The most expensive leak in outbound is not a bad list or weak copy. It is a positive reply that sits unanswered. Most GTM automation stops at the first message: find the lead, research it, write, send. Then the prospect says yes, and the reply waits in an inbox until an SDR gets to it. An agent can close that gap: read every response, classify it, draft the answer, and put it in front of a human to approve in seconds.
Johnathon Kidd, technical lead at Growth Alliance, built exactly this and ran it live on the room at our September event in San Francisco. Everyone in the audience got an email, replied however they liked, and watched the drafted answers come back.
Where the deal dies #
The story is familiar to anyone who has run campaigns. A campaign gets a genuinely positive response. You check a few minutes later and the SDR has not answered. You ping them and move on. An hour later it is still sitting there.
Nobody decided to ignore it. The SDR is working other accounts, the reply landed in a shared inbox, or it came in on LinkedIn while they were on email. At a million emails a month, SDR attention is the bottleneck, and it is spent on the wrong end of the funnel: triaging replies instead of working the conversations that need a person.
The loop, end to end #
Growth Alliance’s system runs in three stages, all on Cargo:
- Leads in. Leads are sourced automatically or pulled from an existing list into a model.
- First touch. A play researches each person, writes the message, and sends it by email or LinkedIn. Then it waits.
- The reply. When a response arrives, the play wakes up, classifies it, and, if it is positive or neutral, drafts an answer. It keeps looping: if the prospect replies again, the cycle runs again.
Two agents do most of the work, and each has one narrow job:
- The research agent turns web search results into one short, factual sentence about the person or their company. One sentence, not a dossier, because that is all a good reply needs.
- The writer agent writes the message from that sentence and the conversation so far.
Classification runs as a seven-step process before anything is drafted. That ordering matters: an agent that drafts first and classifies second will cheerfully write a warm reply to an unsubscribe request.
Johnathon built the whole workflow with the Cargo CLI and Claude, from a single prompt, with minor adjustments along the way.
Keep the human, for now #
In the demo, every drafted reply went to a Slack channel for approval before it was sent. Johnathon’s roadmap is to remove that step, then add calendar booking to the reply, then let the copy improve itself as it sends more.
Our view: remove the approval step one reply class at a time, and only once the classifier has earned it.
- Start with everything approved by a human in Slack. Approving a correct draft takes seconds. The reply still goes out in minutes instead of hours, which is most of the win.
- Track agreement. For each class (interested, wants a meeting, asks a question, not now, wrong person, unsubscribe), log how often the human sends the draft unchanged.
- Graduate the safe classes first. “Send me a time” with a booking link is low risk once it has been approved unchanged a few hundred times. Pricing questions and objections are not.
- Never automate the negative classes. Unsubscribes and complaints go to suppression and a human, not to a writer agent.
The same logic applies to every agent that reaches a person: approval is priced by blast radius, and it is withdrawn class by class as evidence accumulates, not all at once because the demo worked.
What this frees the SDR to do #
The point is not fewer SDRs. It is SDRs spending time where a human is actually needed: the conversation that has turned into a real opportunity, the objection that needs judgment, the account worth calling. The agent handles speed. The person handles nuance.
A note on voice #
Someone asked about AI voice agents for cold calling. Johnathon’s answer was that they are fun to build and the hard part is legal: in the US, AI-voiced calls need the recipient’s explicit prior consent, which is a conversation for your legal team before it is one for your GTM engineer. Email and LinkedIn replies to someone who wrote to you first are a much simpler place to start.
For measuring whether the agents are getting better, see LLM evals for revenue agents.
FAQ #
Yes. An agent can read replies across email and LinkedIn, classify each one, and draft an answer using research on the person and the conversation so far. Start with a human approving every draft, and automate reply types only once the drafts are consistently sent unchanged.
Unsubscribe requests, complaints and anything negative. Those should update suppression lists and go to a person. The first replies to automate are low risk ones, such as a prospect asking for a meeting time.
Because the class decides whether to respond at all and how. Drafting first risks writing an enthusiastic answer to an unsubscribe or a complaint. Classification first routes each reply to the right path.
No. It removes the triage work so SDRs spend their time on conversations that need human judgment, such as real opportunities, objections and accounts worth a call.
AurelienSept 30, 2026