The eighteen months before a first model
The model took nine weeks. The eighteen months before it were the engagement.
The program
| Sector | Government |
|---|---|
| Scope | Data ownership, lineage, one model to production |
| Model build | 9 weeks |
| Everything else | 18 months |
| Operations owner | A named team, agreed in month four |
| Status | In production, with an on-call rota |
The agency had eleven pilots and nothing in production, which is a very common number and always means the same thing. The pilots were not bad. They were unownable.
Eighteen months went on three things. Finding an owner for four data sources, three of which had none — not a team, not a system, a person willing to say in a meeting that this was theirs and that they would be told before it changed. Establishing lineage well enough that a decision could be explained to somebody who had complained about it, which is a much higher bar than being able to explain it to an engineer. And negotiating an operations team who would agree to be woken up.
That third one took the longest and nearly ended the program twice. A team that is asked to carry a model somebody else built, on data they do not own, at three in the morning, is right to refuse. What unlocked it was giving them the rollback: one command, tested, that puts the previous decision logic back in under a minute, with no meeting required.
The model went into production in the ninth week after that was signed. The first time it was wrong, the on-call engineer rolled it back at 04:40, went back to bed, and told us in the morning. That is the whole point.
We were asked twice to skip the ownership work and start on the model. Both times we said we would rather not start.
>
— Delphine Ostrowski, Founder and Chief Executive
The second chair
The chair beside the monitoring desk is squared up and in use, with a jacket over the back of it. That chair is the outcome; the model is an implementation detail.
