The AI deal memo and the work product privilege question
When a partner asks an AI tool to draft a deal memo, the lawyer's mental impressions and the AI's output are mixed in the same document. Who is the author for privilege purposes is a real question.
A partner is preparing a deal memo for an audit committee on a proposed acquisition. The partner has views on the regulatory risk, the integration challenges, the valuation framework. The partner also has Claude open in another window, and uses it to draft the structural sections of the memo: deal overview, structure analysis, market context. The partner edits the AI's draft, adds the partner's own analytical sections, and finalizes the memo.
The memo is later requested in discovery. The producing party asserts work product. The requesting party argues that the AI generated sections are not the lawyer's mental impressions and therefore not entitled to opinion work product protection. Both arguments have legitimate roots in the doctrine. Neither has a settled answer.
This is the new privilege question for corporate practice and most lawyers are not yet ready for it.
The doctrine in brief
Federal Rule of Civil Procedure 26(b)(3) protects documents prepared in anticipation of litigation by a party's representative. The protection has two tiers. Ordinary work product (factual material, undirected research) is discoverable on a showing of substantial need. Opinion work product (the lawyer's mental impressions, conclusions, opinions, legal theories) is nearly absolutely protected.
The distinction matters because most deal memos contain both. A factual summary of the target's business is ordinary work product, possibly discoverable. The partner's analysis of regulatory risk and recommendation on deal structure is opinion work product, almost never produced.
Where AI complicates this
When the partner uses AI to draft a factual summary, the AI's output is a finished work product. The partner edits it, but the editing might be light. The factual section is then largely the AI's product with the partner's review.
The opinion work product analysis traditionally rests on the proposition that the lawyer's mental impressions are reflected in the document. If the document was substantially generated by an AI model and lightly edited, is the lawyer's mental impression reflected in it? The answer depends on the depth of editing, the specificity of the prompt, the iterative refinement, and the lawyer's actual creative contribution.
A prompt of "draft a regulatory risk analysis for this acquisition" produces an AI output that is the AI's work, not the lawyer's. A more directive prompt that articulates the lawyer's theory and asks the AI to draft against that theory produces something closer to the lawyer's work. The line between those two is not clearly drawn.
What courts have said so far
Very little, as of 2026. The few cases that have touched on AI generated content in discovery have mostly been about authentication and reliability, not about privilege. The doctrinal framework for handling AI assisted documents in privilege analysis is being developed in real time, mostly through bar association ethics opinions and through individual judges in specific cases.
The pattern that is emerging is that courts are willing to extend work product protection to AI assisted documents if the lawyer's role in directing, editing, and finalizing the document is meaningful. They are skeptical when the document looks like raw AI output that the lawyer has rubber stamped.
What corporate counsel should do now
Treat the prompt as part of the privilege analysis. The prompt is often the clearest articulation of the lawyer's mental impressions. If the prompt is preserved, the work product argument is stronger because the lawyer's direction is documented. If the prompt is not preserved, the argument that the lawyer directed the AI's output relies on testimony alone.
Edit substantively, not cosmetically. A deal memo that has been substantively edited by the lawyer, with the lawyer's analysis woven into the structural sections the AI drafted, is closer to a traditional work product document than a memo that has been lightly cosmetic edited. The substantive editing is what makes the document the lawyer's.
Document the iteration. If a memo went through multiple AI assisted drafts with the lawyer's refinement at each step, having a record of that iteration supports the work product claim. The record should be preserved in a way that does not itself become a problem under discovery.
Train associates on the distinction. The associate who treats the AI as a drafting partner and the AI as a research assistant produces different documents with different privilege exposure. The training is not technical. It is doctrinal.
A tool that addresses one slice
I built a small tool that helps think through this. /tools/deal-memo-classifier/ takes a paragraph or section and asks structured questions about how it was produced: was it AI generated, was the prompt directive, was it substantively edited, does it articulate a theory or merely report facts. The output is a rough classification of the paragraph's likely work product status. It is educational, not legal advice, and saves nothing.
The broader point is that the privilege framework that the corporate bar relies on every day was built for a world where the lawyer wrote the document. The introduction of AI as a drafting partner changes the underlying assumptions of that framework in ways that have not yet been fully worked through. The lawyers who think about this before their first discovery dispute will be in better shape than the ones who do not.
Walter Allison is a corporate attorney in Denver. He writes here about M&A, private equity, and venture capital structure.
Follow on LinkedIn ·
Firm bio ·
More posts