Qualifier + Closer
work while you sleep.

Run the math: you launch cold outreach to 1000 addresses, you get 150 replies. Out of those 150, 120 are three recurring questions: "how much", "how does it work", "do you have a case study". And 30 are "not interested / unsubscribe". Real human work hidden in there — maybe 10 conversations, the rest is mechanical Q&A.

The problem is that a human still spends 5-8 hours on those 150 replies. Opens each one, reads the context, replies in tone, doesn't mix up leads. Meanwhile new replies pile up, response goes out 6 hours later, the lead has already cooled off.

TG:ON's AI Agents solve exactly that problem. They don't "replace sales" — they take 80% of the routine off your plate and hand the human only those 10-20 conversations where a human is actually needed. Below — how it's built under the hood.

The difference between SaaS-chatbot and LLM-based agent

When people say "AI for sales", 90% picture Intercom/Drift — scripted decision trees: "if user wrote X → reply Y". That's not AI, that's if-else with a UI. Works on FAQ, breaks on any non-standard question.

An LLM-based agent works differently. It reads the whole conversation context (last 20-30 messages + lead metadata), reasons via a system prompt, and generates the reply from scratch every time. If the lead asks "do you have a Notion integration?" — Qualifier isn't matching that to a tree, it just knows the product from its system prompt and answers on point.

Concrete example:

The difference is tangible. The lead just closes the first reply. The second one continues the conversation, because they got a real answer plus a follow-up question that moves the funnel.

How Qualifier scores a lead

Qualifier is the first agent in the pipeline. Its job: read the incoming reply, understand the lead, assign a score. Here's how:

1. Context window: last 20-30 conversation messages + metadata (lead source, channel, previous outreach text, time of first contact).
2. Scoring signals: intent (are they clearly interested), budget (did they mention money / "how much"), timeline ("now / next month / later"), role (do they decide themselves or "I'll ask the team").
3. Output: score 0-100 + tags (hot / warm / cold), plus a recommendation: auto-reply / handoff to Closer / escalate to human.

Scoring formula (simplified):

In FAQ-mode, Qualifier answers standard questions itself: pricing, basic tech questions, how to start a pilot. If a question is beyond its competence — escalates further down the pipeline.

Conversation handoff and deal closing

When Qualifier tags hot, the conversation is seamlessly handed to Closer. From the lead's perspective — nothing changes: same tone, same account. Under the hood, what changes is the system prompt and the loaded context: Closer gets the full pricing sheet, case studies, and payment flow.

Example conversation where Qualifier handed off and Closer picked up:

Closer doesn't "push". Its system prompt includes a rule not to pressure: if the lead says "too expensive", it offers Starter ($49); if "not right now" — it politely closes the conversation with an offer to return in 2-3 weeks. On clear buying signals ("I'll take Pro", "let's get it going") — it sends a payment link and flags the conversation as closed_won.

Five providers, different trade-offs

TG:ON isn't locked to one LLM. You pick the provider and plug in your API key — it's stored locally on your machine, never goes through our server. Comparison:

Typical pilot setup: Claude for Closer (closing quality matters) + DeepSeek/Gemini for Qualifier (higher volume, quality less critical than at closing). Economics — 3-5× cheaper than "everything on GPT-4o" with comparable end result.

When AI goes silent and calls a human

An AI talking to a real customer is risk. One hallucination ("we offer a 60-day money-back" when in reality it's 14) = legal problem. That's why Qualifier/Closer have a hard-layer of hard-stops built in:

• Keyword blacklist: refund, complaint, lawyer, police, sue, contract, GDPR, chargeback, threat. Any of these in a lead's message → AI stops immediately, conversation status goes to escalated, operator is notified.
• Intent detection for financial topics: if the lead talks about refunds / complaints / competitor comparison in the format "you promised X, got Y" — auto-escalates, even without an exact keyword match.
• Max N turns: by default, AI runs a max of 15 messages in a single conversation. After 15 — soft handoff to human ("want to make sure you get the right answer — passing you to a colleague").
• Confidence threshold: if the LLM itself returns an answer with low confidence (model signal on uncertainty) — we don't send it, ask a human to review.

Escalation flow visualized:

Numbers, not promises

Through 2025-2026 we ran Qualifier+Closer on 40+ pilots — from B2B SaaS to infoproducts, 2 fintech clients. Median numbers:

Variance by niche:

• Infoproducts / courses: 45-50% auto-closed (FAQ-driven dialog, few nuances, AI handles it).
• B2B SaaS ($50-200 ticket): 35-40% auto-closed (more feature Q&A, but Closer carries through).
• B2B Enterprise ($1K+): 15-20% auto-closed (many stakeholders, AI only scores, human closes).
• Fintech: 10-15% auto-closed (regulatory, many hard-stops, AI mainly pre-qualifies).