When the Meeting Minutes and the AI Transcript Don’t Match: Litigation Insights From Delaware Chancery

You join a virtual meeting and a banner appears: The meeting is being transcribed by AI. You scan the attendees, glance at the agenda, and run a quick risk-benefit analysis. Who turned it on? Should I say something? Will someone else? Has everyone consented? Will the transcript be accurate and secure? And is there a risk if the AI transcript and the meeting minutes differ?

On August 28, 2026, the Delaware Court of Chancery shed light on risks that arise from differences between the official minutes and AI-generated transcripts of the same board meetings. Its post-trial opinion illustrates what may happen when AI transcription is running during a sensitive meeting, effectively generating an alternate record of the proceedings. The official minutes may emphasize the rationale for a decision in one way, and an AI-generated transcript may provide additional context, suggest a different emphasis, or just be wrong.

This decision follows a line of cases that began earlier this year with decisions that addressed the privilege ramifications of AI use, followed by opinions in which a party’s ChatGPT logs became key evidence, first in a Delaware earnout dispute, then in a federal challenge to government grant terminations. Sidley has been advising clients that AI transcription likely would be the next area of focus, and this decision confirms that it is.

Organizations that use AI tools to transcribe, summarize, or interpret meetings should consider what records those tools create, where those records are maintained, and how they may be used if a dispute later arises. Those considerations are relevant not only to board meetings, but also more broadly to management and other business meetings for which an AI-generated transcript may become the principal written record of what participants said. For organizations with employees who speak different languages, live AI interpretation tools carry additional risk because many of them transcribe what is said before translating it.

Key Considerations

For organizations using AI to transcribe, summarize, or interpret meetings, several questions are worth considering in light of this decision, which we explore in more depth below:

• When should AI transcription be permitted?
• Is there a written policy in place governing AI transcription?
• What review process, if any, is in place for AI transcriptions?
• Are AI transcripts kept as corporate records, and if so, for how long?
• Do directors, executives, and employees know the risks of AI transcription?
• What do AI vendors’ terms allow, who has consented, and where does the data go?

Case Overview

In ATG Capital Opportunities Fund LP v. Lane,¹ an investor tried to replace the entire board of a public company by nominating its own slate of directors. The board declined to accept the nomination, citing information it believed the investor was required to disclose but had not. The investor sued, and the case went to trial.

The court had two written records of what happened in the boardroom: the official minutes and AI-generated transcripts of the same meetings. The minutes described the board’s decisions in the way minutes typically do, recording the action taken and the stated reasons for it. The AI transcripts purported to capture the discussion itself. At one meeting, for example, the minutes recorded that a defensive measure was adopted to protect stockholders; the AI-generated transcript recorded the chairman describing it as “necessary in order for the board to remain in its position.” Notwithstanding common warnings about inaccuracies in AI-generated material, nothing in the opinion suggests that the admissibility or reliability of the transcripts was contested; they were joint trial exhibits, cited alongside testimony without qualification.

After reviewing the record, the court ruled for the investor on its challenge to the board’s rejection of the nomination notice. In the court’s view, the board’s concerns were matters for stockholders to weigh in the election rather than grounds to exclude the nominees from the ballot. Although the AI transcripts did not decide the case, they informed the court’s account of the board’s deliberations, and the court cited them in the portion of the opinion assessing the board’s motivations.

Practical Takeaways

1. Decide deliberately when and how AI transcription can be used, and treat board and other sensitive meetings with particular caution.

Set the ground rules for use of AI transcription. Organizations should decide intentionally whether and when transcription tools may be used, what controls apply, how to provide notice or obtain consent, and whether and how to maintain the output. Where an organization draws the line will depend on its risk tolerance, its regulatory exposure, and the purpose of the meeting. What is acceptable for one organization may not be for another, and the same is true from meeting to meeting within a single organization. The purpose of the meeting, whether privilege may attach, and the seniority of those involved all bear on that judgment.

Board and committee meetings carry higher risk. Minutes of board and board committee meetings are not supposed to be transcripts. Boards and their committees act only through majority agreement, and it is unlikely that every comment or view expressed in a meeting was agreed upon or influenced the final decisions in the same degree. As evidence of corporate action, minutes are among the first documents requested in litigation. Well-drafted minutes provide important evidence of the board’s diligent attention to important matters. Minutes showing that a report was presented, a discussion followed, and a decision was reached support the presumption under the business judgment rule that the board acted on an informed basis and in good faith. The role of the corporate secretary is to create a record of what was decided, and all members of the board or board committee are given the opportunity to review and approve the meeting minutes before they become the permanent record of what occurred.

A verbatim record of every hypothetical or devil’s advocate question may be read out of context (and, to the extent deliberative comments are meant to be provocative rather than setting out a firmly held viewpoint, inaccurate) when it surfaces years later. And recording every remark could chill candid debate. AI transcription tools create a record that may be taken out of context without emphasis on what all members of the board or committee agree should in fact be the record of the meeting.

Privileged and litigation-related discussions call for similar care. Where counsel is present and providing legal advice, the case for disabling transcription is particularly strong. A purportedly verbatim, vendor-generated record of a privileged discussion could invite disputes over waiver, complicate privilege review and logging, and, unlike counsel’s own notes or draft minutes, may be less likely to be treated as work product. And, of course, any AI tool used in this way should be an enterprise tool (i.e., contained within the organization and not used to train an external model) rather than a consumer or other commercial tool that lacks comparable privacy, cybersecurity, and other protections.

Management and executive meetings are not exempt. Although this case is about a board meeting, most AI transcripts are not generated in the boardroom. They are generated in management calls, a pre-meeting discussion, and other meetings that are rarely minuted, so the transcript can become the only written record of that particular event. Those records are generally discoverable in ordinary commercial litigation, employment disputes, and regulatory investigations, and they capture the informal, exploratory way some people actually talk.

More senior participants and more consequential decisions generally mean the record is more likely to be scrutinized in the future. That doesn’t mean more ordinary meetings are necessarily safe; they are equally discoverable. But the risk calculus may be different, and a policy that treats every meeting the same ignores the differing risk profiles and sensitivities of certain meetings.

Global operations widen the exposure. For global organizations with multinational teams, there is an added layer. The live AI interpretation feature, which may rely on AI transcription, is often used to facilitate team meetings across languages. This may make it even harder to get people to agree to turn the tool off, regardless of countervailing risks. But it also means the resulting record could include a machine translation that may not reflect what was actually said. Multi-jurisdictional organizations also have more complex privacy concerns.

2. Articulate a risk-based policy for use of AI transcription.

An organization should consider implementing a policy on how AI transcription may be used throughout the organization, but this step is particularly important with respect to board and committee meetings, executive sessions, or any discussions of sensitive or privileged matters, including matters involving litigation, contests for control, transactions, or personnel.

Where the organization permits transcription of board, committee, or other sensitive meetings, that decision should be made with input from the corporate secretary and counsel and reflected in a written policy that addresses, at a minimum:

• which meetings and sessions, if any, may be transcribed, and which may not;
• whether and under what circumstance individual directors may run personal AI notetakers, which create records outside the company’s control, security, and retention framework and present an additional layer of risk; and
• who is authorized to enable or disable transcription during a meeting, and how participants are informed.

3. Decide whether AI-generated transcripts will be kept as corporate records, and consider the risks of doing so.

Delaware plaintiffs long have looked to obtain emails and texts outside of formal board materials as potential evidence of director motive. In 2025, the Delaware legislature amended Section 220 of Delaware’s General Corporation Law to limit stockholders’ presumptive inspection rights to formal corporate records, including minutes and “records of any action of the board of directors.”² AI transcription has the potential to blur the distinction between formal and informal records. Whether a transcript is a “record” of board action under Section 220 is a question no court yet has addressed. If it is, stockholders potentially could obtain transcripts through a books and records demand, depending on other considerations. In ATG Capital, the transcripts reached the court through ordinary discovery in the litigation; Section 220 could shorten that path considerably. For board and committee meetings in particular, the better approach in context may be not to retain AI transcriptions (assuming there are no applicable litigation holds) after minutes are created and approved by the board or board committee as reflecting the record of the meeting. Companies that allow AI transcription for board, committee, or other sensitive meetings should strongly consider:

• adopting a retention policy that specifies that transcripts are not kept as records (again, absent other circumstances, including litigation holds) and are only used to assist in the preparation of minutes, after which they are promptly deleted, and applying it consistently, because once litigation is reasonably anticipated, transcripts will be subject to hold obligations like any other electronically stored information, and that moment can arrive earlier than expected; the policy should also address, for meetings that are not minuted, whether AI-generated summaries are treated as the working record and, if so, for how long they are retained;
• assessing where transcripts are stored even for a short period of time (including on vendor systems), who can access them, and whether they are being automatically distributed to attendees or others and if so, how to mitigate risks associated with such distribution; and
• considering an immediate review process (close in time to the transcript generation) to improve accuracy and reliability.

4. Train employees (especially directors and executives) to speak knowing a transcript may exist, and draft meeting minutes accordingly.

Speak knowing the transcript may be read verbatim. Employees, and especially directors and executives, should understand that AI transcripts, like other electronically stored information, may need to be preserved if relevant to a dispute, and that a transcript of a meeting they attended could be produced and read back to them. They should assume that remarks in a transcribed meeting could be read verbatim, without tone or context, by a judge or jury, and that hypotheticals, devil’s advocacy, and shorthand may not read as intended. Organizations may consider addressing these issues in training for directors, executives, and other employees who regularly participate in transcribed meetings, using real examples like this case.

Draft minutes knowing a transcript may exist. Minutes are intended to summarize actions taken and the considerations behind them, rather than to record every remark along the way. The summary nature of minutes is a legitimate and longstanding practice. The practice has parliamentary roots – Robert’s Rules of Order long has instructed that minutes should focus on what was done at a meeting, not what was said, because a deliberative body speaks through its votes rather than the remarks of individual members.

Board minutes are where the gap between minutes and transcripts may become more stark, and this case is a powerful reminder of what can happen.

When a verbatim transcript exists, however, the minutes may be read against it, and any material gap may be characterized by an adversary as a discrepancy, even where the transcript itself may be misleading in emphasis or even inaccurate. Minutes should be accurate and deliberate, and approved by the board or board committee members who participated.

5. Extend AI governance to vendors, consent, individual users, and cross-border use.

• Vendor terms. As with other generative AI tools, companies should review the terms governing meeting transcription vendors, including data retention, use of recordings and transcripts for model training, access controls, security, and the company’s ability to delete data. Enterprise versions of these tools typically offer negotiated terms, administrative controls, and contractual commitments regarding the use of customer data; consumer or free versions often do not, and employees may be using them without the organization’s knowledge.
• Consent. Recording and transcribing meetings may implicate federal and state wiretap and recording-consent laws, including laws requiring all parties’ consent in many jurisdictions. A federal court in California recently allowed such claims to proceed against an AI notetaker vendor alleged to have recorded participants without their consent and retained the recordings for its own use.³ Meetings with outside advisors, counterparties, and other third parties warrant particular attention.
• Individual use. Directors, executives, and other employees should be reminded that prompts entered into consumer AI tools, including in the heat of a transaction or contest, are likely discoverable, generally unprivileged, and potentially probative of intent, as the exchanges at issue in Fortis Advisors v. Krafton and United States v. Heppner illustrate.⁴
• Cross-border use. Consent and recording rules differ by country, and transcripts and recordings may cross borders and trigger data transfer and privacy obligations. As noted above, AI translation can produce an inaccurate record that is nonetheless discoverable. The policy discussed above should extend to multinational teams, not only the board.

Looking Ahead

Unauthorized notetakers raise questions that courts have not yet had to answer. Individual participants could run personal notetakers without authorization, creating records outside the company’s systems and retention framework. Whose record that is, who controls it, and what preservation and production obligations attach to it are questions no court has yet had to decide. We expect that will change.

The evidentiary questions have not yet been tested. Courts have yet to work through how AI-generated content should be evaluated under authentication, hearsay, and reliability standards.⁵ Nothing in ATG Capital suggests those questions were raised; the transcripts were joint exhibits. But they will be raised eventually. AI transcription can be imperfect: Comments can be misattributed to the wrong speaker, negations dropped, and tone or hypothetical framing lost, especially during cross-talk. A transcript that is largely accurate may still misstate the one remark that matters or obscure the relative emphasis participants placed on different points, and the error may not be apparent to a reader who was not in the room. Companies that retain transcripts should consider how their accuracy will be assessed and whether metadata or source recordings should be kept for later authentication. In litigation, parties should decide deliberately whether to stipulate to AI-generated records as joint exhibits or to preserve objections to authenticity and reliability.

Conclusion

The ATG Capital case highlights that AI-generated transcripts may be treated as part of the record in assessing the motivations behind corporate decisions, with attendant risks where the AI transcript and the official meeting minutes differ.

As transcription and similar functionality become more common in software used for board, management, and other internal meetings, organizations may want to consider in advance when those tools should be used, what records they create, how those records are retained, and what controls apply to them. Addressing those questions through governance, retention, and training policies before a dispute arises may reduce both litigation and operational risk.