My launch estimates ignore every launch I have seen
💎CORE ARGUMENT
When I estimated the Exceptions Inbox launch I did what everyone does. I looked at this project: this team, these stories, this customer, the risks I could name. I added the pieces, added a margin for the risks, and arrived at 18 February. It felt rigorous. It was a view from inside, and the inside view contains no information about how projects like this usually go.
That information existed, down the corridor. After the launch I went back through eight earlier Halden launches. On average they had arrived six weeks after their first announced date. I had been present for several of them and had sat in their retrospectives. None of it had entered my estimate, because each of those launches came with a story about why it was special, and so did mine.
We slipped on 12 February and launched on 11 March, three weeks late. By the house average I did well. By my own estimate I failed. The second sentence is the misleading one: the estimate was the error, not the project.
What I do now is mechanical on purpose, since judgment is the faculty that fails here:
- Write the inside estimate as before.
- Find the reference class and its record: launches of this size, at this company.
- Start from the record, and move towards the inside estimate only for reasons I could defend to a stranger.
The uncomfortable part is social. An outside-view date sounds like pessimism, or like not believing in the team. I now think the reverse. Quoting a date that eight predecessors missed is the disrespectful act, because it sets the team up to be late against a number that was never real.
Where I might be wrong. Eight launches is a small sample, and I chose which eight after I already knew what I wanted them to say. Some were late for reasons that have since been fixed.
It also suits me. Everyone is six weeks late turns my three weeks into a success, and I notice how quickly I reached for it.
How this note changed
- 23 March 2025: Written the morning after reading the chapter. Angry at myself, which shows.
- 28 March 2025: Took the anger out. Added the three steps while closing the launch project.
- 7 December 2025: Reread for issue 52. The social paragraph is the part I still find hard to act on.
🧱BUILT FROM
🌍WHY IT MATTERS
Every date I give now comes with the record of similar work beside it. If I cannot find a record, I say the estimate is a guess, in that word.
☑️NEXT STEPS
- Collect first-announced and actual dates for past Halden launches
- Put the six-week average in the launch closing note
- Give the Carrier Scorecards build estimate with its reference class
The inside view and the outside view, read twelve days after the launch; and an essay on why the typical estimate sits well below the average outcome.
The shape of the error is in 💎 Late projects are late by a multiple and not by a margin. The general rule is 💎 A number without its base rate is a story.
Planning from the details of the case at hand produces a best-case scenario with small adjustments. The distribution of outcomes for similar past cases is a better predictor, and it is ignored because each case feels unique from inside.
A date is also a tool. An ambitious one focuses a team, and a date padded to the historical average will be filled to the last day. The honest estimate and the useful deadline may be two different numbers.
The slip itself, the issue that told it straight, and the closing note of the launch project.