Glossary
Guardrail
A constraint applied to the agent to prevent answers that are out of scope, off-brand or ungrounded in your content. It covers faithfulness checks, topic blocking and masking of sensitive data.
A guardrail is a constraint that stops the agent doing something the model would happily do. Three kinds cover most of it: staying on topic, staying grounded in your content, and never emitting sensitive data.
Grounding is the one that matters commercially. A faithfulness check compares the generated answer against the retrieved passages and blocks it when it asserts something no passage supports: the difference between a wrong price quoted to a customer and a “let me get someone to confirm that”.
Guardrails cost latency, and in real-time voice there is not enough budget to run all of them. That is a deliberate trade-off, not an oversight: text can afford full checking, audio typically only affords the masking.
See also
PII masking
Removing or replacing personally identifiable information (tax IDs, emails, phone numbers) before the text reaches a model or a log. It is a practical privacy requirement for any agent that talks to customers.
RAG (retrieval-augmented generation)
A technique where the agent first retrieves relevant passages from your knowledge base and only then generates an answer based on them. It is what keeps the answer anchored in your content rather than in the model’s generic memory.
Start the conversation today
Launch an AI agent that qualifies leads and books meetings around the clock. Live in under ten minutes.
A guided demo · your agent live in under 10 minutes