AI Product Manager
When should an AI assistant interrupt you?
Claude Duet and Granola offer two ways to put information in front of people. A meeting-change exercise shows how to decide what deserves a notification.
At 9:42 a.m., a vendor moves a delivery from today to Friday. You’re meeting the client at ten. A message now would give you eighteen minutes to change the plan. The same message during dinner is a particularly irritating way to learn what you already know.
The facts in both messages are identical. A product manager evaluating only their accuracy would give both full marks. A customer would have a different scorecard.
Two products make different promises about attention
Anthropic introduced Claude Duet on October 21 as a research preview of an assistant that can notice developments across connected apps and bring them to you. On November 11, Granola introduced an agent brief refreshed each morning from meeting context. A connected agent can consult it for priorities, people, and projects. The first pattern can reach someone while a decision is still open. The second gives them context when they choose to look. Anthropic’s Claude Duet announcement · Granola’s agent brief.
Write both options into the spec. If the only behavior your team has designed is “send a push,” every fresh piece of information starts to look urgent.
Write three versions of the same moment
- Interrupt now. The delivery date changed, the client call starts in eighteen minutes, and you own the agenda. Show the old and new dates, the source, and a button to open the agenda.
- Include in the morning brief. The delivery date changed for a project you review next week. Include it under that project so you see the schedule in context.
- Stay quiet. The project owner already accepted the revised date and updated the agenda. Retain the event in history so someone can look it up.
Then add a fourth case. Two messages disagree on the new date. Show the conflict to the owner while there is time to resolve it. An assistant that silently chooses Friday could turn an unsettled question into a promise to a client.
Separate the alert from permission to act
“The delivery slipped” and “I’ve emailed the client” are different product behaviors. Anthropic describes showing users the proposed action, its destination and content before certain sensitive actions. For the fictional delivery case, let the owner inspect the source, open a draft, and review the recipients before sending. They may want to phone the client first. How Claude Duet handles approvals.
Score one week of decisions before sending a live alert
Replay a week of past changes with the people who owned them. Preserve what the assistant could have known at the time. For each case, ask the owner what they would have done differently and the latest moment a message could have helped. Record a short reason when the owner and the proposed rule disagree.
For illustration, imagine twelve proposed interruptions. Owners say eight would have changed a decision in time, so eight of twelve proposed alerts earned their interruption. Then review the events the assistant kept quiet. If three more needed timely action, the pilot caught eight of eleven actionable events. The first fraction tells you how noisy the assistant feels. The second tells you what people still have to watch themselves.
If most of the misses belong to one project with unusually tight deadlines, pilot a rule for that project. Making everybody’s phone buzz more often is an expensive way to solve one team’s problem.