AI Litigation Discovery: An England and Wales Playbook

AI can reduce the time spent locating documents, grouping issues and building chronologies. It can also destroy context, expose privileged material or produce a confident proposition that no authority supports. In litigation, speed is useful only when the result remains reproducible, reviewable and under a lawyer’s control.

This article states the position on 31 July 2026 and focuses on civil litigation in England and Wales, particularly disclosure in the Business and Property Courts. Different rules and practice apply in Scotland, Northern Ireland, criminal proceedings, tribunals, arbitration and investigations. It is operational information, not legal advice; the case team must apply the court’s orders and obtain specialist advice.

The rulebook starts before a model is selected

Practice Direction 57AD treats documents broadly: electronic files, messages, audio, video and other records can all be documents, including deleted material and metadata. Parties and their legal representatives have continuing duties to preserve potentially relevant documents, disclose known adverse documents and cooperate on a proportionate process.

PD57AD expressly recognises technology-assisted review, but recognition is not permission to run an unexplained black box. Parties may need to discuss data sources, search methods, review techniques and formats in the Disclosure Review Document. The court can scrutinise proportionality and direct the process.

Outside that regime, Practice Direction 31B remains important for electronic disclosure. It says preservation should be considered as soon as litigation is contemplated, encourages efficient use of technology and recognises native files and metadata. CPR Part 31 and the order in the individual matter remain essential. A model cannot infer the governing disclosure model from a generic prompt.

Preserve first, then build a defensible data map

The legal hold should precede collection and AI analysis. Identify custodians, devices, cloud systems, collaboration tools, personal accounts used for business, mobile messages, archives, backups and departing-worker data. Suspend routine deletion where proportionate and document the decision. Preserve native files and metadata; avoid opening or converting files in ways that alter evidence.

Create a matter data map with:

• source system, custodian and business owner;
• relevant date range, issue and jurisdiction;
• volume, format and encryption state;
• collection method, operator and timestamp;
• hash or other integrity control;
• retention, legal-hold and deletion status;
• likely personal, special-category, privileged or confidential content;
• processor, hosting region and access route.

The map is not a one-off inventory. Update it when pleadings change, a new custodian appears or an adverse document reveals another repository. Connect each collection decision to the issue it serves. Our legal contract review operating guide describes a related evidence-first pattern, but litigation requires additional preservation and disclosure controls.

Choose a bounded AI task

“Prepare the case” is not a testable use case. Start with one task whose inputs and correct outputs can be reviewed: near-duplicate detection, email threading, language identification, proposed issue tags, date extraction or candidate chronology events. Retrieval can help a lawyer locate passages; it should not be treated as proof that no other relevant material exists.

Use deterministic tools for exact deduplication, hashing, date normalisation and export. Use statistical or generative systems only where their uncertainty is visible. Never let a model overwrite originals or promote a document directly into an agreed bundle.

Design review around recall and asymmetric risk

Missing one decisive adverse document can matter more than reviewing hundreds of irrelevant files. Evaluation must therefore reflect the case, not a vendor’s generic accuracy claim. Build a stratified test set that includes key custodians, issue periods, unusual file types, short messages, scanned records, non-English material, privileged patterns and known adverse examples.

Have senior reviewers adjudicate disagreements and record the rationale. Measure recall and precision by issue, custodian and document class, rather than one blended score. Sample documents the system marked irrelevant as well as those it promoted. If active learning or continuous training is used, preserve every training round, seed set, reviewer decision and model version.

The operating sequence should be:

• Agree issues, sources and review protocol with the responsible solicitor.
• Preserve and collect using forensically appropriate methods.
• Process copies in an isolated matter workspace.
• Run deterministic filters and exception reports.
• Apply the tested model to propose, not finalise, labels.
• Conduct senior review of adverse, privileged and low-confidence material.
• Perform quality-control sampling across included and excluded populations.
• Export native files, metadata and audit records in the agreed form.

If the team cannot recreate why a document was included or excluded, the process is not defensible.

Privilege is a legal judgement, not a keyword

Names of lawyers, “privileged” banners and legal vocabulary are weak proxies. Communications involving in-house counsel may mix legal and commercial purposes. Attachments can have a different status from covering emails. Common-interest, without-prejudice and litigation-privilege questions are fact-sensitive.

Use AI only to route likely privileged material into a restricted queue. Combine participant lists, domains, matter numbers and communication patterns with human analysis. Keep a privilege log where required and give senior lawyers a controlled route to reverse coding. Prevent privileged examples from becoming training data for another matter or a provider’s general model.

Accidental disclosure requires a rehearsed response: stop review, restrict access, preserve logs, notify the responsible solicitor and follow the governing rules and order. Do not rely on a vendor’s “delete chat” button as proof of deletion.

Hallucinations and authority verification

The Judiciary’s October 2025 AI guidance warns about hallucinations, bias and confidentiality and stresses personal responsibility for material produced in a user’s name. A fluent citation is not an authority.

Every proposition generated with AI should resolve to an official judgment, legislation database, procedural rule or licensed authoritative source reviewed by a lawyer. Verify the party names, neutral citation, court, date, quoted passage, later treatment and applicability. Keep research notes showing the actual source, not merely the prompt and output.

Chronology output needs the same discipline. A date extracted from an email header, attachment, OCR layer or quoted chain may describe a different event. Preserve the passage and document identifier beside each proposed entry, distinguish document date from event date, and flag conflicts rather than choosing one silently. Witness teams should see the underlying record before adopting the model’s sequence. Never use generated wording to coach a witness into an account that the documents do not independently support.

The Law Society’s generative AI essentials, updated in June 2026, similarly emphasise verification, confidentiality, provider due diligence and professional duties. Our AI [paralegal and legal-tech guide](/blog/ai-paralegal-legal-tech-trends-uk) explores role design; it should never be read as delegating professional responsibility to software.

Confidentiality, data protection and security

Case collections commonly contain personal data, trade secrets, health information and third-party records. Legal professional privilege exemptions under data-protection law are specific, not a blanket suspension of UK GDPR. The ICO’s exemptions guide should be applied to the actual processing and request.

Document purpose, lawful basis, minimisation, access, retention and deletion. Consider a DPIA where processing is likely to create high risk. Segregate matters and clients cryptographically and logically. Prevent public-model submission, cross-tenant retrieval and unrestricted downloading. Test malicious documents for prompt injection, hidden text, macros and links before model ingestion.

Supplier contracts should state processing instructions, subprocessor controls, hosting and transfers, incident notification, return and deletion, audit support, model-training prohibition, export formats and cooperation with court deadlines. Follow the NCSC’s secure AI development guidance, and use our UK data privacy and AI compliance guide for the wider governance workflow.

Supervision and matter-level accountability

The SRA Code of Conduct for Solicitors requires competence, accurate and non-misleading communications, proper service and observance of duties to the court. The SRA’s effective supervision guidance makes clear that an authorised individual retains ultimate responsibility and that AI output needs appropriate human review.

Name a responsible solicitor for each AI use. Record its permitted task, sources, reviewers, quality thresholds, privilege workflow, security owner, export procedure and shutdown route. The case theory, witness evidence, disclosure decisions and submissions remain professional judgements.

A 90-day controlled pilot

Days 1–30 — protocol. Select one closed or safely duplicated matter dataset. Confirm procedural scope, preservation and security. Inventory tools and processors. Build an adjudicated evaluation set and baseline review hours, recall, precision, coding disagreement and privilege escapes.

Days 31–60 — blind comparison. Run the model without changing the production review. Compare its proposed coding with senior-review outcomes. Investigate misses by issue and source. Test permissions, export, deletion, prompt injection, OCR failures and a provider outage.

Days 61–90 — supervised live use. If the closed test passes, use the tool on a bounded live tranche with parallel quality control. Document the protocol, confer where required, sample excluded records and ask an independent lawyer or litigation-support specialist to challenge the evidence trail.

Expansion gates should require:

• no alteration or loss of source evidence or metadata;
• 100% traceability from material output to its source passage;
• agreed recall thresholds met for every high-risk issue and custodian;
• zero unresolved privileged-material exposure;
• complete reproducibility of collections, model versions and review rounds;
• measurable review-time reduction without more substantive corrections;
• successful export in the court-agreed native and metadata format.

Pause immediately if preservation is uncertain, scope changes without retesting, the provider uses matter data for training, a citation cannot be verified, or excluded-population sampling finds a material miss. Suspend generation after a confidentiality incident or unexplained model change. The goal is not an automated lawsuit. It is a disciplined review system in which lawyers can explain every consequential step.

Primary sources checked

• CPR Practice Direction 57AD, Ministry of Justice
• CPR Practice Direction 31B, Ministry of Justice
• CPR Part 31, Ministry of Justice
• Artificial Intelligence: Judicial Guidance, Judiciary of England and Wales, October 2025
• Generative AI: the essentials, Law Society, updated June 2026
• SRA Code of Conduct for Solicitors
• Effective supervision guidance, SRA
• Data-protection exemptions guide, ICO
• Guidelines for secure AI system development, NCSC