How to Build AI Sales Agents That Actually Close

An AI agent isn't a chatbot with a nicer prompt. It's an autonomous worker that perceives context, makes decisions, takes actions in your real systems, and knows when to hand off to a human. Get the architecture right and it becomes your most consistent sales rep — one that never forgets to follow up.

The anatomy of a production sales agent

Every reliable agent we deploy has the same five layers. Skip any one of them and the agent feels impressive in a demo but falls apart in production.

1. Perception — pulls context from your CRM, calendar, and conversation history.
2. Reasoning — an LLM decides the next best action given goals and guardrails.
3. Tools — concrete actions: book a meeting, enrich a lead, send a follow-up.
4. Memory — short-term conversation state plus long-term account history.
5. Escalation — a confidence threshold that routes hard cases to a human.

Define the job before the prompt

The biggest mistake teams make is starting with the prompt. Start with the job description instead. Write down exactly what a great human SDR would do, what they're allowed to decide, and when they'd escalate. Your system prompt is just that job description, formalized.

A minimal agent loop

Under the hood, an agent is a loop: observe state, choose a tool, execute, observe the result, repeat until the goal is met or escalation triggers. Here's a stripped-down version of the control loop we use.

type AgentState = {
  goal: string;
  context: LeadContext;
  history: Message[];
};

async function runAgent(state: AgentState) {
  for (let step = 0; step < MAX_STEPS; step++) {
    const decision = await reason(state); // LLM picks the next action

    if (decision.action === "escalate" || decision.confidence < 0.6) {
      return handoffToHuman(state, decision.reason);
    }

    if (decision.action === "complete") {
      return finalize(state, decision.summary);
    }

    // Execute the chosen tool and feed the result back into context.
    const result = await tools[decision.action](decision.args);
    state.history.push({ role: "tool", name: decision.action, result });
  }

  // Safety net: never loop forever.
  return handoffToHuman(state, "max_steps_reached");
}

Chat, voice, or both?

Guardrails that keep you out of trouble

Frequently asked questions