An AI SDR is software that does part of a sales development rep’s job on its own: it researches accounts, writes and sends outreach, sorts the replies and answers the simple ones. It is faster than a person and far cheaper per message.
A human SDR wins where judgment matters: the awkward reply, and the qualifying conversation that turns interest into a meeting a rep accepts. So most B2B teams should split the job, with software on research and drafts and a person on replies and calls, and test any autonomous setup first.
What an AI SDR actually does
The label covers two different kinds of product. An assisted tool researches and drafts, and a person reads, edits and sends. An autonomous one, usually sold as an AI agent that works toward a goal on its own, sends by itself and answers replies before anyone reads them.
Most of what follows turns on that difference. An assisted tool changes how fast your SDR works. An autonomous one changes who is speaking to your buyer, and that brings in deliverability, the law and your brand in ways a drafting tool does not.
In either mode the work breaks into the steps a person already does: finding accounts, choosing contacts, writing, sending, sorting replies, qualifying and booking the meeting. The useful question is which steps to hand over, not whether to hand over the job.
Where software wins, and where a person does
Taken step by step, the early work rewards speed and breadth and the later work rewards judgment. That is roughly where the line falls.
| AI SDR | Human SDR | |
|---|---|---|
| Account research | Fast and broad across websites, news and job posts | Slower, and better at noticing what the research missed |
| First drafts | Good, and quick to vary by role and industry | Better at the one line that shows real understanding |
| Sending | Any volume at any hour, which is also the risk | Limited by the hours in a day |
| Sorting replies | Reliable on out of office, unsubscribe and “not me” | Needed for anything ambiguous |
| Questions and objections | Uneven, and a wrong answer goes out under your name | Strongest on price, timing and fit |
| The qualifying call | Weak, and AI voice brings legal limits of its own | Where the quality of the meeting is decided |
| Handover to a rep | Can book the slot | Can tell the rep why the meeting matters |
The more a step depends on reading a person, the less of it should go to software. Test an autonomous AI SDR hardest on its answers to questions and objections, because each one goes out in your company’s name.
What buyers do with automated outreach
Buyers are getting more of it, and they are checking it with people. Forrester’s 2027 B2B go-to-market predictions expect buyers to meet more automated outreach and agent-mediated interactions, and to double their human interactions to validate what AI recommends.
The reason, Forrester says, is that AI cannot fully judge a provider’s fit, feasibility, trust or accountability. That is an SDR’s job in miniature. An inbox full of machine-written messages makes the person who can answer a hard question rarer, and worth more.
Sharyn Leaver, Forrester’s chief research officer, framed the year’s challenge as “deliberate choices about what to automate and where human judgment still matters”. For an SDR team, that choice is made step by step, not once for the whole job.
Volume spends your domain reputation
An autonomous AI SDR sends as much as you let it, and every spam complaint lands on the domain it sends from. Gmail’s sender guidelines ask senders to start with low volume to engaged recipients, raise it slowly and avoid bursts.
A tool built for volume will not do that unless someone sets the limits. The same guidelines ask senders to keep the spam rate below 0.10% and never to reach 0.30%, and a fast ramp to strangers is the quickest way past both.
Send from separate domains, warmed before any cold volume, never from your main one, so a bad week cannot reach the mail your company runs on. And count the quieter cost: a buyer who ignored ten generic messages is harder to reach when a person finally writes.
The rules that follow the sender
Buying the tool does not move the legal duty to the vendor. The US Federal Trade Commission’s CAN-SPAM guide says the law makes no exception for business-to-business email, and that you “can’t contract away your legal responsibility” by hiring another company to send it.
Selling into Europe adds a disclosure duty. Since 2 August 2026, Article 50 of the EU AI Act has required providers of AI systems that interact directly with people to make sure those people know they are dealing with AI, unless that is obvious.
The European Commission’s guidelines on Article 50 count a system that writes and sends messages to people as interacting directly. Their example of a disclosure is “an email generated by an AI agent sent to a natural person that features an AI label at the top”.
The line falls between the two modes. A person using AI to help them write is outside the rule, and the mere chance of human review does not take an autonomous system out of it. The duty sits with the tool’s provider, so ask yours how it labels mail to European buyers.
AI voice is stricter. As a May 2026 Congressional Research Service report explains, the FCC treats AI-generated voices as artificial under the Telephone Consumer Protection Act, so robocalls using them generally need explicit, documented consent. Keep cold calls to mobiles human.
Compare cost per meeting held, not per email
AI SDRs are sold against salary, and a human SDR does cost far more than the salary on the offer letter. But the licence is not the whole cost of the software either. Count these on the AI side:
- The licence, and whether it is priced per seat, per contact or per meeting.
- Contact data and enrichment, if the licence does not include them.
- The extra sending domains and inboxes that volume needs, and the weeks it takes to warm them.
- The hours a person spends reviewing drafts, replies and the meetings it books.
- The meetings a wrong answer costs, which no invoice shows.
Then divide each side’s total by the meetings held and accepted by a rep, not by emails sent or meetings booked. Software that looks cheap per email can still be expensive per meeting worth having.
How to test an AI SDR before you trust it
Run it as an experiment, not a rollout. The point is to learn whether it books meetings your reps accept, at a cost you can live with, without hurting the domain or the brand.
- Pick a segment you can afford to lose, never your best accounts, and keep a matched group worked by your own people to compare against.
- Start in assisted mode, with a person approving every message, and loosen it only for message types that keep passing review.
- Send from separate, warmed domains with a daily cap set before launch.
- Read every reply in the early weeks. Replies show what buyers make of the messages sooner than any dashboard.
- Measure meetings held and accepted per hundred accounts against the comparison group, alongside spam complaints, unsubscribes and angry replies.
- Write the stop rule before you start: the complaint rate or rejected meeting rate that ends the test.
Let it run until the meetings have been held, not just booked. A test that ends at the booking measures the calendar, not the pipeline.
Which one to choose
Lean on software when the market is large, deals are small, the offer is simple to explain and speed matters more than nuance. Keep people in front when the list is short and every account counts, as in account-based marketing, or when the sale is complex or regulated.
Most teams sit between the two, which is the case for the split: software takes the research and the first draft, and a person takes every reply that needs judgment and every call. The SDR’s job does not disappear. It moves toward the conversation.
● About the author
Mark Zides is the founder and CEO of Katama. He founded it to do the work rather than sell it. Meet the team.