What is an AI SDR agent?

An AI SDR agent is an automated system that takes on the role of a Sales Development Representative: it qualifies incoming leads using a script (BANT, MEDDIC), conducts an initial dialogue in Telegram or email, and transfers the warmed-up contacts to the manager. Works via n8n + Claude API without writing any code.

How much better is AI SDR than live SDR in terms of conversion?

On a flow of 50 leads per day, CR from lead to qualified lead: 35–55% for AI SDR versus 20–30% for live SDR. The main reason is the speed of the first contact: AI responds in 30–120 seconds, live SDR – on average in 4–8 hours. Speed directly affects conversion.

Do you need a CRM to run AI SDR?

At the first stage, no. A Notion spreadsheet or Google Sheets is enough to store qualified leads. It is better to add CRM integration (amoCRM, Bitrix24) in the second iteration, when the flow is stable and the logic for transferring transactions is clear.

What can't AI SDR do?

Three weak points: complex technical objections (“we already have a contractor, prove that you are better”), non-standard requests outside the script and emotionally charged answers. All this requires escalation to a live manager - the agent must be able to convey the dialogue on time, and not try to close it himself.

All articles

AIJuly 28, 202512 min

AI SDR agent in the sales funnel: automatic lead qualification without hiring an SDR

SDR (Sales Development Representative) is a position that AI is beginning to replace. Analysis: what is AI SDR, how it works in TG and email, what tasks it takes on, where it breaks down, and how to set it up via n8n + Claude API without CRM integration at the first stage. With real CRs from practice.

SDR is the position that the AI begins to replace first. Not because it is the simplest, but because it is the most scripted. Most of the SDR’s work is the qualifying conversation on BANT: there is a budget, the decision is made by himself, the problem is real, the deadline is not “someday”. In my projects AI SDR agent on n8n + Claude API does this in 30–120 seconds after the application and gives a CR of 35–55% per qualified lead.

CR from lead to qualified lead: 35–55% for AI SDR versus 20–30% for live SDR with a flow of 50 leads/day. Setup takes 2–3 days, without a developer. Read more about how I generally build automation in marketing - article about AI agents in n8n.

What is SDR and why is it the first position to replace AI

SDR (Sales Development Representative) is a person who processes the incoming flow of leads: first contact, qualification, setting up a meeting with AE (Account Executive). Typical funnel: lead leaves a request → SDR calls or writes → finds out BANT → either transfers it to work or sends it to research.

Economics of the role: junior SDR in Moscow costs 80–120K ₽/month, processes 40–80 leads per day under normal load. With a flow of 80–100 leads per day, you already need 2 SDRs. The problem is that 60-70% of an SDR's time is spent on routine qualification questions, which are the same for 90% of leads. This is the replacement point.

Why SDR and not another role? Three reasons:

The script is rigid - the same BANT questions in a predictable order
The data is structured - an incoming lead always contains a name, contact, source
The decision is binary - qualified or not, the meeting is scheduled or not

Copywriter, designer, analyst - they have subjective tasks where the script does not work. SDRs spend 15–20% of their working time on such tasks. This makes the role an ideal candidate for partial replacement.

How AI SDR works: from lead to meeting

The data flow diagram looks like this. The lead leaves a request on the website or through a TG bot → webhook is triggered → n8n receives data → Claude API receives context + BANT script → Claude generates the first message → TG bot or email delivers to the lead → lead responds → Claude classifies the response → n8n routes: continue qualification, schedule a meeting or escalate to a manager.

Key parameters of this scheme:

First response time: 30–120 seconds (vs. 4–8 hours for a live SDR)
Number of dialogues in parallel: unlimited (one live SDR - 10–20 simultaneous)
Work 24/7 - a request at 23:47 receives a response at 23:47
Qualification cost: 8–25 RUB per lead (API calls Claude + infrastructure)

Important point: the agent is not trying to close the deal. His task is to qualify and set up a meeting with the manager. As soon as the lead has confirmed BANT, n8n creates a card in CRM or Notion, makes an appointment and notifies the manager with a ready-made context: “Sergey, 34 years old, B2B services, budget 200K+ ₽/month, makes the decision himself, the request is urgent. Meeting tomorrow at 14:00.”

About other automation scenarios that work well next to AI SDR - in 10 ready-made n8n scripts for marketing.

What AI SDR Does Well

Three tasks that the agent performs at the level of a good SDR or better:

BANT Qualification

Budget, Authority, Need, Timeline - four questions to figure out. Claude holds the script better than a tired employee on a Friday night. Prompt: “You are the SDR of Company X. The goal is to figure out 4 BANT criteria in 3-5 message exchanges. Don't sell, don't mention prices. After BANT confirmation, offer a meeting with the manager in the slot from Calendly.”

Primary contact via TG and email

First response speed is a key conversion factor. According to my projects: a lead that receives a response in 5 minutes converts 2.3 times better than a lead that receives a response in 4 hours. AI responds in 30–120 seconds at any time of the day.

Make an appointment

Calendly API + n8n: as soon as Claude returns outcome: qualified, webhook automatically sends the lead a link to a specific time slot. The meeting confirmation is added to the manager's Google Calendar. No “text me on Monday.”

Where AI SDR breaks down

An honest analysis of the limitations that I saw in three real projects:

Complex objections

“We already have a contractor for this task, prove that you are better” - Claude will try to answer, but without specifics about competitors and without the living authority of the manager, this converts worse. Escalation is needed here. Rule: any comparison with a competitor is an exit from the automatic script.

Non-standard requests

The lead asks a question that is not in the FAQ or agent context. Claude is hallucinating or giving an inaccurate answer. Solution: in the system prompt, write explicitly - “if the question is out of context, do not come up with an answer, answer: “good question, I’ll check with the team and get back to you” and escalate.”

Emotionally charged clients

An irritated person who has already had a negative experience with a similar product feels the bot in 2 exchanges and becomes even more irritated. Indicator: negative wording in the answer + short remarks. The agent should escalate this dialogue to the manager immediately—before the situation gets worse.

Stack: n8n + Claude API + webhook - step-by-step setup

I describe the minimum working option without CRM at the first stage. The entire stack is free up to a certain amount:

n8n self-hosted for VPS from 300–400 ₽/month. An alternative is n8n.cloud from $20/month, but for 50–200 leads per day VPS is cheaper.
Webhook node in n8n — entry point. You paste the URL into a form on the website or into a TG bot via BotFather + Telegram Bot API.
HTTP Request node → Claude API. Endpoint: https://api.anthropic.com/v1/messages. Pass the system prompt + history of the dialog in the messages[] format. Model: claude-sonnet-4-6 for a balance of speed and quality.
Code node to parse Claude's JSON response. Check the outcome, BANT fields.
Switch node: qualified → Notion + Calendly; needs_more_info → continue the dialogue; escalate → TG notification to the manager.
Telegram Bot API or SMTP to send a response to the lead.

In the second iteration, add amoCRM or Bitrix24 instead of Notion. Complete setup from scratch - 2-3 days if you have a ready-made BANT script. Without a developer. Learn more about how to build funnels on this stack - article about building a sales funnel in 2026.

AI SDR vs Live SDR: Comparison Table

Parameter	AI SDR (n8n + Claude)	Live SDR (junior)
First response time	30–120 seconds	4–8 hours (working hours)
CR lead → qualified lead	35–55%	20–30%
Parallel dialogues	Unlimited	10–20 at a time
Cost of qualification	8–25 RUB/lead	150–400 RUB/lead (FTE/stream)
Open 24/7	Yes	No (8–10 hours/day)
Complex objections	Bad - escalation needed	Good (with experience)
Non-standard requests	Bad - risk of hallucination	Okay
Settings	2–3 days	2–4 weeks of onboarding
Scaling	Instantly (traffic growth does not change the cost)	Hiring + onboarding for every +50 leads/day

The optimal model is a hybrid: AI SDR qualifies 70–80% of the flow, live SDR only works with complex cases and closes meetings with large accounts. In my practice, this reduces the SDR staff by 2-3 times while increasing the flow of leads.

Case: B2B SaaS, 80 leads/day, AI qualifies 72%

Project: B2B SaaS for document automation, average bill 80–150K ₽/month. Incoming flow: 80 leads per day from TG bot, website and cold outreach. Before the introduction of AI SDR, one junior SDR worked full time - he handled 30-40 leads a day, the rest hung without response for 6-12 hours.

The setup took 3 days: day 1 - description of the ICP and BANT script, day 2 - setting up n8n flow and integration with the Telegram bot, day 3 - test on 20 real leads, editing the system prompt.

Results after 30 days:

AI qualifies 72% of incoming traffic without human intervention
CR per qualified lead: 48% (was 22% for SDR on the same stream)
CR for an appointment: 28% of the total flow (was 11%)
Average first response time: 52 seconds (up from 5.4 hours)
SDR now only works with 28% of the flow - complex cases and large accounts
Qualification cost: 18 ₽/lead (was 220 ₽/lead)

What went wrong at the start: the first version of the prompt did not contain escalation instructions for non-standard requests. Claude tried to answer questions about integration with SAP, which was not in the context - he gave inaccurate data. The fix took 20 minutes: I added an explicit trigger “if the question concerns integrations outside the list, escalate.”

More details on how to build marketing in B2B SaaS in Russia - separate article with metrics and channels.

Ethics and Disclosure: Should You Tell Your Client They're Talking to AI?

A topic that most “AI SDR guides” avoid. My answer: talk. Not directly in the first message - but transparently and without trying to pretend to be human.

This is my practice. First message: “Hi [name]. I am an AI assistant for the [company name] team. I process incoming applications and help find the right expert. A few quick questions and I’ll pass you on to the manager.” No human names, no “my name is Alexey.”

Why this works better than masking:

Expectations are adjusted - the client is not annoyed when he notices “bot patterns”
Trust is higher - B2B audience appreciates transparency and immediately understands why the answer came in 40 seconds
Reduces risk - in the Russian Federation there is no clear regulation for identifying AI in dialogues, but it’s a matter of time

The only argument I've heard against disclosure is "people don't trust bots." According to my data, on 3 projects, the CR qualifications of a transparent AI agent are higher or equal to the masked one. The client answers the questions in the same way, but irritation when the bot is “exposed” nullifies the conversion.

About cold outreach via Telegram, which often comes before the SDR funnel - templates with a response rate of 8–15%.

If you want to figure out how this works on your lead flow, write @dipustovalov. Or via the form on the website. In a 30-minute call, it becomes clear whether AI SDR is suitable for your specific funnel - and how much it will save over the horizon of three months.