Synthetic control answers a common problem: the change happened in one place, so there is no group of similar places to compare with. The method builds one. It weights a set of untreated units so that their combined history tracks the treated unit closely before the change, then treats the gap that opens up afterwards as the effect.
It suits AI rollouts that go live in a single country of a regional business, or in one plant or one business line, with the other units supplying the donor pool. It needs a long, regular, unit-level history for every unit, which is why the baseline should be recorded weekly and by unit before the pilot starts.
The weights must be fixed before the results are seen. A synthetic control assembled afterwards can be tuned to show almost any effect. The method’s best-known use, by Abadie, Diamond and Hainmueller, estimated the effect of California’s 1988 tobacco control programme against a synthetic California built from other states.