An AI SDR agent for cold outreach does one of two jobs. One researches — it finds prospects, enriches each record, and drafts personalized outbound for a rep to review and send. The other runs the whole cycle — it prospects, writes, sends across email and LinkedIn, handles replies, and books the meeting, on its own. The first augments a rep; the second replaces the busywork and takes over the send.
Two jobs hiding in "let an agent do my cold outreach"
"Cold outreach" sounds like one task, but it splits the moment you ask who hits send? One version stops at a prepared draft: it does the research and personalization a human can't scale, then hands the rep a ready-to-review message. The other version owns the whole motion, sending in your company's name to people who never asked to hear from you. The first is a research analyst with a copywriting habit; the second is a full SDR. It's the same silent-filter-vs-agent-that-talks-back split we drew for lead qualification — except lead qualification works inbound leads who raised a hand, and cold outreach starts the conversation outbound with someone who didn't. That raises the stakes on the send, not lowers them.
Job 1 — Research and draft (the safe half)
This is where most teams should start, because a wrong draft costs a rep 30 seconds to fix, not a burned domain or a restricted account. The agent does the unglamorous, unscalable research and leaves the send to a human:
- Enrichment as infrastructure. Clay is explicitly GTM data infrastructure, not an AI SDR: its waterfall enrichment queries a priority sequence of 150+ data providers, falling back to the next source when one returns no match, to build a complete record. Its built-in research agent, Claygent, reads the open web to answer a prompt about each prospect. The output feeds a rep's outreach — it doesn't send anything.
- Custom draft agents. No-code builders like Lindy assemble an agent that researches a lead and drafts a first-touch message you approve, the same enrich-before-you-judge move a meeting-prep agent makes before a call.
- The desk-level version. The capture-don't-commit discipline behind our Gmail triage agent in Issue #001 is the same instinct: let the agent do the reading and the drafting, keep the human on anything that leaves the building.
The output of Job 1 flows naturally into the CRM capture layer, so the research the agent did is logged against the record instead of lost.
Job 2 — Prospect and send (the autonomous digital worker)
When you want the agent to send — not draft, send — you've crossed from analyst to autonomous SDR, and a misfire is now a real message to a real stranger in your name. Two vendors define this end of the market:
- 11x positions Alice as "the first true digital worker in sales": an agent that runs the full outbound cycle — prospecting, research, personalization, and multi-channel outreach across email, LinkedIn, SMS, WhatsApp, and voice — through to booking the meeting, positioned as executing the job rather than assisting it.
- Artisan's Ava is billed as an autonomous AI BDR that identifies and researches prospects, writes personalized outreach, sends multi-channel sequences, handles replies, and books meetings directly on your calendar. HubSpot-native tools like AiSDR sit here too, personalizing email and LinkedIn outreach from CRM data.
Worth noting where the honest vendors draw their own line: Artisan's product page says plainly, "Ava doesn't replace your reps, she replaces their busywork — she books the meetings and your reps close them." Even the autonomous camp keeps a human on the deal itself.
What keeps a human on the send
The send is the irreversible step, and cold outreach at volume runs into two hard constraints that a draft never touches:
- Deliverability. Google's email sender guidelines — in force since February 1, 2024, with enforcement ramping through late 2025 — require bulk senders (5,000+ messages a day to personal Gmail) to authenticate with SPF, DKIM, and DMARC, offer one-click unsubscribe, and keep the spam-complaint rate below 0.10%, never reaching 0.30%. Cross that line and your messages get routed to spam or rejected — for the whole domain, not just the bad campaign. An autonomous sender optimizing for volume is optimizing against exactly this number.
- Platform rules. LinkedIn's policy on prohibited software and extensions bans using bots or automated means to send messages, add contacts, or scrape the platform — and warns that accounts doing so can be restricted or shut down. "Autonomous LinkedIn outreach" and LinkedIn's terms are in direct tension.
A burned sending domain and a restricted LinkedIn account are slow and painful to undo. That's the textbook case for a human-in-the-loop gate: let the agent research and draft freely, but put a hand on the send until the deliverability numbers and the message quality have earned your trust — then widen the autonomy one segment at a time.
Which do you actually need?
If your pain is "my reps spend all day researching instead of talking to people," that's Job 1 — an agent that enriches and drafts, with a human still sending. If your pain is "we don't have enough pipeline and can't hire SDRs fast enough," Job 2 is the pitch, but start it narrow: one well-warmed domain, a tight list, and a human reviewing sends until the numbers hold. Most teams get more durable value from Job 1 than they expect, because the research is the part humans genuinely can't scale and the send is the part they shouldn't fully hand over. New to the idea? Start at Agent 101, or see the no-code AI agent tools that build the research-and-draft layer.
FAQ
What AI agent works best for cold outbound sales? It depends on which job you want done. For research and drafting that a rep still sends, Clay (GTM data enrichment across 150+ sources) plus a draft agent like Lindy is the safe, high-leverage start. For autonomous prospecting and sending, 11x's Alice and Artisan's Ava run the full cycle. Start with the research layer and keep a human on the send until deliverability and message quality hold.
Should I let an AI agent send cold emails automatically? Cautiously, and not at full volume on day one. Google's sender guidelines require bulk senders to keep spam complaints under 0.10% (never 0.30%), and crossing that routes your whole domain to spam. Gate the send behind human review while you build a sending reputation, then automate the segments where a misfire is cheap.
Can an AI SDR replace my sales reps? Not the closing, by the vendors' own framing. Artisan says plainly that Ava "replaces their busywork" and books meetings while "your reps close them." The durable pattern in 2026 is augmentation — the agent does research and drafting at a scale humans can't, and a rep owns the relationship and the deal.
Is automated LinkedIn outreach allowed? It's against the platform's rules. LinkedIn prohibits using bots or automated tools to send messages, add contacts, or scrape data, and warns that accounts doing so can be restricted or banned. Treat LinkedIn as a channel where a human stays in the loop, not one you fully automate.
Do I need code to build a cold-outreach agent? No. The research-and-draft layer is buildable with no-code tools — Clay for enrichment and a builder like Lindy for the draft — and the autonomous platforms are configured, not coded. See the no-code AI agent tools guide for the trade-offs.
Every agent in this series is a real build, with the part that stays human called out plainly. Want the one professionals actually run each week — the tools, the prompts, and where they keep a hand on the send? Subscribe free and get each week's build in your inbox.