Every founder who asks me for an "AI sales agent" has seen the same demo: a chat bubble that answers questions politely. It's impressive for ninety seconds. Then a real lead asks whether you ship to Canada, the bot invents a shipping policy, and someone quietly turns it off.
The gap between a demo and a system is not the model. It's everything around the model.
The agent is the smallest part
A production agent that qualifies leads and books meetings is really five components, and only one of them is an LLM:
- Ingestion — webhooks from the form, the ad platform, WhatsApp, and the CRM, normalised into one lead object.
- Qualification — a scoring pass that decides whether this lead is even worth a conversation.
- Conversation — the LLM, tightly scoped, with tools instead of opinions.
- Calendar state — real availability, real timezone, real double-booking protection.
- Handoff — the moment a human needs to take over, and the trail that lets them.
Skip any one and you get a toy.
Give it tools, not knowledge
The single biggest reliability win is to stop asking the model to know things and start making it call things.
Instead of stuffing pricing into the system prompt, expose a get_pricing(plan) tool that reads from your actual source of truth. Instead of letting it guess availability, give it get_slots(date_range). Instead of trusting it to remember the lead's history, pass the CRM record in as structured context.
If the model can't get a fact wrong, it won't.
The prompt then shrinks to what it's actually good at: tone, clarification, and deciding which tool to reach for.
Qualification before conversation
Not every lead deserves an agent. Running the model on every inbound is how you burn budget and annoy people.
I score first — source, budget signal, company size, intent keywords, whether they've been in the CRM before — and route:
- Hot → agent engages immediately, aims to book.
- Warm → agent engages, aims to educate and capture missing fields.
- Cold → nurture sequence, no agent, no tokens spent.
- Junk → tagged and dropped.
On the last system I shipped this cut LLM cost by about 60% and raised booked-call rate, because the agent was only ever talking to people worth talking to.
The guardrails nobody builds until they get burned
- Refusal on commercial terms. The agent may quote published pricing. It may never negotiate, discount, or promise a timeline. Those trigger a handoff.
- Confidence floor. If it can't answer from a tool, it says so and offers a human. "I don't want to guess at that — let me get you a straight answer" outperforms a confident wrong reply every time.
- Idempotent booking. Booking is the one irreversible action. It goes through a single function with a dedupe key, so a retried webhook never creates two meetings.
- Full transcript into the CRM. When a human picks up, they read the whole conversation. Nothing is more damaging than a rep asking a question the bot already asked.
What "working" actually looked like
The last one: median response time of 28 seconds, day or night. 3× more booked calls from the same ad spend, because nobody sat in a queue until Monday. And — the number I care about most — zero hallucinated commitments, because the agent was never in a position to make one.
That's the whole trick. Don't build a smarter bot. Build a system that gives a modest bot nothing to be wrong about.