The Best AI Meeting Note May Be the One That Expires

On Monday, a project team talks through a difficult choice. Someone mentions a health problem while explaining a delay. Another person tests an unpopular idea, then withdraws it. By Friday, the meeting has become a recording, a speaker-labelled transcript, an AI summary, a list of tasks and perhaps an index that lets other software retrieve fragments later. The conversation ended. Its copies did not.

AI meeting tools are usually sold as a cure for forgetting. Their more important effect is to change what kind of memory an organisation creates. A transient exchange becomes a collection of durable, searchable artefacts with different audiences and different risks. The useful question is therefore not simply whether to record a meeting. It is: which layer deserves to survive, for how long, and who gets to decide?

A practical answer is an expiry ladder. Keep the smallest record that can still support a legitimate purpose: decisions and commitments for as long as the work requires; a checked summary for less time if its usefulness fades; the raw transcript and audio for the shortest justified period. Expiry should be designed alongside capture, not added after an archive has grown.

One meeting becomes several records

The red recording symbol suggests a single file. In practice, an AI-assisted meeting can produce a chain. Audio or video is transcribed. The transcript is divided into speaker turns. A model selects and rewrites material into a summary. Action items may be copied into project tools. Search or retrieval systems may create additional indexes so that later queries can find relevant passages.

Each transformation changes the exposure. Raw audio contains voice and background context. A transcript makes the words searchable and easy to copy. A summary is smaller, but it embeds selection: what the system treated as a decision, objection or promise. An action item travels furthest because it can trigger work without carrying the discussion that qualified it. Deleting the recording does not necessarily remove these descendants.

This is why “the meeting is stored for 30 days” is an incomplete policy. Which meeting artefact? In which system? Does a summary remain in email after the transcript expires? Was an action copied into another database? Can an administrator retrieve a transcript that participants cannot see? A retention rule attached only to the original file can leave the operational memory untouched.

Current products already expose some of these choices. Microsoft says Teams recordings and transcripts expire after 120 days by default, although administrators and users can alter parts of that lifecycle. Zoom lets administrators configure automatic deletion of AI meeting summaries and separately control whether hosts may retain and access the transcript used to create them. Those settings are useful, but they do not decide an organisation’s purpose. A default is a product configuration, not a considered retention period.

More memory can weaken accountability

The strongest argument for keeping everything is serious. A recording can settle disputes, help an absent colleague, support accessibility and reveal that a polished summary omitted a warning. In regulated or contested work, deletion can also destroy evidence that ought to have been preserved. A casual “delete by default” rule can protect the powerful person who controls the minutes rather than the people whose words were misrepresented.

But indiscriminate retention is not the opposite of institutional amnesia. It can create a different failure: nobody knows which record is authoritative. A task list says “launch on Thursday”; the transcript shows that Thursday depended on a safety test; the AI summary omits the condition; the recording is too cumbersome to review. Keeping all four does not resolve responsibility. It distributes it across artefacts.

Long retention also changes the meeting itself. People who know that exploratory speech will remain searchable may avoid asking naïve questions, naming a risk or trying an unfinished thought. This effect will vary by culture, hierarchy and subject, so it should not be asserted as a universal behavioural law. It is nevertheless a design question worth testing: a memory system can improve recall while making the conversation it records less candid.

The NIST Privacy Framework treats retention as part of privacy-risk governance rather than a storage detail. Its Core calls for policies on data processing—including uses and retention periods—to be informed by privacy risk. The UK Information Commissioner’s Office offers a concrete institutional example: its public notice says its Teams recordings and transcripts are automatically deleted after 21 days, while Copilot prompts and responses are deleted after seven. That is not a universal prescription. It demonstrates that different derivatives of the same work can have different lifetimes.

Build an expiry ladder, not a single timer

An expiry ladder begins with purposes, not file types. For each meeting class, a team can name the minimum durable record and the event that ends its usefulness.

• Decision record: the decision, owner, date, constraints and route for challenge. Keep it while the decision governs work, then archive or review it under the relevant records policy.
• Checked summary: a concise account circulated to participants for correction. Keep it while coordination needs it; do not assume it deserves the same lifetime as the decision.
• Transcript: a temporary verification layer for correcting names, conditions and attributions. Expire it after the checking window unless a specific obligation requires preservation.
• Audio or video: the richest and most intrusive layer. Keep it only when replay has a stated purpose that a transcript cannot meet.
• Derived copies: email, chat posts, action trackers, exports and retrieval indexes. Name them explicitly, because an upstream deletion is not a reliable downstream deletion plan.

The ladder needs exceptions. A legal hold, safeguarding concern, formal proceeding or documented accessibility need may require a different path. Exceptions should identify who authorised them, which artefacts are covered and when the decision will be reviewed. “Keep just in case” is not an exception rule; it is the absence of one.

Participants also need a usable signal. At the start of a meeting, the organiser should be able to state what is being captured, what durable record will remain and when temporary layers will expire. After the meeting, people should know how to correct the summary and whom to contact if the capture itself was inappropriate. Agency is not achieved by a long privacy notice that nobody can act on.

Run an expiry rehearsal

A team can test this approach without redesigning its information estate. Choose three recurring, low-risk meeting types: for example, a weekly planning call, a design critique and a routine supplier check-in. Do not include legal, clinical, disciplinary or otherwise sensitive meetings in the first experiment.

For four weeks, map every artefact each meeting produces and where it travels. Set a short transcript-checking window, such as seven days, while retaining the verified decision record under the team’s existing policy. Before temporary material expires, ask participants whether the durable record is sufficient to explain what was decided, what remains conditional and how to reopen the issue. Record corrections, time spent checking, missing context and copies that could not be deleted as expected.

This resembles the change-preview principle described in Alkemata’s experiment for AI agents: consequential automation should expose its proposed effects before they become hard to reverse. Here, the preview is a retention map. It shows what the meeting assistant will create and what deletion will actually reach.

The experiment may show that a transcript needs 14 days rather than seven, that a particular meeting should never be recorded, or that the summary is too unreliable to become the official record. Those are useful results. The purpose is not to minimise data at any cost. It is to match each layer of memory to a clear responsibility.

Before the next meeting assistant is switched on by default, draw the ladder on one page. If the team cannot say which artefact is authoritative, who can correct it and when each derivative disappears, the system is not merely taking notes. It is setting records policy by accident.