Skip to content

E-discovery – how to use predictive coding

27 March 2018

The court considered the use of Computer Assisted Review (CAR, also known as ‘predictive coding’) in the disclosure process. The judgment provides useful guidance on how CAR should be used in practice. A party should not act unilaterally to use CAR in the disclosure process as there is a risk that the decision will be carefully scrutinised by the court at a later date, and that a different course may be ordered. Also, if CAR is to be used, even where its use is consensual, a party should ensure that the technology is properly ‘taught’ and that the criteria for determining relevance are consistently applied at this ‘teaching’ stage: Triumph Controls UK Ltd & anr v Primus International Holding Co & ors [2018] EWHC 176 (TCC), 7 February 2018

In this USD 65 million share sale breach of warranty claim, involving around 450,000 potentially disclosable documents after keyword searches, the claimant did not mention any intention to use CAR in its Electronic Documents Questionnaire (EDQ) or agree with the defendants that it could be used. The claimant subsequently gave the defendants incomplete or vague information about how CAR had in fact been used.  The defendants therefore applied to the court for an order that the claimant undertake a manual review of the documents which had been identified as potentially disclosable but, following the use of CAR and except as part of a very limited sampling exercise, had not been reviewed at all.

The use of CAR in disclosure – discussion, cooperation and agreement between the parties from the outset is generally required

The court held that a party should generally not act unilaterally to use CAR in the disclosure process.

Discussion, cooperation and agreement between the parties is required from the outset unless there is a very good reason why this should not take place (as per Practice Direction 31B, which deals with the disclosure of electronic documents, and the TeCSA/TECBAR e-disclosure protocol, which is used as a guide in TCC cases involving electronic disclosure).

A failure to cooperate can lead to the court requiring that party to carry out further searches for documents or to repeat other steps which that party has already carried out (following Digicel (St Lucia) Ltd and ors v Cable & Wireless Plc & ors [2008] EWHC 2522 (Ch)).

A party intending to use CAR should therefore raise this with the other side as early as possible to avoid potential misunderstandings and limit the chances that the decision to use it will be carefully scrutinised by the court at a later date, and a different course ordered.

Detailed information required about how CAR was used if party acts unilaterally

When a party has acted unilaterally, that party should provide the other side with details about how the CAR was set up and how it was operated. That had not happened in this case, which the judge found to be “unsatisfactory”. He also found it unsatisfactory that the claimants had provided no information to the defendants about precisely how the sampling exercise, which was done following the CAR process to assess the relevance of any documents that the claimants no longer intended to review, had been conducted.

CAR technology must be ‘taught’ properly: best practice involves senior lawyer oversight

A party should ensure that CAR technology, which generally relies on supervised machine learning, has undergone a proper teaching process and that the criteria for relevance are consistently applied at this supervised ‘teaching’ stage. This point had previously been made in the case of Pyrrho Investments Ltd & anr v MWB Property Ltd & ors [2016] EWHC 256 (Ch), where the court had stated that “best practice would be for a single, senior lawyer who has mastered the issues in the case to consider the whole [teaching] sample”.

In this instance, the claimants’ lawyers had involved up to ten paralegals and four associates in the training process and it was not apparent that there had been any senior lawyer overseeing the work. The court noted on this basis that the CAR system may not have been ‘educated’ as well as it might have been, in particular as to the criteria for relevance.

Judge orders manual review

The court ordered the claimants to conduct a manual review of a 25% sample of the 220,000 disputed documents (55,000 in total) within three weeks, with a view to hearing further from the parties at the next scheduled hearing in the matter. The judge considered this to be proportionate in the circumstances, bearing in mind the size of the claim and the trial date being less than six months from the date of the hearing. There was no decision on costs at this stage, although it can be expected that the claimants will be penalised for embarking on a unilateral approach to disclosure.


This case provides litigants with a high-level guide to using CAR, or predictive coding, properly in the disclosure process. There is now more certainty around the use of these technologies and the potential consequences for a failure to follow best practice. This may in turn encourage more parties to make use of CAR in the disclosure process, which could substantially cut the costs involved with this aspect of litigation. 

Further information

This case summary is part of the Allen & Overy Litigation and Dispute Resolution Review, a monthly publication.  If you wish to receive this publication, please contact Amy Edwards,