AI agents for sales 2026: automating lead generation and correspondence

WORKING EXAMPLE: AUTONOMOUS OUTREACH CYCLE IN PYTHON

Below is a skeleton of an agent that goes through the full cycle: lead search → enrichment → email generation → sending → follow‑up planning. The code is stripped down to the essence, but the structure is production‑ready: state in DB, guardrails, step separation.

import time, sqlite3, datetime as dt
from openai import OpenAI  # любой совместимый клиент LLM

llm = OpenAI()
db = sqlite3.connect("sales.db")
db.execute("""CREATE TABLE IF NOT EXISTS leads(
    email TEXT PRIMARY KEY, name TEXT, company TEXT,
    stage TEXT DEFAULT 'new', last_touch TEXT, touches INT DEFAULT 0)""")

def find_leads(icp: dict) -> list[dict]:
    # тут вызов API провайдера данных по критериям ICP
    return data_provider.search(industry=icp["industry"], size=icp["size"])

def enrich(lead: dict) -> dict:
    # обогащаем: email, должность, свежий триггер (раунд/найм)
    lead.update(data_provider.enrich(lead["domain"]))
    return lead

def write_email(lead: dict, followup: bool = False) -> str:
    goal = "мягкий фоллоуап" if followup else "первое касание, зацепка на триггере"
    prompt = f"""Ты SDR. Напиши короткое письмо ({goal}) для {lead['name']},
    {lead['role']} в {lead['company']}. Триггер: {lead.get('trigger','')}.
    Правила: до 90 слов, один вопрос в конце, без штампов, по-русски."""
    r = llm.chat.completions.create(model="gpt-4.1",
        messages=[{"role": "user", "content": prompt}])
    return r.choices[0].message.content

def send(email: str, body: str):
    mailer.send(to=email, subject="Быстрый вопрос", body=body)  # прогретый домен

def save(lead: dict, stage: str):
    now = dt.datetime.utcnow().isoformat()
    db.execute("""INSERT INTO leads(email,name,company,stage,last_touch,touches)
        VALUES(?,?,?,?,?,1)
        ON CONFLICT(email) DO UPDATE SET
        stage=?, last_touch=?, touches=touches+1""",
        (lead["email"], lead["name"], lead["company"], stage, now, stage, now))
    db.commit()

def run_cycle(icp: dict, daily_limit: int = 40):
    sent = 0
    for raw in find_leads(icp):
        if sent >= daily_limit:  # гардрейл deliverability
            break
        row = db.execute("SELECT stage FROM leads WHERE email=?",
                         (raw.get("email",""),)).fetchone()
        if row:  # уже в работе — не дублируем
            continue
        lead = enrich(raw)
        if not lead.get("email"):
            continue
        send(lead["email"], write_email(lead))
        save(lead, stage="contacted")
        sent += 1
        time.sleep(20)  # человеческий ритм отправки

def run_followups(gap_days: int = 3):
    cutoff = (dt.datetime.utcnow() - dt.timedelta(days=gap_days)).isoformat()
    rows = db.execute("""SELECT email,name,company FROM leads
        WHERE stage='contacted' AND touches < 3 AND last_touch < ?""",
        (cutoff,)).fetchall()
    for email, name, company in rows:
        lead = {"email": email, "name": name, "company": company, "role": ""}
        send(email, write_email(lead, followup=True))
        save(lead, stage="contacted")

if __name__ == "__main__":
    icp = {"industry": "fintech", "size": "50-200"}
    run_cycle(icp)
    run_followups()

Run run_cycle and run_followups on a schedule (cron / queue). The DB guarantees that one contact won’t receive two first emails, and the limit keeps domains alive.

GUARDRAILS: HOW NOT TO KILL YOUR DOMAIN AND REPUTATION
Autonomy without limits = ban in a week. Mandatory minimum:

Domain warm‑up. A new domain shouldn’t send 500 emails per day. Start with tens and scale up gradually.
SPF, DKIM, DMARC. Without proper authentication, emails go to spam even before being read.
Limits and jitter. Random pauses between sends mimic human behavior.
Human-in-the-loop on escalations. A reply “send the contract” should not be handled by the agent.
Stop‑words and unsubscribes. Any negative feedback or unsubscribe instantly removes the contact from the cycle.


METRICS AND ROI: WHAT TO LOOK AT
Don’t chase volume of sends. An autonomous agent can easily inflate numbers that mean nothing. Real funnel metrics:

Reply rate — share of replies. Main signal of personalization quality.
Positive reply rate — share of positive replies. That’s what converts to money.
Meetings booked — scheduled meetings. Final goal of the SDR agent.
Cost per meeting — tokens + API + infrastructure, divided by meetings. Compare with the cost of a live SDR.

According to community measurements, a well‑configured agent for cold outreach yields a reply rate several times higher than template mass mailings — thanks to genuine personalization on triggers, not just {{first_name}} substitution.


Frequently Asked Questions
Will the AI agent replace a live salesperson?
No. It removes the routine of the top of the funnel — search, first touch, follow-ups. Closing the deal, negotiations and complex objections remain with the person. The agent hands the manager an already warmed-up lead.
Do you need a developer to build your own agent?
For a custom agent on LangGraph/CrewAI — yes, at least at the level of a Python junior. If there is no code at all, start with a ready-made SaaS, test the hypothesis, and then migrate to your own to save on scale.
Will the agent's emails end up in spam?
They will, if you ignore domain warm-up and authentication (SPF/DKIM/DMARC). With proper setup and reasonable limits, deliverability is no worse than manual mailing.
How much does it cost to launch such a cycle?
Your agent is mainly LLM tokens plus subscriptions for data and email. At moderate volumes, costs are usually multiple times lower than an SDR's salary, but the exact figure depends on the volume of leads and the chosen model.


We break down fresh agent frameworks, enrichment tools, and open-source SDR builds every day. If you want to be the first to see what really works in 2026 — subscribe to the Telegram channel REDDYX AI — new repositories every 30-60 minutes.