One Meeting, Five Records: Decide Which One Wins

The project meeting ends with apparent clarity. The AI notes say the customer pilot starts Thursday. The task tracker says Friday. A follow-up message says, "after QA signs off," while the CRM still carries the date from last month.

Nobody invented anything. Each artifact reflects a different moment in the same conversation. Yet by the next morning, four people can point to four records and reasonably claim that theirs is current.

AI has made it easier to capture and distribute work from a meeting. It has not automatically answered a more basic question: when the records disagree, which one wins?

The answer is not one enormous document that contains everything. It is a small source-of-truth rule that gives each artifact one job.

A tidy folder can still contain conflicting truth

Meeting systems are becoming much better at producing and organizing artifacts. In August 2026, Google announced that Meet could place notes, transcripts, and recordings into a dedicated folder for each meeting. The update turns scattered files into a more persistent, searchable collection. Google Workspace

A few weeks later, Google introduced deeper agentic capabilities across Gmail, Drive, Docs, Sheets, Slides, and Chat. The company described Gemini as an orchestrator that can gather selected context and create work across apps. Google Workspace

Both developments are useful. Together, they expose a governance problem that retrieval alone cannot solve. A transcript, a generated recap, a project task, and a customer record may all be easy to find. That does not make them equally authoritative.

Search answers, "Where did we mention the pilot date?" A source-of-truth rule answers, "Which date is the team expected to act on?"

Give every record one job

Treat the outputs from a meeting as a record stack. Each layer serves a different purpose, and no layer should quietly inherit authority from another.

The recording and transcript preserve evidence. They help people inspect what was said, recover nuance, and correct a summary. They are not practical operating manuals. Spoken discussion contains abandoned options, jokes, false starts, and sentences that only make sense with the next three minutes attached.

The recap provides orientation. It should help an attendee remember the shape of the conversation and help an absent colleague catch up. A recap is allowed to compress. That is precisely why it should point to authoritative records rather than becoming one by accident.

The decision log governs choices. It records the selected option, who had authority to confirm it, relevant conditions, and what would reopen the decision. If the recap and decision log differ, the decision log wins after a human corrects the conflict.

The task system governs execution. It owns the assignee, deliverable, due date, and status. A bullet in meeting notes can explain why a task exists, but it should not become a second task that drifts away from the first.

The relevant business system governs external commitments. A CRM may own a customer promise. A change log may own a release decision. A hiring system may own an interview outcome. Put the commitment where the people responsible for delivering or reviewing it already work.

This stack is not a recommendation to buy five tools. A small team may keep decisions and actions in one project system. The important part is that the destination is named before automation begins to copy polished fragments everywhere.

Write the rule in four lines

For each recurring meeting, add a short publication rule to the agenda or working agreement:

Decisions live in the project decision log.
Actions live in Linear.
Customer commitments live in the CRM.
The transcript resolves wording disputes; the decision owner resolves meaning.

Change the systems to fit the team. Keep the grammar.

The fourth line matters most. A transcript can show that someone said, "Thursday should work." It cannot decide whether that sentence was a commitment, a forecast, or a polite response to a proposal. Evidence can narrow a disagreement, but authority still belongs to the people and process that made the choice.

This rule also gives an AI agent a safer boundary. Instead of asking it to "send the follow-up everywhere," you can give it explicit routes: add the approved decision to one log, create accepted tasks in one tracker, and link both from the recap. When a required destination is unclear, the agent should stop and ask rather than inventing a new home.

End with a three-minute publication pass

The best time to prevent record conflict is while the people who created it are still present.

In the final three minutes, show the live recap and review only the items that will travel into another system. For each one, ask:

1. Is this a decision, an action, an external commitment, or context only?
2. What is its authoritative destination?
3. Who will confirm that the published record is correct?

Then publish once. The recap can link to the destination, but it should not maintain a separate editable version of the same commitment.

Imagine the team agrees to start the pilot Thursday if QA approves by Wednesday at noon. The decision log receives the conditional choice and decision owner. The task tracker receives the QA check, owner, and deadline. The CRM receives Thursday only after the condition is satisfied. The recap links to each record and explains the sequence.

That is slightly more deliberate than turning every sentence into an instant update. It is also much easier to trust.

Corrections should move outward from the source

Conflicts will still happen. Someone may hear the wrong date, an agent may miss a condition, or the team may change its mind in a later conversation.

When that occurs, correct the authoritative record first. Then update or mark the derivative artifacts. Do not silently edit a recap while leaving the task and customer record untouched.

A lightweight correction receipt can read:

Pilot start corrected from Thursday to Friday in the decision log. QA sign-off moved to Thursday noon. The task and CRM record now match. Supersedes the October 2 meeting recap.

This preserves a useful history without forcing future readers to compare every version themselves. It also makes automation failures visible. If the same kind of conflict keeps appearing, the team can repair the route rather than repeatedly repairing the output.

For high-consequence meetings, add one further guardrail: downstream automation pauses when two authoritative destinations disagree. A customer email should not be sent merely because the recap contains a date while the CRM contains another.

Better memory needs clear authority

A searchable meeting history is valuable because work rarely fits inside the calendar block. People return later to ask why a choice was made, what changed, and what they owe next.

But memory becomes less useful when every artifact presents itself as the final word. The goal is not to erase the transcript, shrink the recap, or prevent helpful integrations. It is to preserve the difference between evidence, explanation, decision, and execution.

Caspi supports that broader meeting loop with live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. Those capabilities can help a team catch conflicts during the conversation and carry context forward afterward. The team still needs to name where authority lives.

One meeting may create five records. That is manageable when each one has a job, every commitment has a home, and nobody has to guess which version became real.