A week of meetings can leave you with five documents, three half-finished action lists, and one uncomfortable question on Friday afternoon: what did we actually decide?

AI is useful here when you make it work from approved notes and transcripts, force it to separate facts from inference, and keep every owner, date, decision, and blocker traceable to a source. The goal is not a prettier meeting summary. It is a small handoff pack another person can inspect and continue from.

In about 45 minutes, you can create four practical assets: a decision ledger, an action tracker, a risk/dependency map, and a concise manager or teammate update.

Capability check — August 11, 2026: OpenAI says ChatGPT Projects are available across free and paid subscriptions and can keep related chats, files, and project instructions together. OpenAI's current Free Tier FAQ also lists file uploads and data analysis, subject to account limits. Google says NotebookLM can answer from uploaded sources with inline citations; standard access currently allows up to 50 sources per notebook. Exact limits, workplace controls, and availability can change.

Before you start: use only material you are allowed to process

Meeting notes can contain customer information, employee performance details, roadmap plans, security issues, pricing, legal discussions, or other confidential material.

Use a company-approved AI account when your employer requires one. Remove information that is not needed for the task. Do not upload private recordings, customer data, HR material, credentials, or confidential documents merely because they might improve the summary.

For a low-data route, paste only the relevant, redacted notes instead of connecting an entire drive, inbox, or workspace.

1. Build a decision ledger that refuses to guess

Outcome: One table showing what was actually decided, what is still open, and where each item came from.

Best fit: Employees who attend several project, product, client, planning, or cross-team meetings in one week.

Inputs and tools: Approved meeting notes or transcripts from the week. A ChatGPT Project is useful for keeping files and instructions together; NotebookLM is useful when you want source-grounded answers with inline citations. Pasted text works too.

Use a simple schema:

TopicDecisionStatusSourceConfidence
Release dateMove to Sept. 8ConfirmedTuesday launch meetingHigh
Vendor choiceNot decidedOpenThursday procurement notesHigh

The most important field is Status. Confirmed, Proposed, Open, and Conflict should not be blended together.

Steps

  1. Add only the notes relevant to the work being handed over.
  2. Ask for one row per decision or unresolved decision.
  3. Require a source meeting, date, or note location for every row.
  4. Use NOT STATED when the source does not contain a decision.
  5. Open the original notes and manually verify every row marked Confirmed.

Useful prompt:

Use only the material I provide. Build a decision ledger with: topic, decision, status (CONFIRMED / PROPOSED / OPEN / CONFLICT), source meeting/date, exact supporting note, and anything that still needs verification. Never convert a suggestion into a decision. If no decision was recorded, write NOT STATED.

Time and cost: About 10 minutes. A free-tier text or file workflow is practical for a small weekly pack, subject to current usage limits.

Privacy, accuracy, and copyright limits: A transcript can be wrong, incomplete, or produced without permission. Check your organisation's recording and data rules. AI can also compress nuance too aggressively, so compare every important decision with the original notes before treating it as settled.

2. Create an action tracker without inventing owners or deadlines

Outcome: A clean list of actions that distinguishes assigned work from vague follow-ups.

Best fit: Teams where action items are scattered across meeting notes, chat messages, and personal to-do lists.

Inputs and tools: The verified source pack and decision ledger.

A useful action table looks like this:

ActionOwnerDue dateEvidenceStatus
Send revised mock-upSamAug. 14Wednesday design notesAssigned
Check analytics impactUNASSIGNEDNOT STATEDFriday reviewNeeds assignment

Steps

  1. Extract every explicit commitment, request, and follow-up.
  2. Separate ASSIGNED from NEEDS ASSIGNMENT.
  3. Keep NOT STATED for missing deadlines rather than estimating one.
  4. Add the source sentence or note that supports each action.
  5. Mark items as complete only when a source actually says they were completed.

Useful prompt:

From these verified notes, create an action tracker with: action, owner, due date, status, dependency, and source. Do not infer an owner from who mentioned the task. Do not invent a deadline from phrases such as “soon” or “next.” Use UNASSIGNED and NOT STATED when necessary. Flag any action that appears in two sources with conflicting owners or dates.

Time and cost: About 10 minutes.

Privacy, accuracy, and copyright limits: Names, deadlines, and responsibilities can affect real work. An AI mistake here can create friction between colleagues, so keep the source beside each item and confirm anything consequential before sharing the tracker.

3. Turn blockers into a risk and dependency map

Outcome: A short list of what could stop the work, what it depends on, and which unanswered questions deserve attention first.

Best fit: Projects with several teams, vendors, approvals, data dependencies, or unresolved technical choices.

Inputs and tools: The decision ledger, action tracker, and the same approved notes.

The useful distinction is between explicit risk and inferred risk.

For example:

  • Explicit: “Legal approval may not arrive before Friday.”
  • Inferred: “If legal approval is late, the launch could slip.”

Both can be useful, but they are not the same kind of evidence.

Steps

  1. Ask for blockers, dependencies, unresolved questions, and external approvals.
  2. Label each item EXPLICIT or INFERRED.
  3. Require a source for explicit items and a stated reasoning chain for inferred items.
  4. Rank by immediacy and reversibility, not by dramatic wording.
  5. Manually check the top three before adding them to a status update.

Useful prompt:

Build a risk and dependency map from the source pack. For each item show: risk/dependency, type (EXPLICIT / INFERRED), evidence, affected action or decision, next verification step, and who must answer if the notes explicitly name someone. Do not turn uncertainty into a prediction. Keep unsupported concerns out of the main list.

Time and cost: About 10 minutes.

Privacy, accuracy, and copyright limits: AI can make an ordinary unknown sound like a serious project risk. Treat inferred items as questions to investigate, not as facts. Avoid feeding sensitive commercial, security, legal, or personnel risks into a tool your organisation has not approved.

4. Draft the manager update or handoff from verified facts only

Outcome: A concise update that another person can act on without rereading the whole week.

Best fit: Friday status updates, holiday handoffs, cross-team transitions, manager check-ins, or covering for an absent colleague.

Inputs and tools: Only the verified decision ledger, action tracker, and risk map. Do not draft from the raw meeting pile again; that reintroduces ambiguity you already removed.

A useful structure is:

  • Done: completed work that has evidence.
  • Decided: confirmed decisions.
  • Next: assigned actions and deadlines.
  • Blocked: verified blockers and dependencies.
  • Needs decision: unresolved items that require a person to choose.

Steps

  1. Feed the three verified tables back into the AI.
  2. Ask for a 150–250 word update for the real audience.
  3. Require every sentence to map to a ledger row.
  4. Remove background detail that does not change the reader's next action.
  5. Check names, dates, commitments, and tone before sending or posting it yourself.

Useful prompt:

Draft a concise handoff for [manager / teammate / project lead] using only these verified tables. Structure it as Done, Decided, Next, Blocked, and Needs Decision. Do not add context from general knowledge. Do not soften unresolved issues into completed work. Keep names, owners, dates, and commitments exactly as recorded. At the end, list any sentence you could not trace directly to a source so I can remove it.

Time and cost: About 15 minutes including human review. A normal free-tier text chat is enough once the verified tables are small.

Privacy, accuracy, and copyright limits: Do not let an AI send the message automatically unless your organisation has explicitly approved that workflow. A polished update can still contain a wrong commitment, confidential detail, or recipient-sensitive wording. Human review is the final publication gate.

The 45-minute weekly reset

ActivityTime
Decision ledger10 minutes
Action tracker10 minutes
Risk/dependency map10 minutes
Manager update or handoff15 minutes

The reliable chain is:

raw notes → source-backed decisions → explicit actions → verified risks → concise handoff

That is much safer than asking an assistant, “Summarise my week,” and trusting whatever it chooses to emphasise.

A free or low-cost route

You do not need a dedicated meeting-intelligence subscription for this workflow.

  • ChatGPT Free: OpenAI currently lists file uploads and data analysis for free-tier users, subject to usage limits.
  • ChatGPT Projects: Available across free and paid subscriptions; useful for keeping the week's files, instructions, and chats together.
  • NotebookLM: Standard access currently supports up to 50 sources per notebook and source-grounded chat with citations.
  • Plain text + spreadsheet: Still the lowest-data fallback. Paste redacted notes, then keep the final ledgers in a local sheet or document.

The useful thing AI saves is not the responsibility of deciding what happened. It saves the repetitive work of extracting, normalising, comparing, and formatting evidence so that you can review it faster.

Final five-minute check before sharing

Before sending the handoff, verify:

  1. every decision marked confirmed really appears in the notes;
  2. every owner was explicitly assigned;
  3. every deadline was actually stated;
  4. every blocker is labelled as fact or inference;
  5. no confidential or unnecessary personal information leaked into the final update.

If one of those fails, fix the source table first rather than polishing the final prose.

Sources

Checked August 11, 2026:

AIProductivityFutureOfWorkMeetingNotesWorkflows
Written and reviewed by /lico

Just writing down my thoughts, interests, and the things I learn along the way.