Nearly every export boss asks us the same question: "Won't it quote a wrong price?" — and it's the right question. One invented, wrong price costs you more than a buyer; it costs you your reputation in the industry. Most AI tools answer "you review it before sending", which hands the risk straight back to you and leaves the automation half-finished.

Our answer is different: we turned "dare we automate?" into an engineering question, not a luck question. Below are the three safety gates every auto-sent reply in TradovaAI must pass.

Gate 1: Blacklist pre-filtering — spam never reaches the AI

A sender blacklist plus a keyword blacklist stops ads, mass blasts and competitor probing before anything enters the AI. This layer saves more than compute — it saves your reps' attention: whatever makes it past this gate is at least a message worth reading.

Gate 2: Intent whitelist — only low-risk actions get through

Even a genuine inquiry isn't all fit for AI to answer. The system runs intent recognition first:

  • Cleared for auto-reply: price-list requests, catalog requests, lead times, specs, stock — questions with verifiable answers, where a stumble does no real damage;
  • Always routed to humans: complaints, legal matters, price negotiation, contract terms, exclusive distribution — every one of them needs a person to decide.

The whitelist is yours to define. Tighten it or widen it as you like — the boundary rules live in configuration, never in the AI's on-the-spot judgment.

Gate 3: Confidence threshold — guesswork never enters the send queue

Every reply the AI generates carries a confidence score and cited sources: which knowledge-base entries it drew from, which document each figure came from. Replies below the threshold are automatically downgraded to drafts and land in the human review queue — never sent to a buyer directly.

That's what "traceable answers" actually means: not a slogan, but a mechanical check before sending — say only what you can verify; when in doubt, stay silent.

The fallback: one sentence short beats one answer invented

Beyond the three gates there is one iron rule: when a question isn't covered in the knowledge base, the AI says so honestly and hands the conversation — full context attached — to your team. The buyer never has to repeat themselves, and your rep already knows the whole story when they step in.

Why the line between "auto" and "human" must be mechanical

Plenty of tools run AI replies and human replies as two disconnected worlds: the AI blunders through a round first, and a person swoops in once the customer is annoyed. We built it the other way around — define what the AI may do first, and everything else belongs to people by default:

  • AI handles: first responses, common Q&A, screening and scoring, follow-up nudges — high volume, repetitive, low risk;
  • People handle: special pricing, negotiation, complaints, contracts — low frequency, high value, high stakes.

Every auto-send leaves an audit trail you can trace: who sent what, when, and on what basis — one lookup, after the fact.

The one-sentence summary

Full-auto is only safe when the things that must not be automated are blocked by the mechanism itself. The harder the guardrails, the further you can open the throttle.

If you'd like to watch these gates handle real inquiries, book a demo — we'll test live with your own product data, including a few questions your documents deliberately can't answer.