Ranking people and writing to one of them

The first job is ordering: given a list, which names should be worked first. It is a scoring problem, and it is reasonable to ask a model to help with it. The second job is composition: given one person and one reason, what is the sentence that makes the message worth reading. A score can inform the second job, but it cannot perform it.

The quoted failure is a composition failure. A tool that writes the same email to everyone has not failed to rank; it has failed to notice that the recipients differ. Hiring more SDRs would not have fixed that either, which is why the comparison with headcount is partly a distraction.

The swap test for “generic”

Generic has a simple test: replace the company name, the role and one product reference with another plausible company's, and see whether the message still reads as if it were written for someone. If it does, the message was assembled from a template and a score, not from anything the sender knew about the reader.

That is why personalization tokens do not rescue a generic message. The reader can tell the difference between a sentence that contains their name and a sentence that could only have been written to them. The second kind requires a fact: something the reader said, shipped, asked or complained about.

What a score cannot supply

Fit, intent and likelihood are rankings of people against a model of the product. They compress a lot of context into a number, and the number is useful for deciding where to spend the next hour. It contains no sentence, no detail and no reason that would survive being read by the person it describes.

An agent asked to write from a score alone has to fill the gap with the product description, which is exactly the material every other recipient gets. That is the mechanism behind the sameness: not laziness, but the absence of anything recipient-specific in the input.

Where WarmQuant sits

WarmQuant is not a lead-scoring or list-prioritization tool, and it does not send anything. It works on one discussion at a time: the writer supplies the thread text, the confirmed facts about the product, one real limitation, their relationship to the product and the goal of the reply, and receives a private English draft to review and send by hand.

The unit of work is deliberately one person and one reason. If the supplied discussion does not support an honest reply, the tool can return a short reason to skip instead of producing a message. That is the opposite of throughput, and it is the only way the sentence can contain something the score never had.

What this note does not prove

The post names no vendor, gives no numbers and does not say how the trial was configured. A tool that produced identical emails may have been used with a thin brief, or with no per-recipient input at all; the post does not settle which.

Nothing here shows that AI outbound cannot work, or that scoring is useless. The narrower claim is that the two jobs fail differently, and that a team evaluating a tool should ask which job it is actually doing before comparing it with an SDR.

How this note was made

This field note follows the workflow WarmQuant supports: start from a public discussion, keep the source attached, separate what the thread shows from what a supplier claims, and prepare the lesson for review. It was written for WarmQuant’s blog and checked against the cited sources before publication. It is not a customer testimonial, independent research or proof of demand.

Sources & further reading

Read the original before replying. Sources can change or be removed. An excerpt is not the full thread, and a public post is not verified demand.