Every founder we talk to has tried an 'AI SDR'. The story is always the same: they paid $400 a month, the tool sent 2,000 messages, and they booked zero meetings. Worse, three of their best prospects unsubscribed in disgust. The conclusion they jumped to was 'AI outreach doesn't work'. The actual conclusion should have been 'spray-and-pray outreach doesn't work, and AI just makes it faster'.
The pattern that does work
Our recipe has three parts and we won't ship an agent without all three. First, a tight ICP filter that runs before any outreach happens — usually a combination of firmographics, technographics, and a recent-trigger signal like a funding round, a new hire, or a job posting. If a prospect doesn't clear the filter, the agent never even drafts a message.
Second, real context per message. The agent reads the prospect's last three posts, pulls recent news about their company, and checks for a mutual connection. The message references something specific that the prospect actually said or did in the last two weeks. No 'I noticed you're the VP of Sales at Acme' openers — those die instantly.
Third, a fast handoff to a human at the first reply. The agent never tries to book the meeting itself. The moment a prospect replies, a real account executive takes over the thread and the agent's role is done. This is the part most tools skip and it's the part that makes everything else work.
What the numbers look like
Across the last six deployments we ran with this pattern, the average was 14 qualified meetings booked per agent per month, at roughly one-eighth the fully-loaded cost of a junior SDR. The reply rate on first-touch messages was 11% versus an industry benchmark of 1.5% for generic AI outreach. The unsubscribe rate was 0.4%, which is lower than most human-written sequences.
What we deliberately don't automate
We never automate the second message in a thread. We never automate connection requests on LinkedIn. We never automate replies once a human has taken over. We never let the agent invent product claims — every claim it can make is pulled from an approved facts file. These constraints feel limiting but they're what keeps the trust intact.
The architecture, briefly
Under the hood: a nightly job pulls signals from Apollo, Clay, and a few first-party sources, scores them against the ICP filter, and writes the survivors to a queue. A drafting step uses Gemini 3 Pro with a prompt that includes the prospect's recent activity and a strict 80-word limit. A human approves the first batch each morning (this step disappears once trust is established, usually after two weeks). The send step goes through the client's own mailbox via Smartlead so reputation stays with them, not us.
The takeaway
Agents win when they replace the boring 80% of the work — research, drafting, scheduling logistics — and let humans handle the moments that actually matter, which are the conversations. Anything else is noise, and noise costs you both money and reputation.