Automation applied to an unreliable dataset does not reduce error; it distributes it faster and with more authority. The output reads well, which is precisely the risk.
We start with the narrow, well-governed data an organisation can defend, and expand scope only where the record supports the claim being made.
