The rapid adoption of artificial intelligence (AI), particularly generative AI (GenAI), is transforming how professional services and financial services firms operate. AI tools are increasingly being deployed across a range of functions, from document review and transaction monitoring to customer service, compliance processes, and internal reviews and investigations. The Bank of England and Financial Conduct Authority’s (FCA) third survey of AI and machine learning in UK financial services found that 75 percent of firms have already adopted some form of AI, with a further 10 percent planning to do so within three years.

However, the use of AI raises significant legal and regulatory issues for firms to navigate. In a regulatory context, and with lawyers increasingly using GenAI tools, one potential area of risk concerns how those tools are being used in matters involving the in-house legal team or external counsel and the potential for this to impact legal professional privilege. Understanding the issues is essential for firms seeking to harness the benefits of AI whilst managing legal and reputational risks.

This is the first in a two-part series examining privilege and AI. In Part 1 we set out: (1) a recap on key privilege principles and interaction with GenAI tools; and (2) privilege considerations in internal firm-led reviews and investigations. Part 2 in this series will address: (3) privilege risks when responding to FCA regulatory inquiries and investigations; and (4) practical implications and recommendations for firms.

1) Recap on privilege and interaction with GenAI

Legal advice privilege applies to documents that are confidential communications between a lawyer and their client, made for the dominant purpose of giving or receiving legal advice. Documents protected by legal advice privilege may be withheld from disclosure to third parties. This protection extends to all forms of communication with a lawyer, including emails, WhatsApp messages, and oral communications. Legal advice privilege also covers documents that record or evidence legal advice.

Litigation privilege protects communications between lawyers or their clients and any third party created for the dominant purpose of obtaining advice or information in connection with existing or reasonably contemplated litigation (or other adversarial proceedings). Litigation privilege is wider than legal advice privilege as it extends privilege protection to communications with and documents prepared by third parties, so long as those communications and documents are for the dominant purpose of actual or contemplated litigation/proceedings. Importantly, the relevant proceedings must be adversarial, not investigative or inquisitorial. Whether investigations and inquiries can be treated as adversarial for litigation privilege purposes will depend on the nature of the investigation or inquiry and/or the stage it has reached.

A fundamental requirement of privilege is confidentiality: For a document to be privileged, it must be confidential. This often becomes an issue when, for example, privileged legal advice is circulated too widely and privilege is lost as a result. This principle is central to understanding the risks that arise when using GenAI tools.

Privilege is also recognised under the Financial Services and Markets Act 2000 (FSMA) as confidential documents and communications qualifying as ‘protected items’ under section 413 FSMA are exempt from production, disclosure or inspection by regulators like the FCA or Prudential Regulation Authority (unless withheld with the intention of furthering a criminal purpose). Generally speaking, a legally privileged communication will be a ‘protected item’ for these purposes. As a result, regulators cannot compel the disclosure of privileged or protected items (although they can ask a firm or individual to waive privilege and provide such documents on a voluntary basis).

With this in mind we consider below how privilege might apply when using a GenAI tool both in terms of putting data in and getting data out.

Putting data in: If privileged material - such as legal advice previously received from external counsel - is entered in a public AI system (such as a publicly available chatbot), this may give rise to a loss of confidentiality and waiver of privilege over that material and it could also lose its ‘protected item’ status under section 413 FSMA. A solicitor sharing confidential information could also face issues with regards to compliance with the SRA's professional conduct rules.

Using a closed or bespoke AI system with the benefit of contractual protections is lower risk but consideration should be given to who has access to the material and the extent to which confidentiality has been maintained or whether privileged material has been shared beyond the original internal “client” group and can no longer be treated as ‘protected’.

More detail on this risk can be found in our briefing: ‘Legal professional privilege in the generative AI era’.

Getting data out: It is becoming increasingly common for employees in professional services and financial services firms to use GenAI tools to seek legal guidance. A claim to privilege in relation to the resulting output would face significant challenges for a number of reasons including that a GenAI tool is not a lawyer.

However, prompts and outputs generated by a lawyer using a secure, closed AI system may be covered by legal advice privilege if they form part of the lawyer's privileged working papers which reflect the trend of legal advice being given. Again, see our more detailed briefing for further considerations.

2) Internal firm-led reviews and investigations: AI and privilege

Well-governed firms routinely consider conducting lessons learned reviews once an incident or issue has been identified that may indicate a weakness or error in the firm's systems and controls, governance, or culture. These internal reviews are an important part of good governance and regulatory compliance, enabling firms to identify root causes, remediate issues and demonstrate to regulators that appropriate steps have been taken.

However, when AI tools are deployed to support these internal reviews and investigations, careful consideration must be given to the preservation of legal professional privilege at each stage of the process. The privilege considerations in the context of internal reviews differ from those that apply in litigation because firms are unlikely to benefit from litigation privilege unless it can be argued that adversarial proceedings are in reasonable contemplation (such as threatened prosecution1). Firms conducting internal reviews must instead rely on the narrower protection of legal advice privilege (which does not extend to third parties but is limited to communications between a lawyer and their client).

Firms should therefore exercise caution when deploying AI tools in the context of internal investigations, ensuring that consideration is given at the outset to how the tools will be used and by whom and to what extent information shared with, and generated by, the tool can be treated as privileged and/or a ‘protected item’. A clear protocol should be implemented within the privilege circle with appropriate confidentiality safeguards in place.

Establishing the terms of reference

One of the initial steps in any internal review should be documenting the scope of the review in a terms of reference (or equivalent) document. This requires proper consideration of the "exam question" the firm is considering for the purpose of the review. At this stage, it is essential that consideration is given to whether the review will attract legal privilege.

Key questions to address include: Will the review "working group" include in-house legal counsel who can help assert privilege over communications and records created by that group? Will the firm be advised by external lawyers? Is the dominant purpose of the report to obtain or provide legal advice?

If an AI model is used to draft the terms of reference and if that model is an "open model" (i.e., not restricted to certain personnel within the business and available to the public), then any ‘instructions’ to the AI model and anything generated by it is unlikely to be privileged for the reasons discussed above: first, the GenAI model is not a "lawyer" - one of the fundamental requirements for legal professional privilege to apply; second, it would be difficult to argue that the document is "confidential" - a cornerstone requirement of legal professional privilege - when it is held within a publicly accessible system.

This illustrates the importance of establishing, from the outset of an internal investigation, appropriate guardrails around the creation of documents during the review, particularly where AI models are used, to preserve privilege and ‘protected item’ status under s413 FSMA.

Document collation and review

The next stage in an internal investigation is typically document preservation, searching and review. Many firms have already started using AI tools to upload documents into a secure repository (such as Harvey's "Vault" document storage function) to facilitate review. The primary attraction of using AI in this context is to reduce the time spent on first-pass document review, enabling investigators to identify relevant materials more quickly and efficiently.

However, this approach may carry privilege risks. Depending on the tool used and who has access to the vault, if no ‘triage’ review is conducted first to identify and segregate privileged documents, there is a risk that privileged materials uploaded into the AI system could lose confidentiality and privilege and cease to be classified as ‘protected items’ under s413 FSMA. Consideration needs to be given to how privileged material will continue to be protected.

Firms should therefore carefully review the terms of service of any AI tool used in document review. Consideration should be given to whether the service provider has any right to access, process, or retain documents uploaded to the system, and whether there are adequate contractual protections in place to preserve confidentiality and prevent use of client data for training purposes. Equally important is the implementation of appropriate internal access controls: firms should ensure that access to the AI tool or document repository is restricted to those individuals who have a legitimate need to review the materials, and that privileged documents are not made available to a wider audience than necessary to preserve their confidential status.

AI-generated review outputs

The outputs generated by AI tools during internal investigations will typically include timelines, chronologies, summaries of documents, or thematic analysis pertinent to the investigation. Where the AI-generated output is produced by a lawyer for the dominant purpose of providing legal advice in connection with the internal review, it may attract privilege on the basis such documents comprise the lawyers’ working papers. Nevertheless, this is still subject to the confidentiality principle highlighted above. Output generated by AI tools being used by other members of the working group is less likely to be privileged. There may be an argument where an item has been created for the purpose of seeking legal advice but there could be challenges such as the output being viewed as a communication from a third party (i.e. the AI tool) and not the ‘client’, which would also jeopardise the output’s ‘protected item’ status under s413 FSMA.

The same privilege considerations apply when using GenAI to generate investigation reports setting out the findings of the review. If the report is prepared for the dominant purpose of obtaining or giving legal advice, it may attract legal advice privilege. However, if the dominant purpose is to inform the board or management of factual findings, privilege may not apply.

Board-reporting

A further risk arises where privileged legal advice - including in relation to privileged outcomes of internal or regulatory investigations - is referenced or summarised in board minutes or other governance documents. If such material is disclosed in a board paper that is circulated widely within the organisation, privilege over the underlying advice in relation to the investigation findings may be waived. Even if the board paper is marked as confidential, wide circulation may undermine the claim to privilege. This risk is heightened where organisations use AI-powered tools to automatically transcribe board meeting minutes or to generate summaries of board discussions. If privileged material relating to an investigation is discussed during a board meeting, the AI tool will capture and transcribe that content, creating a written record that may be stored, indexed, or circulated beyond the intended audience. Such automated transcription could inadvertently waive privilege over the underlying advice in relation to the investigation, particularly if the transcripts are not subject to the same confidentiality controls as other privileged communications.

Best practice is to keep privileged legal advice separate from factual reporting to the board. Where legal advice is communicated to the board, ideally it should be set out in a separate, clearly privileged document with restricted circulation, rather than embedded in the board papers or minutes. This approach ensures that if disclosure is required to a regulator or in legal proceedings, the risk of inadvertent waiver over embedded privileged advice is mitigated, and such advice can be readily identified and withheld, whilst the remainder of the board minutes can be disclosed without redaction (if appropriate).

AI transcription tools

Firms should also exercise caution when using AI transcription tools to record internal meetings or calls during an investigation. Common use cases include transcribing witness interviews or investigation working group meetings to discuss the progress and preliminary findings of the investigation – each of which carries privilege and disclosure risks. A few key risks to consider:

  • First, AI transcription tools create a verbatim record of discussions, which could be disclosable in subsequent litigation or regulatory proceedings. This is particularly significant for transcripts of preliminary discussions where views are still being formed: if early conclusions differ materially from the final investigation outcome, a disclosable transcript could undermine the firm's position or create inconsistencies that are difficult to explain. Witness interview notes are unlikely to attract privilege in any event (unless prepared by lawyers for the dominant purpose of obtaining legal advice), and a verbatim AI transcript of such interviews would almost certainly be disclosable.
  • Second, there is a risk of waiver over privileged content in the AI-generated transcript if the model is programmed to automatically distribute to all participants in a call or meeting (some of whom may fall outside the client working group), or to store in a location accessible to a wide audience within the organisation. This should be a particular consideration for managing meetings to update stakeholders outside of the investigation working group. The broader the distribution, the more tenuous the claim to confidentiality and privilege becomes.
  • Third, even if a lawyer is included in the discussion, only those parts of the transcript concerning communications for the purpose of obtaining or providing legal advice will be protected. Simply copying or otherwise including a lawyer in a communication or discussion does not, of itself, attract privilege.

Where discussions involve sensitive legal matters, firms should carefully assess whether AI transcription is appropriate. Preparing notes through lawyers, with circulation restricted to those within the client or lawyer groups, may be preferable in order to preserve privilege.

Footnotes

  1. 1

    Director of the Serious Fraud Office v Eurasian Natural Resources Corporation Ltd [2018] EWCA Civ 2006

Contact