AI-driven restructuring and redundancy: the UK legal duties

"AI restructuring" now appears in board packs and press releases, and increasingly in tribunal claims. The legal framework has not changed: what has changed is how easily an employer can end up with a selection process it cannot explain.

This guide separates two distinct issues — AI as the reason for redundancies, and AI as a tool in the redundancy process.

AI as the reason: is it a genuine redundancy?

Redundancy under section 139 of the Employment Rights Act 1996 requires that the requirement for employees to do work of a particular kind has ceased or diminished. Automation is a classic redundancy reason: if AI absorbs the work, the requirement for people to do it genuinely diminishes.

Two traps recur:

• The work has not gone, only moved. If the tasks are redistributed to remaining staff or offshored, the employer must still show the requirement for that kind of work has reduced. A role rebadged with AI in the title and refilled shortly afterwards invites a claim that the dismissal was not by reason of redundancy.
• Redundancy used as a cover. Where AI adoption is used to remove particular individuals, the real reason may be capability, conduct, or something unlawful. Contemporaneous documents — including internal messages about who "won't adapt" — are disclosable.

Age discrimination is the standout risk. Assumptions that older workers cannot or will not adapt to new tools show up in restructuring papers with some regularity, and they convert a redundancy exercise into a discrimination claim.

Consultation duties

• Individual consultation is required for a fair dismissal in every redundancy, however small. It must happen at a formative stage, with genuine consideration of alternatives.
• Collective consultation under section 188 TULRCA applies where 20 or more redundancies are proposed at one establishment within 90 days: 30 days' consultation before the first dismissal, rising to 45 days for 100 or more, with duties to consult on ways of avoiding dismissals, reducing numbers and mitigating consequences.
• Notification to the Secretary of State (form HR1) accompanies the collective threshold. Failure is a criminal offence.

Where the driver is AI adoption, "ways of avoiding dismissals" has real content: redeployment, reskilling, phased implementation and reduced hours are all things a tribunal will expect to see genuinely explored. An employer whose consultation papers treat the technology decision as immutable is inviting a protective-award claim.

Selection criteria that survive scrutiny

Criteria should be objective, capable of evidence, and applied consistently. AI adoption tempts employers towards criteria that are none of those things:

• "AI skills" or "digital adaptability" — legitimate in principle, but it must be assessed on demonstrable evidence and against training actually offered. If only some staff had access to the tools, the criterion is unfair and potentially discriminatory.
• System-generated productivity data — attractive because it looks objective. Before using it, check what it actually measures, whether it accounts for reasonable adjustments, part-time hours, caring-related patterns or different role mixes, and whether staff knew it would be used this way.
• Attrition-risk or "flight risk" scores — never appropriate for selection. They are predictions about a person's intentions, not measures of their contribution.

Document who scored, on what evidence, and what moderation took place. Scores generated by a system still need a human who can explain each one.

Using AI inside the redundancy process

If a tool ranks employees for selection, you are making an employment decision informed by automated processing. That triggers the UK GDPR duties covered in our guide on automated decisions: a DPIA, transparency in the staff privacy notice, meaningful human involvement, and a route to contest the outcome. It also creates a discovery problem — the employee is entitled to their data, and if the model's inputs are embarrassing, they will be read out at hearing.

Using generative AI to draft consultation documents or dismissal letters is lower risk but not zero. Fabricated figures and copy-pasted templates that contradict the actual process are a recurring reason for procedural unfairness. Whoever signs it owns it.

A defensible process

1. Record the business case for the change, including the role of AI, before individuals are identified.
2. Define the pool and criteria, and evidence why each criterion measures what you say it does.
3. Consult genuinely, at a formative stage, on the change as well as the outcome — including retraining and redeployment.
4. Score on evidence, moderate, and keep the working papers.
5. Offer suitable alternative employment where it exists, and search across the group.
6. Give individual outcomes with reasons and a right of appeal.
7. Run an equality impact check on the provisional pool and outcome before finalising.

The practical takeaway

AI does not create a new redundancy regime. It creates new evidence — productivity dashboards, model scores, internal messages about who will adapt — and that evidence is disclosable. The employers who come through cleanly are the ones whose paper trail shows a business case, real consultation, and criteria a human can explain without reference to a black box.