What an AI SDR agent is, and what it is not

AI SDR agent

AI SDR Agent: An AI SDR agent is software that runs sales development work — prospecting, research, outbound messaging and reply handling — end to end, without a human approving each step.

The load-bearing word is agent, not AI. Anthropic's December 2024 write-up draws the line where it belongs: an agent is a model directing its own tool use in a loop, with a stopping condition, rather than a fixed path a developer wired in advance. Apply that to sales development. You get a system that picks which account to research next, decides whether the research justifies a message, writes it, sends it through a mail API, reads the reply, and chooses between booking, nurturing and dropping. Most products sold under the label do two of those six.

Where the term came from

"SDR" predates the AI part by over a decade. Aaron Ross's Predictable Revenue popularised splitting the sales role so that one group prospects and another closes, which is why "sales development rep" became a seat you could hire for and therefore a seat vendors could promise to replace.

The compound "AI SDR" arrived as a product category in 2023. Regie.ai moved from AI sales copy toward agent framing, AiSDR took the category name as its brand, and 11x and Artisan sold "digital workers" with human first names — Artisan's is Ava. Salesforce made it enterprise vocabulary in September 2024 when it announced Agentforce at Dreamforce with a packaged SDR Agent. Eighteen months, pitch deck to SKU. That is about how long a category name lasts before it stops meaning anything precise.

What it is not

The neighbour people confuse it with is the sales engagement platform — Outreach, Salesloft, Apollo sequences. Adding an LLM that writes the body copy for step 2 does not convert a sequence into an agent. The sequence still decides what happens next, and it decided before the lead existed.

Thing	Who decides the next action	What it sends	Typical failure
Email sequence	You, at setup	Fixed steps on a calendar	A bad list, sent faithfully
Site chatbot	Intent script	Nothing outbound	Dead-ends when off-script
AI copy assistant	Human, every time	Nothing	Fluent email to the wrong person
AI SDR agent	The model, per lead, at runtime	Mail API or sequencer, via tool call	Unbounded loop; confidently wrong research

The two things people mean

"AI SDR agent" names two structurally different products. They fail differently.

Type A — the replacement seat. A vendor-operated agent with its own name, dashboard and sending infrastructure, priced against headcount or per meeting booked. 11x and Artisan are the clearest examples of the framing. You hand over an ICP and a domain. The loop runs inside their system, so your audit trail is whatever their UI shows you — which is the part that will hurt the first time a prospect forwards you an email you cannot explain.

Type B — the scoped component. A tool-calling loop you run yourself that does one SDR job inside a stack you already own: qualify inbound form fills, enrich and route, research an account before a human writes. Tools are exposed through function calling or something like Model Context Protocol, so tool selection happens at runtime rather than being hard-wired in order. Scope is narrow on purpose. Published walkthroughs of this pattern — qualifying inbound leads from a CSV, for instance — share one trait: the loop is allowed to do exactly one job.

Type A sells you an outcome and owns the loop. Type B gives you the loop and leaves the outcome to you. Type B is the form that keeps the decision log inside a system you control. A vendor that says "agent" and means "our sequencer now writes the subject line" is neither.

How to tell whether something qualifies

Four tests. All four, not three.

Can it change the plan mid-run? If step 3 is always email 3 regardless of what step 2 learned, it is a sequence with a language model inside it.
Does it choose its tools? Runtime selection from a tool registry counts. A pipeline where enrichment always precedes scoring, in that order, does not.
Is the stopping condition enforced outside the model? Unbounded loops are the most common production failure mode for agents, and the fix is a hard step cap and a spend cap in the orchestrator — not an instruction in the prompt asking the model to be reasonable. The case for capped agents comes back to this.
Is there a row per decision? If you cannot reconstruct why lead 4,312 got that email, you cannot manage the thing. You can only hope.

Send authority is the authority to destroy your domain. A Google Workspace account caps at 2,000 messages per day, and anything over 5,000 messages a day to Gmail addresses puts you in Google's bulk sender rules: SPF, DKIM, DMARC, one-click unsubscribe, and a spam complaint rate held under 0.3%. An agent that decides its own send volume can cross that line in an afternoon, and CAN-SPAM applies to the output whether a human or a model wrote it.

I built Metadata.io from $0 to $15M ARR selling automation to demand gen teams, and I now run agents in production on my own projects. Spend and send are the two authorities I hold back longest. Research, drafting, scoring, routing, reply triage — delegate those. The irreversible actions stay behind a cap, and some of them stay human.