Customer-facing agents need trusted context, not just more context
Sales and customer support agents can look strong in demos but still be risky to use directly with real customers. The core problem is not only how much context the agent has, but whether that context can be trusted. Customer records often mix what the customer actually said, what a team member guessed, what was true months ago, and what the model invented.
A retriever may treat all of those as the same kind of information, so an agent can act on an outdated note or a as if it were a real promise. In one real case, an agent congratulated a customer for expanding with the platform because it found a real deal note from an account that had already churned two quarters earlier. The note existed, but it was no longer true.
A better setup adds , freshness, action boundaries, and proof to customer knowledge. The agent needs to know where a fact came from, whether it is still current, which actions need , and what evidence it used.
Key points
- Customer-facing agents often still need a step before contacting real customers.
- Customer context can include facts, guesses, outdated notes, and .
- A retriever may not know which pieces of information are reliable or current.
- Useful customer knowledge should include , freshness, action boundaries, and proof.
- Sending messages or writing to a CRM should have stricter checks than drafting a response.