Model drift
Definition
Degradation in an automated system's output over time as the underlying model is updated, or as the real-world inputs move away from what it was tuned on. It arrives quietly, without an error message.
A classifier tuned in January against a provider's model can behave differently in June after a silent version update. Nothing breaks; the output is simply worse, and nobody notices until a downstream number moves.
The defence is a small held-out set of examples with known correct answers, run on a schedule, with an alert on the failure rate. It is a day of work and it is the difference between finding out from a test and finding out from a customer.
Definitions are free.So is the cost review.
A short call, then a written review inside 72 hours: what your marketing costs, what it produces, and what is worth fixing first.