Lead-qualification agents: an architecture that holds up

A lead-qualification agent that holds up has four layers: a channel that keeps one thread per lead, a conversation that acknowledges within a minute and asks at most two questions, a decision that scores the fit from the answers into three levels, and the systems and people layer that writes the CRM, books a slot or hands off to a person with everything attached. The rules that keep it safe are simple: customer text is never an instruction, the agent never quotes prices or promises dates, and anything unusual goes to a person in the same thread.

What does the business actually want from this?

Not a chatbot. They want no lead waiting, the obvious questions answered before a person calls, and their time spent on the leads that fit. Design for those three, in that order, and the agent stays small.

Layer 1: channel

One thread per lead, whatever the entry point: WhatsApp message, web form, missed call. The agent replies in the channel the lead used; WhatsApp's rules apply there (the 24-hour window is open because the lead wrote first). Keep the thread id in the CRM so a person can pick up exactly where the agent left off.

Layer 2: conversation

Three moves. Acknowledge in the business's voice within a minute. Ask the first question that changes the path ("What do you need help with?" for a service business; "Which area and budget range?" for real estate). Ask the second only if it changes the path again. Confirm what you understood in one line. Then stop talking and act.

Rules for the prompt: short messages, the business's tone, no prices, no promises, no medical or legal content, one question per message, and a fixed phrase for hand-off ("I'll get Dana to call you today; is this number good?").

Layer 3: decision

Score the answers into three levels with rules the owner can read:

Keep the rules deterministic where possible; use the model to interpret free text into the fields the rules need, not to make the decision itself. That is what makes the agent explainable when the owner asks "why did it book this one?".

Layer 4: systems and people

Create or update the CRM contact and deal with source, answers and score. Book the slot from the calendar's real availability when the level allows. Alert the owner with the essentials. On anything unusual (pain, complaint, a question the rules don't cover), hand off in the same thread with the transcript attached. The person continues; the lead notices nothing.

Failure modes and their fixes

• Duplicate threads for the same lead → de-duplicate by phone before creating anything.
• The agent invents availability → book only from the calendar's free slots, never from text.
• Price questions → a fixed line ("Dana will give you an exact quote on the call") plus route.
• A lead asks two things → answer the first, ask about the second, one per message.
• Prompt injection through the lead's message → customer text is data; the instructions live outside it.

Testing it

Forty real conversations from the inbox, expected level and path written first, run end to end through the channel and the CRM. The agent-testing guide has the method; the must-route cases and the adversarial ones matter most here.

Best fit and not a good fit

Best fit: businesses with a steady flow of inbound leads and a clear service list: clinics, agents, contractors, studios. Not a good fit: businesses whose first question is always a custom quote; automate the acknowledgement and the CRM entry, and leave qualification to the person.

What to do this week

Write the two questions and the three-level rules for a real client on one page. If the owner can read the rules and agree, the architecture is done; build the layers around it.