What it is
Model drift is the slow drop in an AI model's accuracy as real-world data and behavior move away from what it was trained or tuned on. Over time this causes degraded responses, more hallucinations, and lower grounding accuracy. It can be caused by shifting data (data drift) or changing definitions of correct outcomes (concept drift).
Why it matters
Unchecked drift quietly erodes agent reliability and customer trust long after launch, making ongoing monitoring essential rather than a one-time governance task. Mid-market teams without dedicated MLOps staff are especially at risk of missing it.
Key components
- data drift detection
- concept drift monitoring
- retraining triggers
- accuracy baselines
How it connects
Teams running Agentforce watch for drift through agent telemetry and MELT monitoring, and correct it by re-grounding agents on updated Data Cloud sources or retuning the model.
Good to know
Set up scheduled accuracy audits and retrieval-quality checks rather than relying solely on user complaints to catch drift.
Related terms
Agent Observability
The practice of inspecting, debugging, and understanding AI agent behavior at runtime by consuming agent telemetry — traces, metrics, logs, and events — through dashboards, alerts, and evaluation tools.
Agent Telemetry
The runtime data emitted by an AI agent — every decision, tool call, input, output, latency, and cost — used to monitor reliability, quality, and spend in production.
AI Governance Framework
The set of policies, roles and controls that decide which AI a company uses, on what data, and who answers for the results.
MELT (Metrics, Events, Logs, Traces)
The four-pillar framework for organizing observability data: Metrics (numeric signals), Events (state transitions), Logs (per-step context), and Traces (end-to-end request chains).
