Privilege and Generative AI: Function, Not Form

Winona Chan sets out the fundamentals of legal privilege and how to protect it when using generative AI tools.

Generative AI (GenAI) is now embedded in day-to-day business workflows, from drafting documents to analysing information and summarising meetings. For in-house legal teams, this raises a critical question: how can legal professional privilege be preserved when AI tools are used across an organisation?

The answer lies in applying established privilege principles to new workflows. AI does not change the law on privilege, but it does change how documents are created, shared and stored. This creates new risks, particularly around confidentiality and the blurring of legal and commercial work. This article outlines a practical framework for managing those risks.

Function over form

A persistent misconception is that privilege can be created simply by involving lawyers: for example, by copying in the legal team or routing documents through them after the event. That is not the position in English law.

Privilege depends on function, not form. The key question is whether a communication was created for the dominant purpose of seeking or giving legal advice, or whether its disclosure would reveal that advice. It cannot be applied retrospectively to documents that were originally created for commercial purposes. This matters in an AI context, where businesses increasingly generate polished analyses and summaries before any lawyer is involved. If those materials are created outside a legal advice context, they are unlikely to attract privilege later.

The “two lanes” model

A useful operating model is to divide work into two distinct lanes:

• The business productivity lane where AI is used for routine drafting, summarising and internal analysis. The working assumption here should be that privilege does not apply.

• The legal lane where work is carried out for the purpose of obtaining or giving legal advice. This lane should involve lawyers from an early stage, use approved tools, and operate within controlled processes.

Keeping these lanes separate is critical. Problems arise where AI outputs mix business and legal content, or where legal issues are explored in general business workflows rather than through a structured legal intake process.

Confidentiality

Confidentiality is fundamental. If it is lost, privilege cannot be claimed.

AI tools present varying levels of confidentiality risk. Public chatbots should generally be treated as high risk for privileged material, as prompts and outputs may be stored, processed or accessed by third parties under provider terms. By contrast, enterprise AI tools and specialist legal platforms can present a lower risk if properly configured, particularly where access, retention and data use are tightly controlled.

The key questions for any tool are practical:

• Where are prompts and outputs stored?
• Who can access them?
• How long are they retained, and are they searchable?
• Are inputs used to train or improve the model?
• Is data shared internally or externally?

AI changes the document trail

One of the most significant impacts of AI is the expansion of the document trail. AI tools generate a wide range of artefacts, including prompts, chat histories, intermediate drafts, logs and, in some cases, transcripts and summaries. These materials are often automatically stored, searchable and widely accessible. If they are not privileged or do not reveal privileged advice, they may be disclosable. Businesses should therefore treat AI-generated artefacts as part of the record and manage them accordingly. This includes limiting retention, controlling access and avoiding the creation of unnecessary material.

Several recurring scenarios illustrate how privilege issues can arise in practice:

• Where employees use an AI tool to ask legal questions and receive answers without any lawyer involvement, the starting position is that privilege will not apply. There is no lawyer-client communication, and the exchange is unlikely to meet the dominant purpose test. More defensible models involve AI acting as a delivery channel for lawyer-approved content or as part of a structured intake process that escalates queries to a lawyer. Even then, the position remains fact-sensitive.

• If a business team uses AI to prepare a document intended for legal advice but never sends it, privilege will depend on whether the document was created for the dominant purpose of obtaining advice and kept confidential within a legal process. This is a narrow and fact-sensitive argument, and in practice it is safer to route such materials to the legal team.

• Where AI is used to draft a briefing that is then sent to a lawyer, the analysis is more straightforward. Privilege is more likely to apply, provided the document was created for the purpose of obtaining legal advice and confidentiality is maintained. However, mixed-purpose documents combining legal and commercial content can weaken the claim.

• Privilege cannot be created retrospectively. A document produced for commercial purposes does not become privileged simply because it is later shared with a lawyer. What can be protected is the subsequent lawyer-client communication around it, provided that this is clearly framed and kept confidential.

Meeting notes and AI transcription

AI transcription tools introduce particular challenges. Verbatim transcripts capture everything – legal and non-legal – and may be automatically shared or indexed. Meeting records are not privileged simply because a lawyer attended. Privilege applies only to the extent that the record reflects legal advice or the process of obtaining it. In mixed meetings, this may mean that only parts of a transcript are privileged. A practical approach is to avoid full transcription for mixed meetings, use lawyer-led notes for legal segments, and tightly control circulation and storage.

Conclusion

AI does not alter the underlying rules of privilege, but it does make them harder to apply in practice. The ease with which AI generates, stores and shares information increases the risk that confidentiality will be lost or that legal and commercial work will become blurred. The most effective response is not to rely on technical solutions alone, but to adopt clear operational disciplines: separating business and legal work, using controlled tools and processes, and involving lawyers at the right point.

In short, privilege is not a product feature. It is something that must be actively preserved.