awards for different protected characteristics?
• How often do claimants succeed, and does representation make a difference?
• Do outcomes vary between regions?

These questions are usually answered from experience or anecdote. With every published decision in a single dataset, they can be answered from the decisions themselves.

Testing the approach

Towards the end of the series, I will test how this approach compares with cloud-based AI services. I will prepare a number of fictional client scenarios and give them, unchanged, to the major AI services, in both their free and paid versions, and to the local tool. Each answer will be anonymised and marked blind against an answer key prepared and sealed in advance.

A note on reliability

The figures and findings in this series come from automated extraction. I describe the checks applied to each tool, and its known limitations, in the relevant articles.

Two limitations apply throughout. Nearly three in ten of the published judgments are very short documents, such as dismissals on withdrawal, containing almost no text. These are excluded from the analysis, although a sample check suggests little of substance is lost as a result. In addition, claims that settle, are withdrawn or are decided orally without written reasons will largely be missing from the register, so the figures reflect published decisions only.

Nothing in this series is legal advice, and no figure should be relied upon without reference to the underlying judgment.