Where we are
Northwind is ten weeks in. M3, raw ingest signed off, has moved from 10 August to 24 August. Six of nine sources are landing, data quality sits at 96.4% against a 99% gate, and the POS extract vendor Veridian has had ticket VN-2291 open for eleven days. Last week's steerco decided two things and half-decided a third. You were in the room, you took notes, and by Thursday nobody will remember which was which. That gap - between what was said and what was recorded - is where programmes quietly lose weeks.
Every meeting produces exactly four things 6 min live
Three of them are worth keeping: the decisions, the actions, and the open questions. The fourth is the discussion - the forty minutes of argument, context-setting, tangents and re-explaining that got the room to those three - and it is worth deleting. Nobody has ever gone back to a transcript to find out what was agreed. They go to the decision log, or they ask you, or they guess.
LiveWhy this is the second thing AI lands on3 min▶
Status reporting is the first place delivery teams point AI. Turning a meeting into "who does what by when" is the second, and for the same reason: it is pure collation, it happens every week, and everyone hates it. It is also the job where a model is genuinely strong - it does not get bored at minute thirty-four, and it does not quietly drop the action that was assigned to you.
- It reads all of it. You stopped taking notes when the vendor conversation started, because you were arguing in it. The transcript did not.
- It has no stake. A model has no reason to soften the action assigned to the loudest person in the room, and it is fast enough to be same-day - which matters more than polish, because an action list that arrives Thursday is a historical document.
The action that was assigned to nobody, twice. On a platform programme, "someone needs to chase the vendor on the extract SLA" appeared in three consecutive sets of minutes, reworded each time, owned by no one. In week four the vendor missed the SLA and the sponsor asked who had been chasing it. Three people said they thought it was someone else, and all three were reading the same minutes. The minutes were accurate. They just recorded a wish.
Self-studyThe one-hour rule2 min read▶
Process the meeting within an hour of it ending, before the next one starts overwriting your memory. Not because the model needs your memory - it does not - but because you are the only person who can tell whether what the transcript says happened is what actually happened. That check gets harder every hour.
Within the hour you can still hear the tone: when Elena V. said "fine, let us do that" it closed the topic, and when Marcus L. said "sure" he had stopped arguing rather than agreed. By tomorrow you have only the words, which is exactly what the model has, so you have lost the thing you were adding. By next week the action list is archaeology and you will write it from what should have happened.
The decision log, and the six fields it needs 6 min live
A decision log is not minutes. Minutes are a record of a meeting; a decision log is a record of the programme's direction, and it should be readable by someone who attended none of the meetings. Six fields, one row per decision, never deleted - superseded decisions stay in and get marked, because "why did we do it that way" is asked far more often than "what did we do".
LiveThe six fields, and the one everybody skips3 min▶
| Field | Why it is there |
|---|---|
| Id - D-14, D-15, sequential | So the status report, the plan and the escalation can all point at the same thing in three characters |
| Date - when it was taken | So you can see what was known at the time. A decision is judged on the day it was made, not on what happened after |
| Who decided - a person, not a forum | "The steerco decided" means nobody decided. A forum cannot be asked to reconsider; Elena V. can |
| What was decided - one sentence | If it takes a paragraph, it is two decisions or it is not a decision |
| Rationale - one line | The reason, not the argument. This is the field that saves you in eight weeks |
| What it cost - what was traded away | The field everybody skips, and the only one that stops the decision being reopened as if it were free |
Northwind's log has two live rows, both taken in the last week of July. Read them as a pair - the first is a scope trade, the second is a cost trade, and both were made by a named human who can be asked about them.
D-14 · 31 Jul · Elena V., COO and sponsor. Descope real-time POS to nightly batch until M5. Rationale: the real-time path depends on Veridian, and R-07 makes that date unsafe. Cost: store operations lose intraday sales visibility for roughly two milestones.
D-15 · 30 Jul · Sofia K., data architect. Keep the legacy warehouse running in parallel until M4 sign-off. Rationale: there is no rollback path otherwise. Cost: two platforms to run and reconcile, and the team carries double support until M4.
Eight weeks later, from zero. A new regional director joined a retail data programme and asked, reasonably, why the POS feed was nightly instead of real-time. Nobody in the room had been in the meeting where it was traded away. The delivery lead spent two weeks re-running an argument that had already been settled, lost it the second time, and the team rebuilt a real-time path they did not need - at the cost of the milestone behind it. One row in a log, with the cost field filled in, would have ended that conversation in ninety seconds.
Self-studyWhere the log lives, and who writes it2 min read▶
- One log per programme, not per meeting. Decisions taken in a corridor, on a call or in a thread go in the same place as steerco decisions. The forum is not the point; the direction change is.
- It goes in your context pack. From b1, alongside the charter, plan, risk register and owner list, refreshed weekly. It is what lets a draft say "as agreed in D-14" instead of proposing work you already cut.
- You write it, the model drafts it. The model proposes rows from the transcript; you confirm the room actually decided, and your name is on the log. And never delete a row - mark it superseded by D-nn, because what you changed your mind about is a programme asset and the first thing a post-mortem asks for.
The most dangerous line in any minutes 5 min live
It is not an error. It is a decision that was never actually taken. The room drifted toward something, nobody objected, the meeting ran out of time, and the notes record agreement. Six weeks later the plan depends on it and the person who would have objected says, truthfully, that they never agreed to that. A model makes this failure worse, not better: asked to extract decisions, it will find decisions, because that is what you asked for.
LiveHow to force the distinction3 min▶
Two moves, one in the room and one in the write-up.
- In the room, say it back. "So the decision is nightly batch until M5, Elena, and you are taking that?" It takes seven seconds and it either confirms the decision or reveals that there was not one. Both outcomes are wins; the second is the bigger one.
- In the write-up, make the model produce a fourth list. Not just decisions, actions and open questions, but everything that sounded like a decision and was never confirmed. That list is where your next escalation is hiding.
Then keep RAID hygiene tight. Anything parked gets an owner even though it has no decision - a parked item with no owner is a risk with no risk id. Anything recommended gets a named decider and a date by which they will decide, or it is not a recommendation, it is a hope. And anything that turns out to be a risk goes into the register with an id, the way R-03, R-07 and R-11 did, rather than living in a set of minutes where nobody will look for it.
"We agreed the contractors, didn't we?" In a Northwind-shaped programme, a steerco spent ten minutes on contract engineers for the hiring gap, ran out of time, and moved on. The notes said "agreed to bring in contract support". Nobody had approved a budget, and the sponsor had said "let us look at that" - which the note-taker heard as yes. Four weeks later the hiring gap was still open, no contractors had been engaged because no one had funded them, and the milestone behind it moved. The word "agreed" did all that damage.
Self-studyBefore you record anything - consent and policy2 min read▶
The practical bit, and it is not optional. A transcript is a recording of people who may not have expected one.
- Get consent before recording, at the start of the meeting, out loud, and let people say no. In some jurisdictions this is a legal requirement rather than a courtesy; in all of them it is the difference between a tool and a trust problem.
- Check what your organisation allows before a transcript goes anywhere near an AI workspace. Some permit the meeting tool but not the export, some permit neither, some have a policy nobody has read. Finding out is a fifteen-minute job and it is yours.
- Watch what is in the room. Steercos wander into people, performance, vendors and commercial terms, and the b1 house rules were clear: no personal data, no client-confidential terms, nothing under NDA. A transcript is exactly where all three arrive by accident. If in doubt, work from your own notes - everything tonight works on notes, and the transcript is faster rather than necessary.
Transcript to decisions, actions and open questions ★ 14 min · everyone builds
Here is a genuine forty-second slice of the Northwind steerco - people talking over each other about the ingest slip. Read it first and try to extract the decisions yourself. Then run the prompt and compare. The interesting part is not what you both found; it is the fourth list.
Paste the notes or transcript into your Northwind workspace, with the context pack loaded. The pack is what lets the model turn "Priya" into Priya N., vendor manager, and "the 24th" into the M3 gate date.
Run the prompt below and read the fourth list first - the things that only sounded like decisions. Then take every "TBC - ask" line to the actual people. That is rail two: the dates come from the team.
Add the confirmed decisions to the log with ids, and only then write the follow-up message in Demo 2.
The same-day message that closes ambiguity ★ 8 min · build your own
Extraction is half the job. The other half is a short message, sent the same day, that turns your reading of the meeting into the room's reading of the meeting. It works because of one line: if you disagree, say so by a stated time. Silence after that is agreement, and everybody knows it.
Send it the same day. Tomorrow it is a summary; today it is a correction window.
Keep it to one screen. Decisions, actions, open questions, and the deadline to object. No preamble, no thank-yous, no recap of the discussion.
Name the objection deadline explicitly and make it real - end of the next working day is usually right. Then send it to everyone who was in the room and everyone whose week it changes. Dan R. was in the room; his BI team was not.
Try it yourself - this week ◐ 30-45 min total
- Start a decision log for a real programme, six fields, and backfill the last five decisions from memory. The cost field will be the hard one, which is the point. Then run your next meeting through the four-list prompt within an hour of it ending, and keep the fourth list even if it is uncomfortable.
- Send the same-day message with a stated objection deadline, once, and look at what came back. Then say one decision back out loud in your next meeting: "so the decision is X, and you are taking it?" About a third of the time there was no decision.
- Check your organisation's policy on meeting recordings and transcripts before you export one. If nobody can tell you, that is a finding worth raising.
Official sources covered
Taught from the PMI delivery canon, the by-hand versions of these artifacts in the PMO course, and PMI's guidance on human oversight of AI-assisted work. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Your extraction returns a decision recorded as "the steerco agreed to keep legacy running". What is wrong with it?
"The steerco decided" means nobody decided. A forum cannot be asked to reconsider or explain itself in eight weeks; Sofia K. can. D-15 has a name on it for exactly that reason.
2 · The model returns an action with no stated owner. What should it have written?
An inferred owner is an invented commitment, and "the team" is how an action survives three sets of minutes untouched. The gap is the finding - go and get a real name.
3 · Why does the extraction prompt demand a fourth list of things that only sounded like decisions?
Ask for decisions and you get decisions - fluently. Forcing a separate list of unconfirmed agreements turns a smoothing machine into a check on the room, and it is the cheapest escalation you will ever run.