AI SDR Agent

Sending at volume was solved a decade ago, and the sequencing tools that solved it are cheap and good. Capacity was never the constraint. The constraint is that a message which could have been sent to four hundred companies reads to every recipient exactly like a message that was, and no amount of additional sending improves that.

Relevance is a research problem, not a writing one. Knowing that an account has just opened a second site, or is advertising for the role a product makes unnecessary, or has publicly committed to a platform the product connects to, means somebody read about that company. That reading is the part that does not scale by adding people, and it is the part most sales automation quietly skips.

Which is why an AI SDR that only drafts messages is the same sequencer with better grammar. The interesting question is not whether software can write the email. It is whether the email had anything true and particular to say, and that is decided several steps earlier.

Discovery assembles the account set from a definition of who is worth talking to: industry and size, the technology a company already runs, a hiring pattern, a funding event, a change of premises. The definition is the business's own and it is explicit, because an account set nobody can inspect is a source of results nobody can explain.

A research agent then reads each account — the site, recent announcements, job advertisements, whatever public sources the business considers meaningful. What it returns is not a summary. It is the specific facts a message could be built on, each with the source it came from, so a person reviewing the draft can check the claim rather than trust it.

Contact enrichment finds the individual and a deliverable address, and scoring ranks what is left, so that effort goes to the accounts where the research found something rather than to whatever the discovery step happened to return first.

Drafting works from those facts rather than from a template with a variable dropped into it. The difference shows up in the reply rate and in nothing else, which is why the research is where the cost belongs. Everything lands in the CRM as it happens, and where the business wants it, a person approves before anything is sent.

Prospect sources and enrichment providers supply the raw account and contact set; headless browsing handles the public reading that no provider sells. A model does the research comprehension and the drafting. Orchestration holds the state of each account, retries the lookups that fail, and keeps the sequence from starting again on a company already in conversation.

The sending itself runs through the outreach platform the team already uses. Replacing it buys nothing, and the sender reputation it has accumulated over years does not transfer to anything new. The CRM stays the system of record throughout, rather than becoming somewhere results are copied to afterwards.

The approval gate is where a business decides how much of its own voice it is prepared to hand over, and it is a setting rather than a stage of maturity. Some teams review every message for a quarter and then stop; some review the first message to any account and never the follow-ups. Both are defensible. What is not defensible is a system with no gate available.

Suppression has to be checked before writing rather than after a complaint. An existing customer, an account with an open opportunity, a company a colleague is already talking to, somebody who asked not to be contacted again — all of that is in the CRM, and a system that reads it only at the point of sending has already drafted the message that should not exist.

Sending volume is held against what the domain can carry, because the fastest way to lose an outbound channel is to use it as though reputation were unlimited. And every account carries its record: what was found, what was written, what was sent, what came back, and which accounts the system passed over.

It does not close anything. Every part of this stops at a conversation with somebody who has shown interest, and the qualification it performs is a filter rather than a judgement. Pipeline gets handed to people, along with the research that produced it.

It does not repair a wrong definition of the customer. Applied to the wrong account set, all this achieves is well-researched irrelevance delivered faster than before, and the system will not notice — it has no way to know that the accounts it was pointed at were never going to buy.

And it does not place the business outside the rules that govern contacting people. Consent, identifying who is writing, and an unsubscribe that works are obligations that do not soften because the sender is software. A system that sends more is a system that can breach them at a rate a person never could.