On this page
Few areas of B2B have attracted more AI hype than sales development. “AI SDRs” promise to research prospects, write personalised emails, send them, handle replies and book meetings, all without a person. Some companies have tried replacing their SDR teams with them.
The early results were sobering. In a survey presented by Tomasz Tunguz at SaaStr Annual in 2024, 83% of respondents reported no tangible results from AI SDRs, and only 3% had generated real revenue with them. Many teams saw the same pattern: far more emails sent, lower reply rates, and damaged sending domains.
Our view, based on building outbound systems for clients: AI is transforming sales development, but not in the way the “autopilot” pitch suggests. The teams getting results didn’t buy a bot. They built a system around their people.
What an SDR actually spends time on
Break an SDR’s week down and most of it isn’t conversation. It’s preparation:
- Finding the right accounts and the right people in them.
- Researching each account to find a reason to reach out.
- Cleaning and updating CRM data.
- Writing and personalising messages.
- Logging activity and following up.
The conversations and the judgement calls, the parts that actually create pipeline, often get the least time.
Where AI is genuinely excellent
Account research. A language model can read a company’s website, recent news, job posts and annual report, and turn them into a short, relevant brief in seconds. This used to take a person ten to twenty minutes per account.
Signal summaries. When an account triggers a buying signal, AI can explain what happened and why it matters in a sentence the rep can act on.
Data work. Normalising job titles, matching contacts to accounts, spotting duplicates and filling gaps from public sources. Tedious for people, well suited to models, and easy to check with code.
First drafts. Given a good research brief and clear guidelines, AI writes a solid first draft that a rep can edit in a minute instead of writing from scratch in ten.
Reply handling. Sorting replies into interested, not now, wrong person, unsubscribe and out-of-office, then routing each correctly, is a near-perfect AI task.
Where full autonomy struggles
Volume without judgement hurts deliverability. When AI makes sending nearly free, the temptation is to send much more. Mailbox providers measure complaints and engagement, and enforce strict limits (see our cold email infrastructure guide). Autonomous systems that optimise for volume can burn through sending domains quickly.
Plausible isn’t the same as relevant. Models are good at writing messages that sound personal. Buyers have become good at spotting them. A message that mentions a podcast the prospect appeared on, but connects it to nothing that matters to them, reads as automated.
Errors are public. A wrong fact about a prospect’s company, a mangled name, or a reply that misreads an objection goes straight to a potential customer. One bad message can close a door that a person would have kept open.
Conversations need context. Live negotiations, objections and multi-person buying committees require judgement and accountability that buyers expect from a person.
The model that works: AI prepares, people decide
The setups we see performing best put AI on preparation and keep a person on judgement:
- AI finds and prioritises accounts based on fit and signals.
- AI researches each account and writes a brief with a suggested angle.
- AI drafts the first message, grounded only in facts from the brief.
- A person reviews, edits and sends, usually in a minute or two per message.
- AI sorts replies and prepares the follow-up; a person handles real conversations.
In this model one SDR can work far more accounts, with better research than they could manage alone, and every message that reaches a buyer has been approved by someone accountable for it. That’s the principle behind human-in-the-loop design.
How to measure success
Messages sent per day is a vanity metric with AI, because it’s nearly free to increase. Measure instead:
- Positive replies and meetings per hundred messages. Is quality going up?
- Pipeline created per SDR. Is the team more productive?
- Bounce and complaint rates. Is the infrastructure healthy?
- Time spent per account. Is preparation time going down while quality holds?
Our advice
Don’t ask “can AI replace our SDRs?”. Ask “which hours of our SDRs’ week could AI take over, so they spend more time on conversations?” The first question leads to fragile systems and damaged domains. The second leads to a team that books more meetings with less effort.
We build AI-assisted outbound systems: research, signals, drafting and reply handling, with people in control of what’s sent. Book a call to see how it would fit your team.


