Guardrails are the rules and controls that keep an agent useful and safe. They should be designed before go-live, not after the first complaint.
Six guardrails every revenue agent needs
- Scope: which segments, lead sources and countries the agent may contact — and which it may not (strategic accounts, open opportunities, existing customers in a renewal).
- Grounding: approved sources for every claim — product, pricing, case studies. No invented features, prices or customer names.
- Volume and pacing: daily caps per domain and mailbox, follow-up limits, quiet hours by time zone.
- Suppression and consent: unsubscribe lists, do-not-contact flags, regional consent rules, competitor and partner exclusions.
- Hand-off: clear triggers for routing to a human (buying intent, pricing questions, complaints, legal questions) and a response-time target.
- Oversight: a named owner, a weekly sample review of messages, alerting on complaint or spam spikes, and a kill switch.
Design patterns that help
Salesforce's engineering team describes its lead-nurturing agents as constrained to predefined workflows and structured outputs instead of open-ended generation, with every response grounded in context like previous interactions and engagement signals, a priority model (replies first, then outreach and nudges) and a queue that prevents duplicate outreach.
Approval levels
Start with human approval of messages or batches, then move to autonomy for the parts that prove reliable. Keep approval for high-risk content (pricing, legal, executives).
Measure the guardrails
Track: messages blocked by rules, hand-offs and their response time, unsubscribes, spam complaints, and errors found in sample reviews. Rising numbers are an early warning.