The question arrives in the first fifteen minutes of almost every call: can an agent replace an SDR? It is the wrong question, and answering it politely has cost us a great many hours. The agents we have built that are still running a year later share exactly one trait, and it is not autonomy. It is that none of them press send.
Sending was never the bottleneck.
Volume was solved around 2018. Sequencers, warm-up pools, multi-inbox rotation — a team of three can put fifty thousand emails into the world this month without a single AI feature, and most of the market has spent the years since discovering exactly how little that is worth. The bottleneck moved upstream and stayed there: knowing which two hundred of eight thousand accounts deserve a message this week, and having something true to say to each of them.
That work is reading. Reading job posts, pricing pages, changelogs, funding notes, the LinkedIn activity of four people. It is the part of the job every SDR does badly, not because they are bad at it but because there are four hundred accounts and eleven working hours. Reading at volume is the one thing a language model is genuinely, unambiguously good at. Pointing it at the sending step instead is like buying a research assistant and having them lick the envelopes.
The bottleneck was never how many emails you can send. It is how many accounts you can genuinely understand before Friday.
Give the agent a job description, not a personality.
The briefs we write for agents look like the briefs we would write for a junior analyst on their first week. Narrow, with a definition of done and a permitted way to fail:
Read the careers page, the pricing page, and the last three product announcements. Return up to three facts suggesting this company is building a revenue team this quarter, each with the URL it came from. If you find none, return NONE.
That last sentence is the whole trick. An agent allowed to return NONE is worth ten agents that always find something to say. Most disappointing AI outbound is not a model failure — it is a specification failure. The agent was told to personalise, so it personalised, and now four hundred prospects have received a compliment about a blog post nobody at the company remembers writing.
Every claim carries a source, or it does not ship.
A model that invents a funding round is not a quality problem. It is a brand liability, and it lands in the inbox of the exact person you were hoping to impress. So we enforce provenance at the field level: a fact without a URL never reaches the drafter. In practice that kills roughly a third of what the researcher produces, and it is the healthiest third of any pipeline we run — because what survives is checkable by the rep in eight seconds, which is the only reason reps ever trust an agent's output at all.
Evals are what separate a demo from a system
Fifty accounts, labelled by a human who knows the ICP. Every prompt change runs against that set before it goes near production, and it ships on two numbers: precision on should we reach out, and factual accuracy of cited claims. Without an eval set, an agent's quality is whatever the last person to read three outputs believes it to be, and prompts get rewritten at eleven at night on the strength of a vibe. With one, a prompt change is a pull request with a diff and a score.
A pipeline of narrow agents, not one big one.
# One job per agent · one output contract each. researcher.brief(account) -> facts[] + source_url | NONE qualifier.decide(brief) -> tier | reject(reason) drafter.compose(brief, offer) -> subject + body | NONE human.approve(draft) -> send crm.append(account, brief, decision, draft, ts)
Every arrow is a place to measure, to cache, and to stop. A single SDR agent that does all four steps inside one prompt is untestable: when replies dry up in week six, there is nothing to bisect. Split it, and the same collapse is a five-minute diagnosis — the researcher's source list went stale, the qualifier's tier threshold drifted, or the offer simply stopped being interesting.
The agents that survive are the boring ones. They read, they cite their sources, and they know how to say nothing.
Hugo Renault · Founding Partner, Mercator AI
Where the human stays.
The human owns the offer, the ICP, the approval, and every reply. The agent owns the reading, the drafting and the logging. This is not a transitional arrangement we expect to remove next year — it is the design. The reply is where the money is, and a reply is a conversation with a person who has just decided to spend some of their attention on you. Handing that to a bot to save four minutes is the most expensive saving in go-to-market.
Approval also has a second function nobody mentions: it is a continuous eval. A rep who rejects six drafts in a morning is telling you something precise about the research brief, and if you log the rejections you get a labelled dataset for free.
What it actually replaces.
Not the SDR. It replaces the four hours a day an SDR spends in browser tabs, and the silent sixty per cent of the list that never got researched because there was never time. The output is a team of five doing the research volume of a team of twenty, on a list where nothing goes untouched merely because it was tedious.
That is a smaller claim than the one on the billboards, and it is the one that survives contact with a real quarter.
Hire them like analysts. Review them like software.
Every agent we run has a job description, instructions living in a repository rather than a CRM note, an eval set, and a named human owner. It gets reviewed when it drifts and retired when its job disappears. Nothing about it feels like the future. It simply works on a Tuesday, and it still works in March.
How we scope an agent before writing a prompt
One job, one output contract. Researcher, qualifier, drafter, logger — each with a typed output and a permitted NONE. If it cannot be measured at the arrow, it does not get built.
Provenance or it does not ship. Every fact reaching a draft carries the URL it came from. Claims without a source are dropped, not softened — it costs about a third of the output and buys the reps' trust.
An eval set before production. Fifty human-labelled accounts. Prompt changes ship on precision and factual accuracy, never on how the output reads at eleven at night.
The human keeps send, and keeps the reply. Approval is the design, not a training-wheels phase — and every rejected draft is a labelled example you get for free.
END · NAMIA