AI in Witness Bundle and Proof of Evidence Preparation: What to Automate and What Not To
By the ChamberOne team 6 minute read
Witness-heavy litigation is exactly the kind of work where AI can save hours: collating documents, building chronologies, spotting date discrepancies, flagging passages in one statement that sit awkwardly with another. But it is also work where a misjudgement about what to hand to a machine creates the most acute professional risk. Courts in England and Wales have already found evidence prepared with AI assistance to be less than satisfactory, and regulators across jurisdictions are tightening their expectations. This article sets out, task by task, what AI genuinely helps with in building a witness bundle and a proof of evidence, what it cannot reliably do, and where the credibility assessment, the instruction-taking and the final narrative must stay entirely with the lawyer.
Key points
- AI can legitimately help with document collation, chronology-building, bundle indexing and cross-referencing inconsistencies across statements, but must not generate or rephrase witness statement content.
- In England and Wales, Godwin v Godwin [2026] EWHC 923 (Ch) shows that even using AI for grammar and presentation caused the court to treat the evidence with caution when original drafts were not retained.
- The CJC's October 2026 Final Report recommended that witnesses of fact be required to declare that AI has not generated or influenced the substantive content of their statement; these recommendations are not yet in force.
- If AI transcription is used for interview notes, retain the original recording: AI transcription can misrender what was said, and the recording is the only means of verification.
- Brief expert witnesses in writing on AI use limits before they prepare their reports; courts are increasingly requiring disclosure of any AI that has influenced an expert's methodology or conclusions.
What a witness bundle and a proof of evidence require
A witness bundle is the collection of documents a witness will be referred to when giving evidence: correspondence, contracts, photographs, records, prior statements. Its value depends entirely on completeness and accurate pagination. A proof of evidence is the lawyer's working document setting out what the witness is expected to say, structured around the issues in the case.
In England and Wales, CPR Part 32 and Practice Direction 32 set the framework for witness statements. Rule 32.4 defines a witness statement as a written statement signed by the person containing the evidence that person would be allowed to give orally, and Rule 32.8 requires compliance with the Practice Direction. For trial witness statements in the Business and Property Courts, Practice Direction 57AC goes further, requiring statements to be in the witness's own words and demanding transparency about how the statement was prepared. Those requirements create a direct tension with the use of generative AI in drafting.
What AI can usefully do: organising and cross-referencing
The preparatory work before a witness even sits down with a lawyer is where AI saves the most time with the least risk. AI tools can be asked to collate and chronologise documents, reducing preparation time considerably. Grouping correspondence by date, identifying which documents relate to a particular event, flagging where the same date appears in different document sets — these are well within what current tools handle reliably.
AI can also compare witness statements, build event timelines, and flag credibility issues across prior testimony. For the bundle itself, AI can assist with indexing, checking that all documents referenced in a statement are actually included, identifying duplicate pages and suggesting a logical ordering. These tasks benefit from systematic attention to detail but require no judgment about what the evidence means.
- Use AI to collate, chronologise and cross-reference documents before instruction-taking
- Use AI transcription to capture an interview verbatim, retaining the recording for verification
- Use AI to compare statements and flag date discrepancies or inconsistencies
- Use AI to build a draft bundle index and check that cited documents are present
- Allow AI to generate or rephrase the substantive content of a witness statement
- Use AI to coach a witness on what to say or how to frame their answers
- Let AI produce a narrative proof without the lawyer conducting a full independent interview
- Submit AI-assisted content to court without keeping the original drafts for scrutiny
Transcription: a permissible use, with conditions
AI transcription is one area where the procedural rules in England and Wales expressly leave room for assistance. AI could be used to transcribe interviews conducted by lawyers with the witness before editing the text into a draft witness statement, keeping the text as close to verbatim as possible to retain the witness's own words. The finished statement still requires the witness's review and sign-off, but transcription from a recorded interview at least anchors the document to something the witness actually said.
If AI transcription is used, the underlying recording should be retained. If the accuracy of the transcription is ever challenged — including whether the AI has introduced content that was not spoken — the original recording is the only way to verify it. AI transcription tools can misrender what was said, particularly with technical terminology, names, or accented speech, and those errors require checking before the transcript is used as a source for any formal statement.
Outside England and Wales, practitioners should check local rules. Courts in Australia, Singapore, the United States and India each have their own requirements governing witness statements and affidavits, and what is permissible in one jurisdiction may not be in another.
What courts have said when AI touches the substantive content
Two cases decided in England and Wales in 2026 show clearly where the line is drawn. In Godwin v Godwin [2026] EWHC 923 (Ch), delivered in April 2026, witnesses used ChatGPT to assist with grammar, spelling and presentation of their witness statements. The judge concluded there was no good reason for either witness to have used ChatGPT, being sophisticated people who could have used standard word-processing tools instead. Because the original drafts were not retained, the court could not be sure the statements were in the witnesses' own words, and approached the evidence cautiously. The case was ultimately sent for retrial.
In R v FGD [2026] EWCA Crim 918, it emerged during trial that a prosecution witness had used AI tools to prepare for cross-examination. The Court of Appeal described this as witness coaching by an AI tool and confirmed that coaching witnesses in their evidence, by any means, is not permitted.
The Civil Justice Council's Final Report on the Use of AI for Preparing Court Documents, published in October 2026, recommended that the CPR be amended to require a witness of fact to declare that AI used in preparing their statement has not generated or influenced its substantive content. The report defines generating substantive content as the production or modification of substantive content by an AI tool, and notes that using AI to generate or influence substantive content can be viewed as analogous to witness coaching. These are recommendations to the Civil Procedure Rule Committee and are not yet in force.
Expert witnesses: a related problem
Expert witnesses face parallel scrutiny. The CJC's October 2026 report also recommended that experts declare where AI has generated or influenced the substantive content of their report, with enough information disclosed to allow the opposing party and the court to scrutinise the methodology. The report identified examples of experts including AI-produced fictitious content in reports and experts instructing AI tools to produce reports that vindicate the instructing party.
For the instructing solicitor, this creates a practical duty. When briefing an expert, it is prudent to provide written guidance on whether and how AI may be used. At a minimum, the expert should be reminded that client confidential information must not be supplied as inputs to AI tools, and that any AI-generated content must be checked against underlying sources for accuracy.
Courts in the United States are grappling with related questions. In Kohls v Ellison, No. 24-CV-3754 (D. Minn. Jan. 10, 2025), a court excluded an expert's testimony in part because of hallucinated citations found in the expert's materials. Separately, in a New York surrogate's court decision, Matter of Weber (N.Y. Sur. 2024), a court held that AI-based evidence should go through a Frye hearing before being admitted. Together, these cases illustrate a consistent international direction: courts expect transparency about AI use in evidence, and the burden of demonstrating reliability falls on the party relying on the material.
What AI cannot supply: credibility, instruction-taking and the final narrative
No AI tool can assess whether a witness will be believed. Credibility depends on how a person presents, the consistency of their account under pressure, and the small hesitations that give experienced lawyers an instinct about where the case is strong. An AI summary of a statement can tell you that a witness claims to have been in a particular location at a particular time. It cannot tell you whether that claim will survive cross-examination.
Instruction-taking is similarly irreplaceable. Sitting with a witness, asking open questions, allowing the account to emerge in the witness's own words, and testing it gently against the documents is not a task that can be delegated to a tool. The CJC's October 2026 report explicitly noted that an unbiased interview approach helps guard against the creation of false memories through leading questioning.
The final narrative — the structure of a proof of evidence that explains how the evidence fits together and identifies where the witness's account may be vulnerable — requires legal judgment that sits entirely with the lawyer. AI can produce a first draft outline as a prompt to check whether anything has been missed. It should not be the source of the analysis itself.
Practical steps for litigators
The boundary that emerges from the case law and the CJC's recommendations is clear in principle: AI may assist with the administrative and organisational work around witness evidence, but it must not generate or influence the substantive content of what a witness says. That means structuring the workflow so that AI assistance happens before the interview and in building the bundle, while the lawyer conducts the instruction-taking and shapes the final account.
- 1
Use AI before the witness interview: document collection, chronology, cross-reference of prior statements and disclosed documents.
- 2
Conduct instruction-taking personally, with open questions, and record or note it.
- 3
If AI transcription is used, retain the recording and check the transcript for accuracy before it is used.
- 4
Do not use AI to rephrase, restructure or improve the witness's own words in a formal statement.
- 5
Keep a clear written record of exactly how AI was used at each stage of preparation.
- 6
Brief experts in writing on the limits of permissible AI use and the disclosure requirements that apply.
Keep records of how every witness statement was prepared. The CJC's report noted that witnesses will need sufficient information from their lawyers regarding the preparation of the statement in order to make the proposed declaration about AI use, which means lawyers must be able to give that account accurately.
Before using any AI tool on client witness materials, check the tool's data-handling terms. Sending witness statements or interview transcripts to a third-party AI service raises confidentiality, privilege and data-protection questions that vary by jurisdiction.
Sources
- Civil Justice Council, Use of AI for Preparing Court Documents: Final Report (October 2026)
- Lewis Silkin: CJC final report on AI and court documents — declarations for witnesses and experts (November 2026)
- Taylor Wessing: AI in focus — AI's possible role in witness statement preparation (July 2024)
- Reed Smith: (A)I believe that the facts stated in this witness statement are true (2026)
- CMS Law: Use of AI in court — can the rules catch up? (2026)
- Doughty Street Chambers: The dangers of using AI to assist with witness statements — lessons from Godwin v Godwin (2026)
- Greenberg Traurig: Expert testimony in the age of generative AI — recent case developments (2025)
- Justice.gov.uk: Practice Direction 32 — Evidence
How ChamberOne handles this
- Documents uploaded to a matter — including PDF, Word and other file types — can be analysed by AI, which lists parties, key dates, clauses and citations, each linked to the paragraph it came from.
- Risk flags on clauses are produced as suggestions that a partner can override, not as final conclusions.
- An AI matter brief can be requested to summarise the matter when a lawyer needs to pick up an unfamiliar file, produced as a draft for review.
- Every AI call is recorded in the audit log with the model used and token counts, so the firm has a record of when AI was applied to a matter.
- The firm can switch AI features off entirely; with AI features off, nothing is sent to an AI provider.
What it does not do
- AI output can be wrong or incomplete, can miss or misread things, and needs checking against the source document.
- The audit log records that an AI call happened; it does not confirm that the output was correct.
- ChamberOne does not generate witness statements or proof of evidence documents, and does not replace the lawyer's interview and judgment.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
Hallucinated citations: what court sanctions teach about checking AI output
The patterns behind fake authorities reaching courts, and the supervision and checking controls that would have caught them.
5 minute readE-discovery and technology-assisted review explained for non-specialists
How predictive coding and generative review differ, why validation matters and how to document the process.
5 minute readBuilding a litigation chronology with AI
Extract dated events, link each to its source page and check for conflicts before the chronology is used.
5 minute read