· By the AgentSDR team
The short answer
An AI SDR is software that does a sales development rep’s job: it contacts prospects, reads the replies, writes answers, follows up and gets a meeting booked. Some run fully on their own, and others draft for a person to approve.
This guide explains how the category works so you can compare tools without the marketing. It is written by the AgentSDR team, which makes an AI SDR, so read our parts with that in mind. For AgentSDR on one page, see the homepage. For the tools side by side, see best AI SDR tools.
What does an AI SDR do?
An SDR’s week is a loop. An AI SDR automates each stage of it:
- Prospecting. Choose who to contact and find a way to reach them. Some tools include a contact database. Others work from lists you import.
- Outreach. Send a first message and timed follow-ups over email, LinkedIn, WhatsApp or phone.
- Reply handling. Read what comes back, sort the interested from the uninterested, and stop outreach for people who opt out.
- Answering. Write the next message in the thread and send it, or hand a draft to a person.
- Follow-up. Keep chasing a lead who goes quiet, on a schedule, until they move or say stop.
- Booking. Get a meeting on a calendar and tell the right person.
A tool might cover all six or only some. That difference matters more than the label on the box. See the AI sales agent page for how AgentSDR covers each, and AI sales assistant for the version that supports a human rep.
Is an AI SDR the same as an AI BDR?
In practice, yes. Sales teams often split SDRs (who qualify inbound or outbound leads) from BDRs (who open new accounts), and vendors borrow both titles. No standard says an AI BDR does something an AI SDR does not. If a vendor draws a line between them, ask what tasks sit on each side.
What pieces does an AI SDR need?
| Piece | What it is | What goes wrong without it |
|---|---|---|
| Lead data | A list of people with a verified way to reach them, and enough context to write to them. | Bounces, wrong people and messages with nothing real to say. |
| Sending accounts | Mailboxes, LinkedIn accounts and numbers it may send from, each with limits and hours. | Spam folders, restricted accounts and bans. |
| Instructions and knowledge | Your offer, tone, rules and facts the AI may use. | Invented claims and generic copy. |
| Reply handling | Classification of replies, a draft for each, and stop rules. | Interested leads missed and uninterested ones chased. |
| A review loop | A queue where a person sees what the AI decided and can correct it. | Mistakes repeat at scale before anyone notices. |
| A CRM or record of each lead | One place for the thread, the stage and the next step across channels. | Two channels contact the same person with different messages. |
How do AI SDRs fail?
Deliverability
Cold email lands in spam when volume jumps, lists are dirty or messages look alike. An AI SDR makes it easy to send more, which makes it easy to hurt your domain. The fixes are boring: send from your own mailboxes, set a daily limit for each, keep sending hours human, leave a gap between sends and verify addresses. Our guide to cold email from Google Workspace covers the setup.
Account bans
LinkedIn and WhatsApp restrict accounts that behave like bots. New WhatsApp numbers that message many strangers on day one are a common case. Limits per account, warm-up for new numbers and an immediate stop for anyone who objects reduce the risk, and none removes it. Read the LinkedIn automation limits and WhatsApp B2B outreach guides before you scale.
Bad personalisation
An AI that has only a name and a company writes about the name and the company. A line about a podcast the person never recorded is worse than no line. Personalisation is only as good as the data it draws on and the rules it follows. Good tools use fields you control, let you say what the AI may claim and show you the draft.
Other failures to expect
- Misread replies. A sarcastic no or an out-of-office read as interest. Look for a stated reason and a way to correct the call.
- Follow-ups that ignore a reply. The sequence keeps going after the person answered on another channel. Check that a reply on any channel stops the others.
- Calls that overstate. A voice agent that does not say it is an AI, or promises things nobody can deliver.
- No stop switch. When something goes wrong you need to pause everything at once.
How do you evaluate an AI SDR?
- Start with a small list of people you would be happy to contact by hand.
- Check that you can see every message before it goes out, at least at first.
- Read the limits: per-account daily caps, sending hours, gaps between sends.
- Send a handful of replies to the system yourself, including an out-of-office and an unsubscribe, and watch what it does.
- Open the classification of each reply and check the reason it gives.
- Find the stop switch and use it.
- Ask where your data lives and who can read it. Self-hosted tools keep it on your server.
- Price the whole stack: the tool, the sending accounts, the data and the AI usage.
How does AgentSDR approach it?
AgentSDR is an open-source AI SDR under the MIT License. It runs on your own server and PostgreSQL database, and its AI uses your own OpenRouter key. Cold email goes from your own Google Workspace mailboxes, LinkedIn through Unipile across many accounts, and WhatsApp from your reps’ numbers or a cloud calling station, each with limits you can edit.
Every reply is classified into your stages with a reason and a confidence score, and the stage’s sequence drafts the answer. By default a person sends it. Sequences can be set to send follow-ups automatically, and to send the immediate reply after a 15-minute hold, and both are off until you turn them on. See the AI CRM, and AI calling for the voice agent, which is a trial feature.
It has no built-in contact database: you import lists or enrich with your own provider keys. Founders running their own outbound are the closest fit: see AgentSDR for founders. For the self-hosting model, see open source and self-hosting, and for the people-versus-software question, AI SDR vs human SDR.


