Could AI adjudicators become a reality?

In the current adjudication landscape, with lengthy and complex disputes veering away from the envisioned exercise in “rough justice”, there is a clear opportunity for the use of new technology to streamline disputes. While the legal sector is seeing an increase in the use of artificial intelligence (“AI”) systems by parties in disputes, formal adoption of AI systems as part of decision-making processes has not yet been observed in the UK.

Artificial intelligence is already beginning to influence the resolution of construction disputes elsewhere in the world. Since November 2025, the American Arbitration Association (the “AAA”) has offered an AI-led dispute resolution process for two-party, documents-only construction arbitrations with the aim of reducing the time and cost involved in these disputes. This development provides a useful reference point when considering whether a similar model could be used in adjudication under the Housing Grants, Construction and Regeneration Act 1996 (the “Act”).

How the AAA procedure works

The rules of the AAA’s AI-led arbitration are relatively short. The referring party submits a brief overview of the issues to be resolved and, within 10 business days, the responding party submits a brief response. There follows a round of submissions from each side setting out their case in full, with documentary evidence and legal authorities. At the end of this process, the AI analyses the material presented to it, identifies the principal issues and prepares a summary of each party’s position. The parties are given an opportunity to check and comment on the AI-generated account of their own submissions, in order to rectify any inaccuracies. The AI arbitrator then produces a draft decision. At this stage, a human arbitrator takes control of the decision, reviewing and correcting the output of the AI before issuing the final decision to the parties.

The final stage is crucial as it displays that overall control of the outcome of the dispute lies with a human arbitrator. The human arbitrator has the opportunity to review all material put before the AI arbitrator, as well as the corrected summaries, so is in a comparable position to a traditional decision maker. The effect of the AI-led arbitration rules is therefore to integrate a specialised AI into the disputes process, rather than entirely replacing the arbitrator.

Could adjudications be decided entirely by AI?

Several key challenges appear to preclude AI from deciding adjudications independently in Scotland or England and Wales.

There is no guarantee that, under the Act, an adjudication decision by an AI adjudicator would be enforceable. While the Act itself does not prescribe that an adjudicator must be a natural person, the Scheme for Construction Contracts and the JCT suite both contain this requirement. The courts may also be reluctant to enforce an AI adjudicator’s decision under the NEC4 suite of contracts as their dispute resolution procedures require that an adjudicator give reasons for their decision. AI systems powered by large language models are programs which predict probable sequences of words in response to a given prompt. While these systems are capable of producing what appears to be a reasoned decision, there is no underlying thought process to examine. The courts may therefore be reluctant to hold that the generated text constitutes a provision of reasons. More broadly, it may be that the courts consider decisions reached by an AI system to be contrary to natural justice as an AI system will inevitably reproduce any biases present in its training data.

Secondly, AI systems may omit relevant material, provide an incorrect interpretation of contractual or statutory provisions, or produce a highly confident answer which is fundamentally wrong. While the courts recognise that adjudications may be decided incorrectly, the interim-binding nature of adjudication does not make flaws in decision making harmless. A poor adjudication decision will impair proper cash flow and increase the likelihood of enforcement challenges and follow-on litigation.

Thirdly, there is a vulnerability to prompt injection, a process by which hidden instructions to an AI system are included within a document, with the intention of forcing the AI system to return a particular result. There have already been attempts to influence the judicial process by prompt injection in Brazil and the United States of America. Were AI systems to decide matters independently, there may be a suspicion that an underhanded use of prompt injection had determined the result.

With this combination of formal barriers and practical weaknesses, it seems that AI’s involvement in adjudication is likely to be kept to a supporting role for the foreseeable future.

Could AI systems take a role in assisting adjudicators?

AI-assisted adjudication under the Act comparable to the AAA’s procedure would face an immediate practical difficulty: access to suitable training data. The AAA states that its model was trained on more than 1,500 construction arbitration awards. That body of material is clearly central to the model, providing it with examples of how claims, evidence and contractual arguments have previously been analysed and expressed in reasoned awards.

Adjudications under the Act are kept confidential, and there is no single institution with a similar body of retained decisions on which to train an AI system for use in the UK. Absent a coordinated effort by adjudicators, parties to adjudications and legal and technology specialists to collect and analyse past decisions on a confidential basis, an AI system for adjudications would have to be trained on court judgments from either enforcement proceedings or final determinations of adjudicated issues. These may provide some useful material but their form differs significantly from that of adjudication decisions.

A model trained on court judgments alone may encounter difficulties as the evidence presented by the parties (other than the occasional excerpt) is not published along with the judgments, so there would be a gap in the AI’s training in relation to the assessment of evidence. With potential AI systems for adjudication being limited in this way, the need for human control of the procedure becomes clearer.

Even with human control, however, there is a risk of automation bias. Once a system has produced a coherent draft, a reviewer may tend towards accepting its framing of the dispute, or even its conclusion as a whole. It is not clear how best to manage this risk. Allowing parties to respond to the AI system’s draft decision before the adjudicator makes their determination could reduce the risk of undue reliance on the AI system’s first draft but, in doing so, parties would effectively be arguing the same issues again, diminishing any time and cost savings made.

These practical challenges notwithstanding, there does not appear to be a formal barrier to adjudicators using AI systems to streamline their work, assuming appropriate safeguards are adopted.

Adjudicators should, however, be aware that, if they are to use AI to assist their decision-making process, they should be open with the parties as to how this is being implemented. Parties will be concerned about the confidentiality of the dispute, and there may be scope to challenge a decision assisted by an AI system on the basis of natural justice if the parties were not made aware of its use. In Babcock Marine (Clyde) Ltd v HS Barrier Coatings Ltd[2019] CSOH 110, it was argued that an adjudicator’s decision was in breach of natural justice, as the adjudicator received significant assistance from a quantity surveyor without declaring that this was done. The argument was that this ran contrary to a provision in the adjudicator’s contract requiring him to declare to the parties any assistance received, and the court held that this was a potential breach of natural justice.

While there is a clear difference between obtaining assistance from a qualified professional and an AI system, analogies could be drawn with this case if an adjudicator was to seek assistance from an AI tool on a matter in which they have less expertise. In any event, it is clear that best practice is to disclose any planned use of AI where this affects the substance of the decision.

The future of AI in adjudications

The initial formal adoption of AI systems in adjudication is unlikely to be as dramatic as the label “AI adjudicator” might suggest. The careful use of secure AI tools to organise documents, summarise submissions or prepare an initial structure for a decision could well be implemented in the current legal landscape, but a sudden move towards independent AI adjudicators is not likely. If AI assistance is to increase going forward, a formal system could make that process more transparent by defining how AI systems are used, allowing parties to see the AI’s product and communicating the extent of the human adjudicator’s control of the outcome. While a specialised AI adjudicator model seems out of reach for now, AI-assisted adjudication appears increasingly plausible and could be beneficial, provided that its introduction is accompanied by transparency, independent judgment and a clear allocation of responsibility for the final decision.

If you have any questions on the topic above, please get in touch with a member of our construction team below or your usual Brodies contact.