AI sends transcripts to all invitees; attendance drops. What now?

[INTERVIEWER] AI sends transcripts to all invitees; attendance drops. What now? Your AI feature automatically sends the meeting transcript to everyone invited. It works perfectly, but attendance drops. What now? The reflex is ripping the feature out. It did exactly what it promised, but a downstream metric moved the wrong way. A strong answer does not rush to remove anything.

It confirms the causal link, names the broken incentive, and fixes the incentive rather than the symptom. This is a second order effects question asking whether you can think one step past the feature just working. The interviewer is probing whether you understand that products change behaviour. That change can undo the goal even when the feature is flawless. A junior PM sees the metric drop and kills the feature.

A senior one asks whether the drop is even bad and fixes the underlying incentive. Establishing cause before action and questioning the metric itself is the whole point. Let me walk you through it. Before touching anything, I isolate the variable. Is attendance actually down outside normal variance, and did it start when the transcript autosend launched? Correlation is not cause, so I run a check.

I compare meetings where the AI sent a transcript against meetings where it did not, acting as an A/B test. I also compare attendance before and after launch for the same recurring meetings. If attendance is flat in the control group and down only where transcripts went out, the causal link holds. If it dropped everywhere, including meetings the feature never touched, something else is going on.

It could be a seasonal dip or a company push to cut meetings. The transcript feature would be a red herring. Never fix a cause you have not confirmed. That check comes first. Assuming the link is confirmed, I name the mechanism out loud. Naming the incentive is the core of this question. People used to attend partly because it was the only way to know what happened in the room.

Now the transcript arrives regardless, so the cost of skipping dropped to near zero. The feature removed the reason to show up. That is the classic second order effect. The direct effect of everyone getting the notes was genuinely good, while the indirect effect of nobody needing to attend was bad. Stating it cleanly tells the interviewer you actually understand what happened rather than just reacting to a red number.

Most candidates skip the next move, which is pushing on the goal before jumping to a fix. Is falling attendance actually bad? Maybe some meetings did not need those people. The transcript freed them up to do real work, which is a win. I segment the meetings to find out. For meetings where attendance genuinely matters like decisions, live discussions, and brainstorms, the drop hurts.

Fewer people means worse decisions and relitigation later. But for status broadcast meetings that could have been a document, the drop is the feature working exactly as intended. The fix should target only meetings where presence has real value. I must not blanket punish a useful feature just because one metric looked bad. Questioning whether the drop is even undesirable is a senior instinct.

Now I lay out options and commit. The first option is restoring a reason to be live. I make attendance provide something the transcript cannot, like live Q&A, influencing a decision in the moment, interactive polls, or breakout discussions. If the only value a meeting had was information transfer, it should be a document. The second option is changing when and how the transcript is delivered.

I stop automatically sending it to everyone the second it ends. I send it only to people who could not attend. I could also require a small action to get it, introducing a light cost. The third option is making it configurable by the organiser. The meeting owner decides whether transcripts autobroadcast. Decision meetings can withhold them to keep presence valuable, while informational ones broadcast freely.

The fourth option is reframing the goal. If attendance was only a proxy for people being informed, and the transcript delivers that at lower cost, maybe attendance was never the real goal. I explicitly question whether we are defending the right metric. Then I commit. I make the autosend feature organiser configurable and default it to sending only to nonattendees.

I add live only value for meetings that need presence. This fixes the incentive without killing a useful feature. The last step is shipping the change as an A/B test rather than just deploying it. I watch attendance recover for decision meetings, but I also watch guardrails like total time saved, transcript usage, and satisfaction. If attendance comes back but everyone hates being dragged into meetings again, that is not a win.

I have just traded one problem for a worse one. I measure whether the real goal improved, meaning informed teams making good decisions, rather than just checking whether the attendance proxy went back up. Let me make it concrete. A meeting assistant automatically emails the full transcript to every invitee the moment the meeting ends. Attendance across recurring team meetings drops 20 percent over three weeks.

To confirm cause, the A/B test shows attendance flat where transcripts were off and down 25 percent where they were on, starting right at launch. It is causal. The second order effect is that attendance was the only way to stay informed. The transcript removed that cost, so people skip. To see if it is bad, I segment the meetings.

The drop is fine for weekly status broadcasts, which should have been a document anyway. It is harmful for decision reviews, where fewer people means worse decisions and topics getting reargued next week. I fix the incentive by defaulting transcript delivery to nonattendees only. I let organisers of decision meetings turn autosend off, and add live only polls and Q&A to make showing up worth it.

I A/B test the fix. Attendance at decision meetings recovers to within 5 percent of baseline. Status meeting attendance stays low, which is intended. Total meeting hours per person falls, which is a genuine win. The metric I ultimately report is decision meeting attendance plus time saved, because that is the real goal underneath the proxy. A sharp interviewer will test whether you would generalise the lesson.

They will ask how you would catch a second order effect before it happened. The answer is that you wargame the incentive at design time. Before shipping, ask one blunt question of any feature that removes a cost. What did that cost used to buy us? Attendance was buying informedness, so removing the cost of attending was always going to reduce attendance.

You could have predicted it on a whiteboard. The discipline is a prelaunch pass where you list the behaviours the feature makes cheaper. You ship the guardrail metric for each on day one, not after the surprise. Here attendance would have been an obvious countermetric to watch from launch. The second push is whether every downstream metric move is a second order effect and how you avoid chasing ghosts.

The filter is causality plus materiality. You only act when the A/B test confirms the feature caused it and the metric actually matters to the goal. A 5 percent wobble on a metric nobody optimises is not worth a fire drill. A confirmed drop on decision meeting attendance is worth investigating. Knowing which downstream moves to chase is judgement. Showing that filter keeps you from overreacting to every twitch.

Here is what makes them lean in. First, you confirmed causality with an A/B test before touching the feature, instead of just assuming the transcript caused the drop. Second, you named the incentive precisely. The feature removed the reason to attend. You then questioned whether the drop was even bad, which is the senior move. Third, you fixed the incentive by delivering to nonattendees, adding live only value, and making it configurable.

You did not just fix the symptom by ripping out a feature that did exactly what it was designed to do. Now for the traps. The first is immediately proposing to remove a feature that worked perfectly. This throws away real value just because one downstream metric twitched. The second is never questioning whether falling attendance is actually a problem.

You end up defending a proxy metric blindly without asking what it was standing in for. The third is treating correlation as proven cause with no A/B test. You then confidently fix a phantom while the real cause sits untouched. To recap, confirm the effect is real and caused by the feature with an A/B test. Name the second order effect and the broken incentive.

Ask whether the drop is even bad and segment the data. Then fix the incentive rather than the symptom and A/B test the fix while watching the real goal. Carry this into the room. A good feature can move a downstream metric the wrong way through an incentive. Confirm the cause, ask whether the drop is even bad, and fix the incentive rather than the symptom.

---