Redfern Schedules and Document Production: Where AI Review Earns Its Keep in Arbitration

A request to produce arrives in an international arbitration, and two reflexes follow. The first is to treat it like a US document request and hand over everything in the custodian's mailbox that touches the subject. The second is to object to all of it and let the tribunal sort it out. Both reflexes cost you, and the second costs you twice, because a tribunal that has to referee blanket objections stops believing the ones that are real.

The IBA Rules on the Taking of Evidence in International Arbitration do not give you discovery. They give you a narrow, reasoned exchange, and the Redfern Schedule is where that exchange is argued and decided.

This post covers what the IBA standard actually requires, why the Redfern Schedule is a scoping instrument rather than a form, what it takes to produce to a granted request precisely, and where AI-assisted review carries real weight in that work. The short version is that arbitration rewards the party who can produce less and explain why, and that explanation has to be built document by document.

The standard is two tests, not one

Article 3.3 of the IBA Rules requires a Request to Produce to identify either a specific document or a narrow and specific category of documents reasonably believed to exist. It also requires a statement of how the documents requested are relevant to the case and material to its outcome.

Read that carefully. Relevance and materiality are two tests, not one phrase. A document can touch the subject matter of the dispute and still fail materiality, because nothing in it moves the tribunal toward a finding. US practice trained a generation of litigators to treat responsiveness as the whole question. In arbitration it is the first half of the question.

The practical consequence is that a custodian-and-date-range population is not an answer to a request. It is the starting pool from which the answer has to be drawn.

The Redfern Schedule decides scope column by column

The schedule, named for Alan Redfern, sets out each request in its own row: the documents sought, the requesting party's justification on relevance and materiality, the responding party's objections, and the tribunal's decision. Many tribunals add a column for the requesting party's reply.

That structure is doing something specific. It forces every dispute about scope to be argued and decided one request at a time, in writing, before production. When the tribunal rules, you are not left with a general obligation to produce relevant material. You are left with a set of individually granted requests, each with its own boundary, each of which you now have to satisfy on its own terms.

This is the part teams underestimate. A schedule with eighteen granted requests is eighteen separate review criteria, not one.

Over-production is not a safe default

In a court proceeding, producing more than you had to is untidy. In arbitration it is a position. You argued for a narrow boundary, the tribunal drew one, and then you produced across it.

The costs are real. The Article 9.2 grounds you relied on, confidentiality in particular, are harder to maintain over material you volunteered. Privileged and commercially sensitive documents travel in the excess. And the credibility you spent on the schedule is gone by the next round of requests, which in a multi-phase arbitration is usually soon.

Precision is not an aesthetic preference here. It is the position you already took.

Objections need evidence, and so does testing theirs

Article 9.2 lets a tribunal exclude documents on grounds including lack of sufficient relevance or materiality, legal impediment or privilege, unreasonable burden, commercial or technical confidentiality, and considerations of procedural economy and proportionality. What the rule does not do is let you simply assert them. A burden objection that arrives without figures reads as reluctance. A confidentiality objection that cannot say what is confidential, and in what proportion of the population, invites a tribunal to grant the request and leave you to redact afterwards, which is the worse outcome.

The same applies in reverse. When the other side's production lands, the question is not whether it looks substantial. It is whether it answers the request as granted. Gaps are visible only against the boundary the tribunal drew, and finding them means reading their production against that boundary rather than skimming it.

Where AI review carries its weight

The reading in a Redfern exercise is large and the judgment is narrow. That ratio is what makes the exercise a good fit for AI-assisted review, provided the division of labour is right.

This is the approach we take with Claira inside Nuix Discover. Each granted request becomes its own criterion, written as a prompt that states the request, defines the relevance and materiality tests in the tribunal's own language, and specifies exactly what the output must contain. Claira runs that criterion across the collected population and returns, for each document, a determination, a written justification, and excerpts quoted from the document itself. Your counsel confirms or overrides every call. The machine carries the reading and the reasoning stays visible, while the decision stays with the team that has to defend it.

Two things matter more than usual in this setting. Case Context holds the matter background, the parties, their roles and the relevant periods, so the eighteenth request is assessed against the same map as the first. And because every determination arrives with a justification and a citation, an Article 9.2 objection stops being an assertion. You can say what is in the population, how much of it, and on what basis, which is the form the objection has to take to survive.

Running the other side's production against the granted request is the same exercise pointed outward. It is how you find the gap before the hearing rather than after it.

We have written before about this logic in a different procedural vehicle, in our article on Article 48 GDPR and the US discovery request, where narrowing before production is what both audiences are actually looking at.

Where to start

Three things, in order.

Treat the granted schedule as a review plan, and write one criterion per request before anyone opens a document. Build your case context once, with parties, roles and periods, so every criterion is applied against the same facts. Then test each criterion on a small benchmark set with known answers before running it at scale, which the custom prompts guide sets out step by step.

Arbitration does not reward the party that produces the most. It rewards the party that can show why what it produced is what the tribunal asked for. If you would like to see what that looks like against a live schedule, book a working session with our team.

The volume is the other side's problem. The boundary is yours.