Bad data is operating risk for agents.
People compensate for bad data with memory, skepticism, and informal checks. They know the CRM owner is stale, the product event is noisy, or the support status lags behind reality. Agents often do not know that. They treat the available context as if it deserves confidence unless the system tells them otherwise.
That changes the cost of data quality. A wrong field no longer only annoys an analyst. It can drive an automated recommendation, a customer-facing draft, a misrouted escalation, a bad forecast summary, or an action in another system.
The context layer should carry quality signals with the facts agents use. Known stale source. Low-confidence match. Duplicate record. Manual override. Unreconciled hierarchy. Missing owner. Instrumentation break. These labels help agents choose safer behavior.
Quality also needs severity by action. A minor typo in an internal note may not matter. A questionable billing status matters before sending a collections email. A low-confidence account match matters before updating CRM. The same data issue can be harmless in one workflow and risky in another.
This series is not generic data-quality hygiene. The question is narrower: what data problems create agent risk, and how should agents behave when they encounter them?
The answer should be operational. Warn, stop, ask for confirmation, escalate, or use a lower-risk action. Do not let the agent turn uncertainty into polished prose and call it done.
This is part 8 of 10 in The AI Context Layer.
