Back to Articles
    Governance & Boards

    Should AI Draft Your Board Minutes?

    Board secretaries have been quietly using meeting AI for a while now, and the output is often better formatted than what they used to produce by hand. Here is the difficulty: minutes are not a record of a meeting, they are a legal record of decisions, and the qualities that make AI summaries impressive are precisely the qualities that make bad minutes. The tool is genuinely useful. It just needs to be pointed at a different job than the one it defaults to.

    Published: August 24, 202614 min readGovernance & Boards
    Board secretary reviewing an AI-generated draft of nonprofit board minutes

    The board secretary role is one of the least enviable jobs in nonprofit governance. You are expected to follow a substantive discussion closely enough to participate, while simultaneously capturing it accurately enough to produce a document that could end up in front of the IRS, an insurer, opposing counsel, or a future board trying to reconstruct why a decision was made. Most people do one of those two things well and the other poorly.

    So the appeal of meeting AI is obvious, and adoption in boardrooms has been quiet and fast. A tool that listens, transcribes, and produces a structured summary appears to solve the problem exactly. The secretary can participate fully, the draft arrives before anyone has left the parking lot, and the formatting is consistent in a way that manual minutes across three different secretaries never were.

    The reason this deserves more thought than it usually gets is that minutes occupy an unusual position among organizational documents. They are the official evidence that your board exercised its duties, and they are frequently the only contemporaneous record of a decision. The IRS expects public charities to keep them permanently. They matter in disputes about director conduct, in insurance claims, in audits, and in any situation where someone needs to establish that a board considered something before acting.

    What follows is not an argument against using AI in the boardroom. It is an argument about which part of the work to give it. We cover what minutes are legally for and why summarization tools push against that purpose, the specific failure patterns that show up in AI-drafted minutes, the separate question of what happens to the recording and transcript, and a workflow that captures most of the time savings without producing a governance record you would rather not explain.

    Minutes Are Evidence of Decisions, Not a Record of a Conversation

    Ask a nonprofit attorney what belongs in board minutes and the answer is narrower than most boards expect. Minutes should establish that the meeting was properly held and that the board acted: the date, time, and place, who attended and who did not, whether a quorum was present, the motions made, who made and seconded them, the vote results, any abstentions and recusals, and the fact that directors left and returned when they had a conflict. Reports received and materials distributed are noted. Actions are recorded precisely.

    What does not belong is nearly everything else. Minutes are not a transcript. They should not attribute opinions to individual directors, should not record who argued which side, should not characterize the tone of a debate, and should not preserve the half-formed idea someone floated before withdrawing it. The consistent advice from counsel who defend nonprofits is that detail about discussion creates risk without creating value, because those passages can be read years later, out of context, by someone whose interests are adverse to the organization.

    That said, minutes should reflect that deliberation occurred, particularly on consequential matters. There is a meaningful difference between recording that the board discussed the executive compensation study, reviewed comparability data, and approved the recommendation, and recording a blow-by-blow of who thought the number was too high. The first demonstrates process, which is exactly what a regulator wants to see. The second is raw material for a plaintiff.

    This is the heart of the problem with using a summarization tool for this job. A meeting AI is built to capture what was said and condense it faithfully. Good minutes require the opposite instinct: aggressive omission of nearly everything said, retention of a narrow set of formal facts, and neutral characterization of the rest. Those are different tasks, and the default output of any meeting tool sits on the wrong side of the line.

    What belongs in minutes

    The formal record of board action

    • Date, time, place, and whether the meeting was regular or special
    • Attendance, absences, and confirmation of quorum
    • Motions recorded exactly, with mover, seconder, and vote result
    • Abstentions, recusals, and directors leaving or rejoining the room
    • Reports received and documents presented, identified by name
    • A neutral note that deliberation occurred on significant matters

    What should stay out

    Detail that creates exposure without value

    • Who said what, and which directors took which position
    • Direct quotes, verbatim exchanges, and characterizations of tone
    • Ideas raised and abandoned, and options considered then dropped
    • Speculation about legal exposure outside privileged counsel advice
    • Personnel and client details beyond what the action requires
    • Editorial framing such as "after a heated discussion"

    The Specific Ways AI-Drafted Minutes Go Wrong

    These failures are consistent enough to anticipate, and most of them are invisible unless someone is reviewing against the standard rather than reviewing for readability.

    Attribution by default. Meeting tools identify speakers because that is what they are built to do, and the resulting draft is full of sentences like "Director Alvarez raised concerns about the reserve policy" and "Director Chen questioned whether the projections were realistic". This is exactly what counsel spend their careers telling boards to keep out of minutes. It is also the hardest pattern to remove after the fact, because the entire draft is organized around it.

    Motion imprecision. The wording of a motion is the one thing in minutes that must be exact, and this is where transcription-based drafting is weakest. Motions are often spoken informally, amended mid-sentence, restated by the chair in different words, and voted on without a clean final reading. A summarizer produces a plausible reconstruction. What the board actually approved and what the minutes say it approved then differ, which is the single most consequential error possible in this document.

    The confidence problem. Where audio was unclear, where several people spoke at once, or where the discussion was genuinely muddled, an AI draft does not leave a gap. It produces a clean sentence. A human secretary in the same position would write a question mark in the margin and follow up. The draft's uniform confidence hides exactly the places that needed checking.

    Failure to note procedural facts. Quorum, recusals, and a director stepping out of the room during a conflicted vote are essential and are frequently not stated aloud in a way a transcript captures. Someone physically leaves; nobody announces it. These omissions matter enormously, because the recusal note is often the whole point of the record on a conflicted transaction.

    Executive session leakage. If the tool keeps running when the board moves into executive session, personnel discussions, legal advice, and compensation deliberations end up in a transcript held by a vendor and potentially in a draft circulated by email. This risk is significant enough that we treat it separately in our discussion of AI notetakers and executive session.

    Helpful invention. Language models fill gaps. A draft may state that the board reviewed a document it did not review, or that a report was accepted when it was merely presented, simply because that is the usual shape of such a paragraph. On a governance record, a fabricated procedural step is a serious problem, and it reads entirely naturally.

    What a reviewer must verify against the source

    Never approve a generated draft on a read-through alone

    • Every motion matches the wording the board actually adopted
    • Vote counts, abstentions, and recusals are correct and complete
    • Attendance and quorum are accurate, including late arrivals and early departures
    • No individual director is named in connection with a position or opinion
    • Nothing describes a procedural step that did not occur
    • Executive session content is absent entirely

    The Transcript Is a Separate Problem From the Minutes

    Even if your minutes end up flawless, using a recording tool creates an artifact that did not exist before: a complete, searchable, verbatim account of everything said in the room. Boards adopt these tools thinking about the minutes and rarely think about the transcript, which is the item most likely to cause trouble later.

    A transcript is discoverable. In litigation, in a regulatory inquiry, or in a subpoena, a verbatim record of board discussion is exactly what an opposing party would most like to have, and its existence is not something you can undo once the meeting is over. Boards that have spent years being careful about what goes into minutes can undermine that discipline entirely by creating a parallel record that is far more detailed and that nobody made a decision to keep.

    There is also a candor cost. Directors who know the room is being transcribed talk differently, and the difference is not in the organization's favor. The value of a board is in its willingness to raise uncomfortable questions, disagree with the executive director, and think out loud about risk. A permanent verbatim record suppresses exactly that behavior, and the suppression is subtle enough that nobody will report it.

    Then there are the practical questions. Where does the transcript live, for how long, and who at the vendor can access it. Whether it is used to train a model. Whether your retention schedule covers it, or whether it simply accumulates in a cloud account nobody owns. Our discussion of retention policy for AI meeting records works through those questions in detail, and a comparison of tool privacy postures appears in our review of Granola, Fireflies, and Otter.

    The defensible position for most boards is that the transcript is a working aid with a short life. It exists to help the secretary produce the minutes, and it is deleted once the minutes are approved. That should be a written policy the board has adopted, not a preference someone holds, because an undocumented practice of deleting records looks very different from a documented retention schedule that was consistently followed.

    Decisions the board should make explicitly

    Adopt these before the tool is used, not after

    • Whether recording is permitted at all, and who authorizes it each meeting
    • That the tool stops before executive session, with a named person responsible
    • How long transcripts are retained and what triggers deletion
    • Who may access recordings and drafts before approval
    • That the approved minutes are the sole official record of the meeting
    • Whether directors are told, every time, that the tool is running

    A Better Division of Labor

    The answer to the title question is a qualified no for the drafting and an unqualified yes for several adjacent tasks. AI should not be producing your minutes from a transcript, because the shape of that output is wrong at a structural level and fixing it takes longer than writing correct minutes would have. But there is real work around minutes that AI handles well, and most of the time savings people are chasing live there.

    Start with a strict template. Minutes for a given board should have identical structure every time, with fields for each required element. The secretary fills the template during or immediately after the meeting from their own notes, which takes far less time than drafting prose because the structure does the remembering. A template is also the single best defense against omission, since a blank field is visible in a way a missing paragraph is not.

    Use the transcript as a lookup tool rather than a source. When the secretary is unsure of the exact wording of a motion or the sequence of votes, searching the transcript for that specific moment is enormously helpful and carries none of the risk of generating the document from it. The secretary is still the author. The recording is a reference they consult, the same way they might have asked a colleague what the chair said.

    Put AI to work on the completeness check instead of the drafting. Give it your approved minutes and ask what required elements are missing, whether any individual director is named in connection with an opinion, whether any motion is described imprecisely, and whether anything reads as characterization rather than fact. This is a review task with clear criteria, it improves the document, and it never touches authorship.

    Automate the follow-through, which is where boards genuinely lose value. Extracting action items with owners and deadlines, drafting the summary that goes to committee chairs, building the agenda for the next meeting from unfinished business, and tracking whether prior decisions were implemented are all tasks with no legal weight and real administrative cost. This is the same territory covered in our guide to preparing board meeting packets with AI.

    Finally, use AI to help the board understand its own history. Minutes going back a decade are a genuinely useful corpus, and searching them for when a policy was last revised, what the board decided about a recurring issue, or whether a conflict was previously disclosed is slow by hand and fast with the right tooling. That use is purely retrospective, touches no official record, and is often what new directors need most. Boards building broader capability here will find our overview of board AI literacy a useful companion.

    Give AI these jobs

    Real time savings, no governance risk

    • Checking approved minutes for missing required elements
    • Flagging attribution, characterization, and imprecise motions
    • Extracting action items, owners, and deadlines for follow-up
    • Building the next agenda from unfinished business
    • Searching years of approved minutes for prior decisions
    • Summarizing board materials before the meeting, not after

    Keep these human

    Authorship of the official record

    • Drafting the minutes themselves, from a template and the secretary's notes
    • Recording motion language, which must be exact rather than reconstructed
    • Deciding what deliberation to note and how to characterize it neutrally
    • Anything touching executive session
    • Judgment about what to leave out, which is most of the skill

    The Case for Small Boards, Where the Tradeoff Is Genuinely Harder

    It would be easy to write this advice from the perspective of a well-resourced organization with a professional secretary and counsel on call. Most nonprofits are not that. In an all-volunteer or very small organization, the alternative to AI-assisted minutes is frequently not careful manual minutes, it is minutes that arrive three weeks late, or that consist of four lines, or that do not get written at all.

    Against that baseline, a tool that produces a structured draft is a genuine improvement, and saying otherwise would be dishonest. Incomplete minutes are their own governance failure, and a board with no reliable record of its decisions is worse off than one with a record that needs editing. The right advice for these organizations is not abstinence, it is a tighter and simpler process.

    The practical compromise is to let the tool produce whatever it produces, then treat that output strictly as raw material for a template rather than as a draft to be edited. The secretary transfers the required facts into the template and discards everything else. This takes fifteen minutes, it is faster than writing from scratch, and it structurally prevents attribution and narrative from surviving into the final document, because nothing gets copied across unless it belongs in a field.

    Small boards should also be more conservative about the recording itself, not less. An organization without counsel is exactly the one that cannot afford a discoverable verbatim transcript, and it is also the least likely to have a retention policy governing what happens to it. Turning the recording off after the draft is produced, every time, is a simple rule that removes most of the exposure. Boards in this position may also find our guide to AI for small nonprofit boards useful for the broader picture.

    Conclusion

    Should AI draft your board minutes? Not from a transcript, and not as the author of record. The document has a narrow legal purpose that runs directly against what summarization tools are designed to do, and the resulting draft fails in ways that are difficult to see and expensive to discover later. The most dangerous version of this is not a bad document. It is a polished one that says the board approved something slightly different from what it actually approved.

    But the frustration driving boards toward these tools is legitimate, and the answer is not to leave the secretary struggling. Give AI the completeness check, the action item extraction, the agenda building, the historical search, and the pre-meeting summarization of board materials. Keep the authorship, the motion language, and the judgment about omission with a person. That split captures nearly all of the time savings and none of the risk.

    And decide the transcript question deliberately, as a board, in writing. Whether a verbatim record of your deliberations exists, and for how long, is a governance decision. It should not be an accident of which meeting tool someone installed.

    Put AI Where It Helps Your Board, Not Where It Exposes It

    We help nonprofit boards adopt AI with the policies and boundaries that keep governance records defensible.