Artificial intelligence (AI) has already fundamentally changed how people work. It is excellent at generating, recording and analysing data. It can solve complex problems and enable its users to complete all manner of tasks extremely efficiently and to make decisions. It is already a great tool, which makes it all the more remarkable (or concerning?) that it continues to improve at an apparently exponential rate.

Businesses instinctively want to use this new technology, and so they should. But the widespread adoption of AI within the workplace gives rise to a number of interesting issues in the employment law sphere that employment tribunals will no doubt soon have to grapple with.

Who decided?

Take the call centre worker whose calls are transcribed and scored in real time for sentiment, compliance and productivity. At the end of a third month in the bottom decile, a manager looks at the dashboard and dismisses for capability.

Capability is a potentially fair reason for dismissal, and the reason is not in doubt. The difficulty arrives at s 98(4) Employment Rights Act 1996, and it is the same difficulty an employment tribunal has always had with a dismissing officer who acted on a report he did not understand. Did the manager know what the sentiment score measured? Did he know its error rate, or that the model had been trained on a different customer base with different accents? Did he look behind the number, or did he treat the output as a finding of fact?

Those questions are familiar from any case where an employer relies on an investigation conducted by somebody else. What is new is the question whether the employer will be able to say that it was reasonable to rely upon an investigation conducted by AI, and whether its belief in the AI’s conclusions can be said to be reasonable. In certain cases, we can imagine a situation where an employer will be unable to interrogate the findings of the AI, because the model used is a licensed third-party product whose workings are commercially confidential. An employer who cannot explain how a score was arrived at will struggle to show that dismissing on the strength of it fell within the band of reasonable responses.

The practical consequence may be a disclosure fight. The claimant wants the model documentation, the validation testing, the error rates and the raw scores of the comparator group; the respondent says all of it belongs to the vendor and is not within its control. Employment tribunals have little practice at resolving that argument.

Redundancy that is really redesign

The headline proposition is that AI removes jobs. The emerging picture is less sensational: employers are redesigning roles more often than deleting them, but that gives rise to its own legal difficulties.

Redesign might amount to a contractual variation. It might change what a person is paid to do without changing the headcount, and therefore may carry none of redundancy’s statutory machinery. Imposed without agreement, it can produce claims for breach of contract, unlawful deduction from wages and constructive dismissal.

On the other hand, if an employee accepts a redesign of their role, it may act as a trap for that employee. The case of Glitz v Watford Electric Co Ltd [1979] IRLR 89, EAT illustrates the point. G was employed as a ‘general duties clerk’. Later a duplicating machine was bought by the company and eventually the company made its operator redundant when it reduced its clerical staff and asked G to take on the additional duty of operating the machine. G agreed at first but soon had to be taken off such work as vapour from the machine gave her headaches. In dismissing her appeal, the Employment Appeal Tribunal held that because she had agreed to work the machine, her later dismissal for capability was fair when she proved unable to work it.

We can imagine an employee who agrees to use AI in their role but later realises they cannot understand the technology, leaving them facing dismissal for capability.

The project that was shelved

Another interesting question arises when the technology fails. An employee is dismissed because the employer expects a new AI system to absorb the employee’s former role. Six months later the project is abandoned.

The temptation is to say there was never a redundancy at all. That overstates it. Section 139(1)(b) covers a requirement for employees that has ceased or diminished, or is expected to cease or diminish, and the expectation is judged as at the date of dismissal. An employer whose forecast was genuinely held does not lose the reason because the forecast turned out to be wrong.

The point is evidential. The employer bears the burden of showing the reason for dismissal. A shelved project puts every internal document about the expectation in issue: the business case, the board papers and what the decision-maker was actually told about delivery risk. Where the forecast turns out to have been a slide deck rather than an analysis, the reason begins to look constructed after the event, and an employer who cannot prove its reason will not get as far as arguing that the dismissal was fair.

Whistleblowing and confidence

Two claims run in opposite directions from the same technology.

The employee who reports that a model is discriminatory, that monitoring is unlawful or that outputs are being misused may well be making a qualifying disclosure under s 43B(1)(b). Detriment and dismissal claims follow under ss 47B and 103A. Expect these from data and engineering staff, who see the failure first and often understand it best.

Running the other way is the employee who feeds confidential material into a public model. Although not an employment case, the Upper Tribunal observed in UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC) that uploading confidential documents into an open AI tool places that information in the public domain, breaching confidentiality. An employee who pastes a client list or a set of board papers into a chatbot is in breach of the duty of confidence and, in almost every case, of an express contractual term as well.

Where this leaves us

We do not expect that any of the causes of action will change. They have stood the test of time and can be applied flexibly to properly consider claims that have been affected by AI. What will change is the approach to the evidence. Parties should be asking themselves: what system was used, what it measured, who reviewed the output, what that reviewer understood and what they were able to do about it.

For respondents the answer is documentary. An employer who can produce the impact assessment, the record of what the decision-maker was told and evidence of a human review will be in a materially stronger position. An employer who cannot will find that the employment tribunal’s instinct and the claimant’s cross-examination run in the same direction: the machine produced a number, and nobody in the organisation can say what it meant.