AI for Litigation Support and Disclosure Management
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
You're 7 lessons in — don't lose your progress.
Sign up free to save where you are and earn a verified certificate when you pass.
- Explain how Technology-Assisted Review works — seed documents, iterative training, and responsive/non-responsive classification — and what the solicitor's supervisory obligation over the TAR process requires
- Apply the Pyrrho Investments v MWB Business Exchange standard for reasonable search to a TAR-assisted disclosure exercise
- Identify the ways AI can support case strength analysis — chronology synthesis, evidential gap identification, and litigation risk modeling — and the limits that require lawyer judgment to cross
- Describe the verification obligation that applies to every AI output in a litigation context, including the professional consequences of submitting unverified AI-generated case citations to a court
Modern litigation disclosure has a scale problem that manual review cannot solve at any commercially rational cost. A single mid-size commercial dispute can involve tens of thousands of emails, attachment chains, system exports, and custodian communications. Document-by-document manual review of a corpus that size would consume more time than the matter economics justify and more fee earner capacity than most practices can spare. Technology-Assisted Review has moved from a technique used only in the largest US discovery exercises to a standard tool in substantial English litigation — and the supervisory obligations it creates for the instructing solicitor are now well established.
Technology-Assisted Review: How It Works and What the Solicitor Owns
Technology-Assisted Review — TAR, also known as predictive coding — is a machine-learning approach to document review that trains a classification model on a set of human-reviewed seed documents and then applies that classification to the broader document population.
The TAR workflow
The process begins with a seed set: a sample of documents that a qualified reviewer — typically a senior solicitor or experienced litigator — reviews and classifies as responsive (relevant to disclosure) or non-responsive. That human classification trains the model. The model is then applied to the full document population, predicting the likely responsiveness of each document. The process is iterative: the model's predictions are sampled and reviewed, corrections feed back into the training, and the cycle continues until the model reaches a level of recall and precision that is assessed as reasonable. Documents the model classifies as responsive are reviewed by lawyers before disclosure. Documents classified as non-responsive are typically excluded from disclosure, subject to quality-checking measures.
The Pyrrho standard
In Pyrrho Investments Ltd v MWB Business Exchange Ltd [2016] EWHC 256 (Ch), Master Matthews approved the use of predictive coding in an English litigation for the first time. The judgment established that predictive coding is a legitimate approach to the "reasonable search" obligation under CPR 31.7, provided the process is documented, the training seed set is of adequate quality, the model's performance is monitored, and the opposing party has been notified. The court noted that predictive coding could produce better recall than manual review at significantly lower cost — the concern was not that it was less reliable than manual review, but that it required appropriate process governance to be defensible.
The supervisory obligation
TAR does not reduce the solicitor's supervisory responsibility — it relocates it. Instead of supervising individual document review decisions, the solicitor supervises the process: the quality of the seed set, the qualifications of the reviewers providing the training classifications, the model's recall statistics at each iteration, the protocol for handling borderline classifications, and the quality-check sampling on documents classified as non-responsive. The solicitor who instructs a TAR process without understanding what decisions the model is making and on what basis is not discharging their supervisory obligation — they are merely outsourcing it to software.
When instructing a TAR process, document four things before review begins: the composition and size of the seed set, the seniority and qualifications of the reviewer providing the training classifications, the recall and precision metrics you are targeting, and the quality-check sampling methodology for non-responsive documents. This documentation is the evidence of a defensible process if the opposing party or court later challenges the adequacy of disclosure. Agreeing the TAR protocol with the opposing party in advance — which Pyrrho specifically encouraged — significantly reduces the risk of a later dispute about the process.
AI for Case Strength Analysis
Beyond disclosure management, AI tools are increasingly used to support case analysis — synthesizing chronologies, identifying evidential gaps, modeling litigation risk, and stress-testing arguments. The utility is real; so are the limits.
Chronology synthesis. Building a comprehensive case chronology from a large document set combines factual recall with document retrieval skill. AI can ingest a reviewed document set and construct a working chronology, identify documents attached to each chronological event, and flag apparent gaps in the timeline. This is a legitimate efficiency gain — the starting chronology AI produces is a useful scaffold for the lawyer's own analysis.
Evidential gap identification. AI can identify where the factual narrative has gaps — periods where the documentary record is thin, communication threads that start but do not conclude, representations that appear in correspondence but are not supported by contemporaneous documentation. Surfacing these gaps early is genuinely useful; the significance of each gap and how it affects litigation strategy is a matter of legal judgment.
Litigation risk modeling. Some tools offer AI-assisted analysis of analogous case outcomes, allowing a solicitor to test how cases with similar factual patterns and legal issues have been decided in comparable proceedings. This is a form of pattern-matching on known case law, not a prediction of what a specific judge or tribunal will decide. Using it to stress-test a client's case theory or to prepare a realistic litigation risk assessment for a without-prejudice settlement discussion is a legitimate application. Treating its outputs as a reliable prediction of outcome is not.
Document contradiction analysis. AI tools can compare across large volumes of witness evidence and identify contradictions: a witness statement that describes events differently from contemporaneous emails, or a version of events in a pleading that is inconsistent with disclosed documents. Identifying those contradictions for lawyer review is a useful preliminary task — assessing their significance, deciding how to deploy them, and judging witness credibility remain exclusively matters of legal judgment.
A litigation team is using AI to analyze disclosed documents and has received an AI-generated chronology and a list of apparent evidential gaps. A trainee asks whether the AI output can be sent directly to the client as a case analysis. What is the correct response?
Select one answer.
The Verification Obligation in Litigation
The professional consequences of submitting unverified AI output to a court are now documented. In US proceedings, the case of Mata v Avianca saw lawyers sanctioned after AI-hallucinated case citations — plausible-sounding but entirely fabricated — were submitted in court filings. The lawyers had not verified the citations in primary sources before filing; the AI's confident presentation of fictional authorities was treated as a professional conduct failure, not merely an honest mistake. The pattern that case illustrates is directly applicable to any jurisdiction where AI tools are used in litigation preparation without adequate verification discipline.
The risk is not limited to citations. AI-generated case summaries may inaccurately describe the ratio of a real case. AI-generated procedural summaries may reflect the rules of a different jurisdiction. AI-generated document analysis may misread a key exhibit. In every instance, the professional obligation is the same: verify the specific claim before it is placed before a court or tribunal.
The verification standard in English litigation is high. The duty of candour to the court — addressed in the ethics and professional responsibility lesson — is an absolute obligation. Courts have consistently held that lawyers knew or should have known that AI outputs require verification, and that reliance on an AI tool is not a defense to submitting inaccurate materials to a court. The legal research lesson addresses the specific verification discipline required when AI is used to identify case law.
Before citing any case in a court or tribunal document that originated from an AI research tool, verify it in a primary legal database — Westlaw UK, LexisNexis, the National Archives, or the court's own public judgment service. AI tools can hallucinate case names, citations, dates, judges, and the substance of the judgment. The hallucinated citation is often plausible enough to pass a casual read. It will not pass opposing counsel's check. Submitting fabricated case authority to a court is a serious professional conduct matter regardless of whether the AI generated it or the lawyer did.
TAR in a Complex Commercial Disclosure Exercise — City Litigation Practice
Context
A litigation solicitor was instructed on a substantial commercial dispute involving allegations of misrepresentation across a multi-year commercial relationship. The document population — drawn from email archives, CRM system exports, and board minutes across multiple custodians — ran to several hundred thousand items. Manual review at the matter's fee budget was not viable.
Action
The solicitor proposed a TAR process to the opposing party, circulated a draft protocol, and agreed the review methodology before commencing. A senior associate provided the training classifications on the seed set, and the model's recall statistics were monitored across three training iterations. A litigation support vendor with documented TAR experience was retained and required to confirm in writing that the process met the Pyrrho standard. A quality-check sample of non-responsive documents was reviewed at each iteration.
Outcome
The disclosure process was completed within the matter budget and within a timeframe that manual review could not have achieved. The opposing party did not challenge the TAR process, in part because the protocol had been agreed in advance. The solicitor's post-matter reflection identified that the supervisory work — designing the protocol, monitoring the iterations, and reviewing the quality-check samples — had required more senior input than initially budgeted, and that this time needed to be scoped into TAR matters from the outset.
A solicitor uses an AI legal research tool to identify supporting case law for a skeleton argument. The tool returns six cases with citations and short summaries of each holding. The solicitor checks that the legal propositions make sense, incorporates four of the cases into the skeleton, and files it. The following week, opposing counsel writes to say that one of the four cited cases does not exist. What professional duty has most clearly been breached?
Select one answer.
Exercise
Your Task
Take a disclosed document set from a current or recent matter — or, if you are in training, construct a hypothetical set of ten documents representing a simple commercial dispute. Use an AI tool to build a chronology from those documents, then check every entry in the AI-generated chronology against the source documents it references. Record how many chronological entries were produced correctly without amendment, how many required correction, and what categories of error appeared — wrong date, wrong party attribution, misread document content, omitted event. Then assess whether the AI-generated chronology would have been reliable enough to send to a client without the verification step.
Success looks like
- You have checked every AI chronology entry against its source document rather than spot-checking only
- You have categorized the errors found — date errors, party attribution errors, substance errors — rather than simply counting them
- You can articulate a calibrated rule for when AI-assisted chronology work produces a reliable starting point and when the error rate requires near-complete rebuilding
- You have identified at least one specific type of document the AI misread and can explain why the misreading occurred
Watch out for
- Checking only whether the AI summary sounds right rather than verifying against the actual source document — this is precisely the error pattern the lesson identifies in the citation verification context
- Drawing a blanket conclusion from a small document set that AI chronology work is either universally reliable or universally unreliable — the exercise asks for a calibrated view, not a binary one
Hint
Focus on documents where the same event appears in multiple sources — emails, board minutes, and attachments that all reference the same transaction or decision. These are the documents where AI most commonly introduces errors by conflating different accounts of the same event.
- Technology-Assisted Review is a legitimate approach to reasonable search under CPR 31.7 following Pyrrho Investments v MWB Business Exchange, provided the process is documented, the seed set is of adequate quality, the model's performance is monitored across iterations, and the opposing party has been notified.
- The solicitor's supervisory obligation over a TAR process is not eliminated by using the technology — it is relocated to the design of the process, the quality of the training classifications, the monitoring of recall statistics, and the quality-checking of non-responsive documents.
- AI case strength analysis — chronology synthesis, evidential gap identification, litigation risk modeling — produces useful starting points for qualified legal analysis but cannot substitute for the lawyer's judgment on credibility, strategy, and how a specific tribunal will respond to the arguments.
- Verifying every AI-generated case citation in a primary source before filing is a non-negotiable professional obligation: AI tools hallucinate plausible-sounding but entirely fictional case authorities, and submitting fabricated citations to a court is a professional conduct matter regardless of the tool that generated them.
- The duty of candour to the court is an absolute obligation that AI generation cannot excuse — courts have consistently held that lawyers knew or should have known that AI outputs require verification, and reliance on an AI tool is not a defense to placing inaccurate materials before a tribunal.