Every Friday at 2pm, in about 70% of the product orgs I talk to, there is an AI office hour on the calendar. Someone from a central AI team demos a tool, shares a prompt, or runs a Q&A. The audience nods, then returns to a workflow that is structurally identical to the one they had at the start of the meeting. Kill it.
Why they fail
AI office hours are the 2026 version of the agile transformation. Same shape, same failure mode: a centralized team running a ceremony that signals motion without producing it. I audited this format in seven product orgs in the last year. In all seven, the same pattern held. High attendance. High sentiment. Zero measurable change in shipping velocity, eval discipline, or agent integration into core workflows.
Three reasons it breaks. Broadcast does not change behavior. People change what they do when the tool is in their workflow and the old workflow stops working, not when they hear about it. Prompt libraries are an anti-pattern: a prompt that solved someone else's problem is not a prompt that will solve yours, and a shared library bypasses the craft of writing prompts against your own data and evals. And the tracked metric is attendance, not output, so attendance goes up while shipping stays flat and the program gets declared a success. That is exactly the dynamic that killed agile transformations, which tracked ceremonies instead of shipped value.
What to run instead
Three replacements, ranked by how much they move the org.
Paired shipping sessions. Ninety minutes, two people, one concrete feature shipped with agents in the loop. A driver who has shipped something similar sits with a rider who has not, and they build a real artifact for the rider's team. The eval: the artifact has to be in the rider's actual workflow within 48 hours, or the session was a tutorial and should not count. Rotate the rider weekly so the driver pool grows on its own.
Eval reviews. Thirty minutes a week. Each feature owner brings a scorecard, and the group asks three questions per feature: what moved, what broke, what did we kill. Every feature leaves with one explicit decision, logged.
Kill list reviews. Twenty minutes every other week. The agent stack generates a weekly kill candidate list, and the team decides what to formally retire. If nothing gets killed for three consecutive reviews, the stack is not surfacing real signal.
How to make the switch
Do not kill the invite, repurpose it. Week one, keep the slot and run a paired shipping session instead of a demo. Week two, add a 10-minute eval review at the start. Week three, add a 10-minute kill list review at the end. By week four the meeting produces three artifacts per session. In every org I have watched do this, the meeting either becomes the most useful 90 minutes on the calendar or quietly stops happening. Both are wins.
Pick your next AI office hour. Convert that one slot into a paired shipping session this week and see whether anything ships within 48 hours.