Shipped: auto-generated post-mortem themes
After 200 debriefs the themes cluster predictably. The clustering is now automatic — here's what ships, what it does, and what it refuses to do.
Shipped to production yesterday: auto-generated post-mortem themes.
When a proposal closes as won or lost and you run the post-mortem, the product has always supported a structured questionnaire — what did we get right, what did we miss, what would we change. The new feature clusters the free-text answers across every post-mortem in your tenant and surfaces the themes that recur. This is the cluster-and-retire surface the Win-Loss Intelligence page describes, applied to debrief free-text rather than bid-shaped themes.
What it does
After each post-mortem is saved, a background job runs the embedding over the free-text answers and adds them to the tenant’s theme store. A weekly aggregation then clusters new answers against existing themes and surfaces:
- Recurring themes — any theme cited in 3+ post-mortems in the last 90 days.
- Rising themes — any theme whose citation rate has roughly doubled quarter over quarter.
- Retired themes — themes last cited more than 180 days ago, moved out of the active set.
The output is a dashboard on /analytics/post-mortems with the top 10 active themes, their citation counts, and the proposals that cited them. Clicking a theme surfaces the representative answers.
Why this was the right hundredth feature to ship
Two reasons. First, post-mortem data sits in every mature proposal shop and nobody reads it. Leulu’s practitioner piece on debrief discipline makes the point: lessons don’t embed in workflow unless they’re surfaced at the moment of the next decision. Clustering is how they surface.
Second, we’d been hand-clustering themes across our own proposal retros for 18 months. Pattern was obvious. Automating the pattern is a small change with a high leverage.
What it refuses to do
Three explicit non-features.
No cross-tenant clustering. Your post-mortem data stays in your tenant. We do not cluster across customers and surface “trending themes in the industry.” That would be a different product and probably a bad one.
No automatic win-theme retirement. The dashboard suggests retired themes, but pulling a theme out of the active win-theme library is a manual action by the proposal owner. Automating that step would break the accountability loop.
No “AI-recommended next actions.” The feature surfaces what the team said. It does not tell the team what to do about it. Recommendations without the team’s own judgment would be category-typical AI slop.
Where it lives
Docs: /platform/intelligence/post-mortems updated with the new theme dashboard. API: the themes.list endpoint is in the public API for tenants that want to ETL into BI tools.
See the earlier Saturday 20-minute retro post for the discipline side of this. The feature supports the discipline; it doesn’t replace it.