Beyond Memorials: Will AI Re-engineer Arbitration’s Written Phase?

As artificial intelligence becomes capable of processing, organising and interrogating complex factual material at scale, might it not only change how efficiently the participants each conduct their own roles in an arbitration but also force a more profound structural change in the written phase of the arbitral process? This question was discussed recently on a panel moderated by Robert Hunter at the ICC Germany Arbitration Day 2026 in Frankfurt and it is one that has the potential to move rapidly into the mainstream of arbitration practice¹.

AI is already being used increasingly widely across the arbitration ecosystem. Parties, counsel, experts, institutions and service providers are exploring tools that can search, classify and summarise documents, generate chronologies, identify patterns in large datasets, compare witness accounts with underlying records, and assist with drafting and legal research. The focus of much of the discussion, and in particular of published guidelines, has been on the allowable use of this technology by individual parties and their counsel on the one hand and by tribunal members on the other, in particular with regard to questions of fairness, transparency and confidentiality. At the other end of the spectrum, there is discussion about the potential of AI to supplant the role of human decision-making altogether.

This article considers the power of AI in particular to disrupt the underlying structure of the written phase of arbitration procedure, which is the role of the sequential exchange of memorials — that is, comprehensive statements of claim, defence, reply and rejoinder — as the principal means by which a tribunal is introduced to the parties’ cases, the evidential record and the issues it must decide. It asks whether, in suitable cases, AI could help create a more structured, source-linked and dynamically navigable – and thereby both faster and more efficient – route from a complex factual record to a reasoned decision, while preserving the adversarial character and procedural safeguards of arbitration.

The memorial as a standard feature of adjudication

International commercial arbitration has long settled into a widely adopted a procedure by which the entirety of the facts, evidence, technical opinion and legal argument are first fully set out in a sequential exchange of submissions (or “memorials”) in an initial written phase, followed by a substantially oral second phase in which those fully-pleaded cases are tested before the tribunal in a hearing.

The first written phase has become so familiar that it would be easy to treat it as an inherent feature of a judicial process, yet its almost universal adoption by the international commercial arbitration community was by no means self-evident since it is native neither to the common nor civil law litigation procedures familiar to most parties and their counsel. Rather, it has been borrowed from the practice of international law tribunals, where it had developed over the course of the preceding century as a practical solution to the procedural problem of finding an efficient and fair way to give a delocalised tribunal, relying to a substantial extent on a collaborative procedure, a sufficiently complete account of a complex dispute by the end of a single, compact oral hearing as to enable them to make a reasoned decision.

The memorial performs several functions at once. By identifying the legal and factual propositions that each party is advancing, it both gives the other party notice of the case to be met and, under reply, defines the scope of the issues to be resolved, both functions of the traditional pleading. It goes beyond those traditional functions by also front-loading the presentation and organisation of the documentary evidence as also the material witness evidence and expert opinion to be relied on, typically with opportunity for mutual rebuttal in a second exchange.  It additionally provides a narrative through which the tribunal can more quickly and better understand the context and development of the dispute, at least in terms the parties each wish the tribunal to see it, and it creates a record against which the award can later be assessed, challenged or enforced before a tribunal, foreseeably sitting within an entirely separate judicial system.

This article does not question the underlying necessity of those functions being performed. The question is rather whether, in the age of LLMs, they must all continue to be performed entirely through an initial sequential exchange of two, four, six or even eight cumulative but individual and often long and repetitive compilations of multiple documents.  This question becomes all the more pressing in light of the sustained pressure towards leaner and quicker processes. By way of example, it is hard to envisage how a double sequential exchange of memorials in the traditional way will be practicable in the context of the ICC’s recently introduced “Highly Expedited Procedure”, which allows in practice around two months for the written phase of a proceeding at best, allowing a month or so for the writing of an award.  Such an exchange is challenging even within the ICC’s “standard” expedited process, which now applies by default to disputes up to US$ 4 million in value, certainly enough regularly to capture disputes of real factual and technical complexity.

Over the three decades since the memorial process became largely standardised, information technology and in particular digitalisation has reached into all aspects of an arbitration lawyer’s workflow, including document review, hearing preparation and advocacy. Review platforms have replaced paper files; pleadings and exhibits are exchanged electronically; hearing bundles are navigated on screens; and real-time transcripts and remote participation have become routine.  However, while these technological developments have been used to digitise and scale the underlying model, they have not changed its underlying architecture. Despite the substantial shift from paper to digital displays and the massively increased power and speed of calling up and exchanging digitalised data, the underlying procedural choreography today would be instantly recognisable to an arbitrator or counsel practising in the mid- to late 1990s.  Counsel still prepare and exchange sequentially comprehensive memorials drawing each on their subjective selection of the documentary record, bespoke narrative statements from witnesses and the presentation of technical opinion on selected issues through experts’ reports. As a result, the same project history may appear in multiple iterations in pleadings, witness statements, expert reports, chronologies, opening submissions and closing submissions. It is a significant labour, not just for the tribunal but for all participants, to work through an accumulation of overlapping accounts of the same events multiple times in order to identify what is agreed, what is disputed and what will be material to the outcome.

There are sound reasons for this. Arbitration is adversarial. Parties must have reasonable opportunity to define their own case, to characterise the facts as they see them, and to advance the interpretation of documents and evidence that best supports their position.

Nevertheless, the existing memorial process has significant costs. In particular, it tends to generate repetition both within and between memorials and it creates an incentive and almost an imperative for each party, in the words of one eminent arbitrator, to “deconstruct the reality of the project and construct a new one; in fact, two new realities, one on each side of the dispute”². This demands that the parties devote substantial resources to presenting factual material, a proportion of which will not, in the end, genuinely be contested or material to the outcome and successive competing reconstructions of the same factual record in technically complex or document-heavy cases challenge the tribunal to be able to see the wood – or at least the authentic wood – for the trees. It also often leads to new substantial factual and expert issues emerging only in the second exchange of submissions as late as a year or more into the dispute. The cost is not only monetary: the opportunity costs are also substantial, in terms of both the diversion of productive resources within the parties and of the calendar time required to reach a point of reliable evaluation and judicial determination.

Why AI may provide a different solution

The process of determining commercial disputes is data driven.  This plays to the strengths of LLMs, which are good at reading, categorising and comparing material; surfacing connections between a factual assertion and potentially relevant source material; identifying recurring themes or apparent inconsistencies; and helping users navigate a complex record.

Exploiting these efficiencies does not mean we must accept that AI should decide disputes.  There are legitimate and profound concerns about whether AI is fit to replace legal judgment, to assess a witnesses’ credibility, to evaluate and weigh between conflicting expert opinion, and to assume all other aspects of a tribunal’s adjudicative responsibility necessary to arrive at an authoritative account of the facts or law.  While recognising that humans may also be fallible in the exercise of these functions, in arbitration, as in battle, most of us would rather trust to a human decision-maker with a record of competence and integrity than to a robot controlled by AI.

Nonetheless, users should also be able to expect a proper use of available technology. The more interesting question is therefore whether AI may already, or may at least soon, help parties and tribunals manage and perform workflows within a case differently so as to arrive more efficiently at a properly reasoned decision by means of an equivalently fair process.

The starting point should not be that arbitration needs to become “more efficient” or simply quicker merely because technology makes that possible. A faster procedure is not necessarily more efficient just as an efficient procedure is not necessarily the quickest or the cheapest one and the parties may disagree on the desired quality of output. A process that reduces cost at the expense of procedural fairness, party autonomy or the tribunal’s ability to understand the evidence would be a poor trade.

In some cases, a conventional process of sequential memorials may remain the clearest and most proportionate route from an unstructured complex evidential record to the point at which a tribunal might render a reasoned decision. It allows each party to tell its story in its own words and gives the tribunal a structured account of the case.

In other cases, however — particularly those involving a large, shared pool of project documentation and large operational datasets, above all in disputes where the calendar time to reaching a determination is at a premium — the memorial procedure in its current form may be an inferior process, at least in terms of efficiency.  The process of sequential exchange requires each party repeatedly to compress a large and interconnected factual record into lengthy, static and above all perspectival narratives. The resulting memorials, with their supporting documentation, may be individually persuasive and carefully prepared, but they do not always provide the tribunal – or even the parties themselves – with the most efficient route from proposition to proof to decision.

A shared map of issues and evidence?

It is against this background that the panel at the ICC Germany Arbitration Day considered whether a single, shared, source-linked tool could foreseeably take over some of the organisational work that the process of preparing sequential memorials currently performs, such as identifying material issues, linking propositions to sources, structuring chronologies and exposing factual or technical questions requiring proof, while preserving the adversarial process.

The process might look something like this:

• The parties would continue the practice of exchanging initial statements of claim and defence.  Their precise form and content would reflect their function, which would be concisely to set out the parties’ respective claims, defences and core factual positions for the purpose of defining the scope of the issues to be resolved.  While these statements may identify the material supporting evidence to the extent the party wished, they would not present it.
• Following that initial adversarial exchange, a shared, collaborative AI-enabled process could then synthesise this initial exchange in the context of the common corpus of project documentation. The output would deliberately not be as ambitious as an agreed narrative or a finding of fact, much less a determination. It would be a source-linked map of issues and evidence provisionally identifying the principal issues, the parties’ competing propositions (including identification of those issues still ripe for reply and rebuttal), the material documents said to support each position, the relevant factual sequence, the relevant and material common data and the areas in which witness, expert evidence or missing documentary evidence may be required.

The map would not eliminate disagreement: each party would remain free to challenge the output in all respects, including the scope of the selected material and its framing of an issue, as well as to present its own interpretation of the facts and the legal consequences. Its value would lie in the efficiencies to be reaped from two related aspects:

• accelerating the process of defining the issues and the material evidence by identifying early what is – or should be - agreed, what is peripheral and what is both genuinely contested and potentially material to the tribunal’s decision, at the same time as putting all or at least a substantial proportion of the relevant shared documentary evidence on the record; and
• providing a single narrative and evidential core from which the parties can then more efficiently structure and plead their individual cases.

Take a delay dispute. A shared evidential map would not decide whether it was the contractor or the employer that is responsible for a particular delay. It could, however, objectively and neutrally link each of the alleged delay events to the contemporaneous notices, correspondence and other text-based project records that may be relevant to determining the material facts, together with any drawings, programmes and technical material cross-referenced in those records, while identifying the factual and expert questions that remain to be tested. The parties could still advocate their respective positions regarding how the record should be interpreted and the tribunal would still decide, the case. But they could do so through a more efficient and transparent route from proposition to source to conclusion.

Such a procedure may help the parties and the tribunal much more quickly identify the issues that genuinely require determination, expose evidential gaps earlier and focus witness and expert evidence on the questions that matter. Any reduction in time, cost or repetition would follow from that improved organisation of the case, rather than being the sole objective. It might in particular ensure that the lion’s share of issues could be fully pleaded in the first rather than a second exchange of memorials, which alone might enable many more complex disputes to be resolved within an expedited or even “highly expedited” process. It could equally improve the efficiency not only of the judicial process but also simultaneously of parallel or prior ADR processes such as adjudication or mediation.

There are already some initiatives in this general direction.  For instance, in November 2025 the American Arbitration Association launched its “AI Arbitrator”, initially for certain types of construction dispute. The AI Arbitrator guides parties through a structured submissions process from which it extracts and organises the claims, generates a comprehensive AI-produced case summary and ultimately produces a draft award for human arbitrator review and signature. 

The approach proposed here differs in at least three material respects. First, the tool would be applied close to the outset of the arbitration upon the filing of the first two initiating pleadings. Secondly, rather than extracting claims from successive position statements supported by the parties’ own selection of the evidence, the output would be based on an initial summary of claims and defences together with the entire shared documentary record without pre-selection by the parties, organising the issues by reference to the full record from the start rather than as they emerge. Thirdly, the parties would have complete control over both the selection of the tool and agreement over its specific configuration, increasing party controland transparency.  This third aspect comes with a corresponding requirement for a greater degree of collaboration, a point considered further below.

Conditions for legitimate use

A model of this kind would require careful procedural design.

First, the common corpus of data would need to be defined and identified in each dispute.  The quality and maintenance of that corpus will be material to the usefulness of the output.

Secondly, the parties will need to agree on a tool and a prompt, and on the wider configuration that shapes the output — such as the model used, its retrieval configuration and temperature setting — so that the process is adequately neutral and transparent. Because LLM output is non-deterministic and sensitive to prompt design and document order, agreement on a prompt alone will not guarantee consistent or reproducible results; the process must therefore be designed so that the configuration is held constant for the matter and any factual proposition remains traceable to the material from which it is drawn. A link to a source is not proof that the source supports the proposition, so the parties must be able to test that relationship, and the map should distinguish material that is merely source-linked from material checked for faithfulness. An AI-generated summary that cannot be interrogated should have no role in shaping the tribunal’s understanding of the case.

Thirdly, the parties must retain a meaningful opportunity to challenge the framing of the exercise itself. They may disagree not only about the answer to an issue, but about what the real issues are. They may also disagree about the relevance, completeness or interpretation of the underlying material. A shared map of issues and evidence must therefore remain open to the full adversarial toolbox of correction, supplementation and competing characterisation under the procedural direction of the tribunal.

Fourthly, the shared corpus must meet the confidentiality expectations that are fundamental to arbitration. That raises the practical question of where and how the process is run - for example on a commercial platform or in a self-hosted environment - as well as how access is controlled and how the data and any outputs are ultimately handled and deleted.

Finally, human responsibility must remain central. The purpose would not be to create a neutral technological substitute for advocacy or adjudication. It would be to create a more comprehensive and navigable evidential environment from an early stage in which conventional adversarial advocacy can operate more effectively and efficiently.

There are foreseeable objections to such a procedure. Amongst these:

• It will require a degree of collaboration at the early stage of a dispute that goes beyond anything previously required. Experience of drafting and agreeing the more modest task of a list of issues, formerly a core exercise of drawing up an ICC Terms of Reference, has shown that this can be a difficult and time-consuming exercise.
• There is a common concern that current AI tools are not yet accurate enough to use as tools for legal argumentation, with the risk that the output might be materially deficient and thereby reduce rather than increase efficiency.

However, there are equally and arguably stronger incentives to attempt such a procedure.

First, it is more likely than not that LLMs, combined with appropriate validation processes, will be able to provide a reliable tool to meet these demands shortly, if they cannot do so already. As and when they can, the market will tend to demand that they be so deployed.  If parties and their counsel do not seize the opportunity to use LLMs in such a collaborative approach, there is a real danger that LLMs will be used in other ways through which parties and their counsel will sacrifice a greater amount of autonomy over the process. For instance, there is anecdotal evidence that some arbitrators are already using LLMs to summarise the parties’ submissions and supporting evidence. Even leaving issues of transparency and confidentiality aside, it is reasonable to expect that the parties and their counsel collaboratively (or, where necessary, the tribunal after adversarial argument and with the aid of a neutral LLM expert), will arrive at a superior tool and a correspondingly better output than could be achieved in an ad hoc manner by individual arbitrators alone.

Secondly, as already noted above, there is an increasing move towards emergency and accelerated procedures, both in institutional rules and in arbitration laws. It is hard to see how these will work in a dispute of any complexity without early and collaborative assistance from AI, at least, in the absence of enormous pools of manpower.

Thirdly, such an early collaborative use of AI to produce a non-binding summary of the issues and supporting evidence - even one whose accuracy is known to fall short of certainty, provided the parties are aware of its limitations and verify the material output - could simultaneously provide an early and relatively cheap launching pad for facilitating early resolution through ADR, whether by means of adjudication, mediation or early neutral evaluation, alongside and parallel to the arbitral process, with the benefit that the output would already be a procedural step already incorporated within the anatomy of the ensuing arbitration itself should the facilitation not succeed.

Fourthly, channelling the use of LLMs into a collaborative procedure along the lines outlined above will provide some assurance against some of the specific risks of their unrestrained unilateral deployment. It is not hard to imagine, for example, aggressive uses of LLMs to generate drone-swarms of submission documents or other non-collaborative strategies of dominance that will risk overwhelming the arbitral process. Embracing a culture of collaborative use of LLMs is likely to be more effective against these risks than the repeated iteration of rules or guidelines and attempts to enforce them.

Osborne Clarke Comment

It is too early to predict precisely what an AI-native arbitral workflow will look like. The pace of change is too fast and, even when the tools are there and practices have had time to develop, the appropriate approach in each case will differ depending on the nature of the dispute, the volume and quality of the evidence, the technical issues involved, the sophistication of the parties and the tribunal’s approach to case management.

There will be cases in which the conventional memorial procedure may foreseeably remain the right answer and where AI may be deployed most usefully only as a private tool for counsel, experts or case-management teams. But there may also be a category of complex disputes, particularly where the length of calendar time to determination is at a premium, in which an initial and comprehensive shared, LLM-generated source-linked map of issues and documentary evidence may materially expedite and improve the process of understanding, managing and deciding the case.

These conditions are best not renegotiated from scratch in every dispute. Much of the work could instead be done in advance, through a standing set of default arrangements that parties adapt rather than reinvent, addressing matters such as the choice of a neutral platform and the safeguards needed for confidentiality and verification. An approach of that kind would also make the model accessible beyond the largest and best-resourced disputes, extending its benefits to the smaller or accelerated cases in which the initial set-up burden might otherwise be prohibitive.

AI is unlikely to eliminate the need for human analysis, advocacy and adjudication, at least in the foreseeable future in anything but straightforward cases. But, in the right case, it may allow the written phase to become less dependent on repeated and alternating perspectival reconstructions of the same record. That is a proposition worth testing, not because arbitration should be made faster for its own sake but because LLMs present the opportunity to improve the efficiency of the route by which a tribunal reaches its decision and, if parties and their counsel do not quickly seize the initiative, less satisfactory uses and even norms may emerge that might pose a greater threat to the autonomy and integrity of the arbitral process.

Many thanks to Simon Damschen for his contribution to this insight.

[1] The discussion took place on 6 May 2026 and the panellists were Caroline Zand-Korteweg, Opus 2/Head of Uncover, Audrey Abbot, Senior Director, FTI Consulting, Nathalie Lengert, Head of Commercial Litigation at Nokia and Dr. Arno Gildemeister, Global Head of Legal Risk, TÜV Rhein-land Group. The authors thank them for their ideas in developing that panel session.

[2] Michael Schneider, quoted from a keynote address at GAR Live Construction Disputes 2019 published at https://globalarbitrationreview.com/article/gar-live-lookback-schneider-les-sons-history-and-diversity