1. Home
  2. Resources & Playbooks
  3. Guardrailed sales AI

Playbook 03 · Conversational AI & Control

Conversational AI That Converts Without Rogue Discounts: Code Guardrails vs. Prompt Instructions

Why sales AI must enforce price floors, daily spend caps, approval queues and human handoff in server code rather than relying on LLM prompts.

1-min readPublished Updated By the Supersei team

The number-one reason business owners hesitate to put an AI agent on their customer WhatsApp number is fear: "What if a clever buyer tricks the bot into quoting a 50% discount, promising an impossible refund, or confirming a fake payment screenshot?"

If your AI tool relies only on a system prompt ("Please do not give more than 10% discount"), a determined user can jailbreak it. In Supersei, guardrails are enforced in server code outside the language model:

  • Hard price floors: Sei reads from your locked price catalogue and cannot generate a quote or payment link below the floor price you set.
  • Gateway-verified payments only: Sei never takes a customer's word or screenshot that an order is paid; it checks the Razorpay or Stripe payment status API directly.
  • Approval inbox with undo: Sensitive actions wait in an approval queue for a manager to click Accept, and every automated step can be undone.
  • Instant human handoff: The moment a buyer is upset or asks a custom commercial question, Sei hands the thread to a human rep with a full summary and goes silent on that conversation.
The Supersei Playbook: Configure your knowledge base, floor prices, daily message caps and kill switch in Sei AI Sales Agent.

See AI sales agent in Supersei

From ad click to paid invoice.

Sei answers every lead on WhatsApp, calls and chat in seconds, books the meeting, collects the payment on Razorpay, and shows which ad paid for it.

Nothing charged for 30 days · Cancel anytime · GST invoice

Start free trial Book a demo