AI in the Employment Tribunal
A project applying local AI to every published Employment Tribunal decision: what it involves, how it works, and where it may be useful in practice.
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 anywhere in the UK, and it is rarely examined as a whole.
Over the past few weeks I have been building a set of tools to do exactly that. They run locally on a single desktop computer and use AI models to search, extract and analyse information from the complete register. This series describes what the tools do, how they were built and what they have found. Where it may be useful, each article also notes what would be involved in doing something similar yourself.
The context
AI is already part of employment practice, whether or not practitioners use it themselves.
Litigants in person are using cloud-based AI services like ChatGPT, Gemini and Claude to draft grievances, claims and submissions. In June 2026 the Presidents of the Employment Tribunals issued guidance on applications for interim relief noting that many lengthy submissions appear to have been generated using AI. The guidance does not prohibit its use, but reminds parties that their submissions must be concise, relevant and accurate.
Practitioners are not immune from the risks. In *R (Ayinde) v London Borough of Haringey* [2025] EWHC 1383 (Admin), the Divisional Court considered grounds that cited authorities which did not exist. The Court observed that AI “carries with it risks as well as opportunities”, and that its use “must take place therefore with an appropriate degree of oversight”.
Specialist tools are developing on both sides. Online services now offer AI-assisted drafting of tribunal claims directly to individuals. On the practitioner side, the major legal publishers are building AI into the research platforms that many firms already use.
Two questions about general-purpose AI
My interest in this area began with two questions about how general-purpose AI handles employment law.
The first concerns sources. General-purpose models answer legal questions from what they absorbed in training: a mixture of case law, statute and commentary. Where a secondary source states the law inaccurately, or attributes a proposition to a judgment that does not contain it, the error can be reproduced. Where the model has no relevant source, it may produce something plausible sounding instead. That is one reason this project starts from the judgments themselves.
The second concerns the questions asked. A model only answers the question it is given, on the facts as presented. 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 and less distracting there. On that account, an AI tool may well identify potential claims of harassment related to disability and discrimination arising from disability.
A practitioner would first establish the context. Had the employee raised difficulties concentrating? Had occupational health recommended a quieter workspace? Was the move proposed or imposed? Often the answers show an employer attempting, if clumsily, to make a reasonable adjustment. That does not make the remark appropriate. But it changes the analysis considerably, not least because a harassment claim requires that it be reasonable for the conduct to have had the alleged effect. The skill lies in the questions. How far general-purpose tools ask them unprompted is one of the questions this series will test.
The approach
Three questions have shaped how I have approached this project.
What can be learned from the judgments alone? Much of what practitioners know about how tribunals decide reaches us through commentary, reported cases and experience. I wanted to see what emerges when the starting point is the decisions themselves. I have therefore started from the text of tribunal judgments, so that anything the tools produce can be traced to a specific decision and checked against it.
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. The most powerful graphics card available for a desktop computer has 32 GB of memory and costs upwards of £2,400; the data-centre processors behind the major AI services carry 180 GB each, cost an estimated £22,000 to £37,000, and are operated in their thousands. I was curious whether a single desktop computer could nonetheless produce something of practical value. Running the models locally also means that nothing leaves the machine, which would matter for any use involving client information. Part of the exercise is to establish what is lost in exchange for that.
How reliable can the output be made? AI models produce fluent output whether or not it is correct. I have treated verification as central to the project. Where a model extracts a figure from a judgment, the figure must appear in the text; where it quotes a passage, the quotation must be verbatim; and results that fail these checks are set aside rather than included.
The series is therefore exploratory. Some of the tools have produced results that I think are genuinely useful; others illustrate the limitations of the approach.
What I plan to look at
The series will cover four areas, each chosen because it has a practical application in employment practice.
Injury to feelings awards. This is the most developed part of the project. The tools have extracted more than 2,400 injury to feelings awards from the register, each with its Vento band, the protected characteristic concerned, the date and, where the judgment gives one, the tribunal’s reasons. Unlike the curated tables in the commercial databases, the dataset draws on every published judgment and can be updated as new decisions appear. I will describe how it was built and checked, and a viewer that allows awards to be filtered by band, characteristic, year and amount, showing where a given figure sits among them. I am also developing a way of identifying awards made on comparable facts, profiling each award by the factors the case law treats as significant, such as the severity and duration of the effect on the claimant, so that a realistic range can be suggested for a new case.
Trends in awards. With the awards in a single dataset, it becomes possible to examine questions that are usually answered from experience: how awards differ between protected characteristics, whether they have kept pace with the annual uprating of the bands, and whether representation is associated with different outcomes.
Analysing a client’s problem. The most ambitious part of the project is a tool that takes an account of a workplace problem and carries out the kind of first-stage analysis an adviser would undertake at an initial appointment: identifying the potential claims and the issues to explore, assessing them, and indicating a realistic range of value, with reference to tribunal decisions on similar facts. It runs entirely on a desktop computer. I will set out how it works, what it does well, and where it falls short, which at this stage is in a number of places.
How AI systems handle employment law problems. The question I am most interested in is how well the different AI tools that clients and practitioners already use actually perform on a realistic employment law problem. This will involve preparing several fictional client-intake scenarios, which will be given unchanged to the major general-purpose AI services, across both their free and paid versions, and to the local AI tool. Each answer will be anonymised and marked blind against an answer key prepared and sealed in advance.
Later in the series, I hope to extend the approach to Employment Appeal Tribunal judgments. These are published through The National Archives’ Find Case Law service, which requires a separate licence for automated analysis, so the scope of that work will depend on the terms of that licence.
A note on reliability
The figures and findings in this series come from automated extraction. I describe the checks applied to each tool, and the known limitations, in the relevant articles. In particular, around a quarter of the published judgments are very short documents, such as dismissals on withdrawal, containing almost no text. They are excluded from the analysis throughout, although a sample check suggests little of substance is lost as a result.
Nothing in this series is legal advice, and no figure should be relied upon without reference to the underlying judgment.
If you spot an error in any of the figures, or have tried something similar yourself, I would be glad to hear from you.