Smart replies aren't keyword matching. The AI reads the full conversation history, follows the campaign playbook, and crafts a response that fits the moment — every time, in seconds.
Before generating any reply, the system assembles the full conversation context: base guardrails (role, compliance scope), the campaign layer (persona, questions, product information, branch logic), and the complete message history. The AI reads everything before deciding what to say.
Base guardrails sit above every campaign. They enforce tone and format rules that apply regardless of how a campaign is configured — ensuring consistent quality across every conversation.
Stock openers — 'Great!', 'Got it', 'Thanks for sharing', 'Thanks for clarifying' — are explicitly prohibited. Acknowledgement is reserved for genuinely warranted moments, not used as a default conversation opener.
How it works: Enforced via the base guardrail prompt — applies to every campaign without configuration.
All messages use UK English spelling and GSM-7 characters only in English responses. No markdown, no emojis, no exclamation marks. The AI writes as a professional would write a business SMS.
How it works: Character encoding constraint is part of the base guardrail layer — never configurable at the campaign level.
When a contact sends a message outside the campaign scope — a joke, an unrelated question, or a request the AI can't handle — the AI declines briefly, includes the next campaign question, and returns PAUSE. The conversation continues.
How it works: Off-topic handling tested via golden conversation suite — the directive must be PAUSE, not END, for a first off-topic message.
If the AI returns a malformed response or a response that fails schema validation, the platform substitutes a safe fallback response and flags the turn. No broken message is ever sent to a contact.
How it works: Fallback turns are visually flagged in the operator platform's turn detail view — Friyay operators are alerted on any non-zero parse failure rate.
The Campaign Tester lets you throw edge cases at the AI before any contact receives a message.