Drift (in AI)

When an AI agent's behavior changes over time - usually getting worse - due to LLM model updates, prompt subtle changes, or shifting user data. Drift is silent: the agent does not announce it is performing badly. Detection requires monthly eval-suite runs against a baseline test set. Mitigation: periodic prompt re-tuning and quarterly re-grounding. The agencies that catch drift fast keep accuracy high; the ones that do not lose customers.