Write the Outreach Message as a Filter, Not a Pitch

Write the three sentences the rest of the message hangs on

Three sentences carry an outreach message for an AI-heavy role. The first names the task the person would do with a model, in the vocabulary of the work rather than the tool. The second names what they would answer for when the model is wrong. The third names the evidence the process will ask for, and when.

Everything else in the message is recruiting furniture.

A concrete version, for a role that genuinely is AI-heavy: you would own the evaluation set that decides whether a retrieval change ships. You would sign the client-facing summary when the model got it wrong. Two sentences, and neither one sells.

Someone who does that work recognizes it at once. Someone who does not recognizes it too, which is the entire mechanism.

In August 2026, Indeed Hiring Lab put AI-related postings at 6.3% of its US job postings, past their prior peak of 3.3% in 2022 ⁴. The measure counts postings that mention AI, so it sizes employer vocabulary rather than the work behind it.

A subject line reading AI-heavy sits inside that vocabulary and separates the message from nothing.

Outbound is worth writing this way because it is not there for volume. Source outbound when you need evidence, not when you need more people is the premise underneath the message: the list is short, and it was built on purpose.

What terms actually narrow a sourcing list for AI skill is the step just before this one, and a message that ignores what the search found reads as a mail merge that got lucky.

Why does honest outreach for an AI-heavy role draw fewer replies?

Because a truthful description of the work tends to make the employer look less attractive, and the research on job previews points that way. Realistic job previews, the practice of telling candidates the unflattering parts before they commit, have been measured against turnover for decades.

The side effect matters more to a sourcing message than the headline: in the same analysis where previews lowered voluntary turnover, they also lowered how attractive the employer looked ¹.

In 2011 David Earnest, David Allen and Ronald Landis published a meta-analytic path analysis in Personnel Psychology covering 52 studies and roughly 17,000 people.

Getting a preview went with fewer people quitting on their own, at a correlation of -.07 ¹. Squared, that is about half of one percent of the variance in who leaves.

A truthful preview is a cheap adjustment and never a retention plan.

The attractiveness cost measured -.10 ¹. Ignoring the minus signs, that is the larger of the two numbers, and it lands at the front of the process, which is where outreach lives.

Four meta-analyses across 26 years land in one narrow band, every result small and every result pointing the same way: -.09 and -.06 in 1985, -.06 in 1998, -.07 in 2011 ¹.

Agreement across four separate reviews is worth more than any one of the numbers in it.

None of that research studied sourcing messages. Previews went to applicants and to new hires rather than to strangers on a contact list, and the outcome measured was who quit rather than who replied.

The direction carries. The size does not.

The replies lost are the part worth caring about. A message saying the role means checking model output against a source before anything ships will not appeal to somebody looking for a role built around writing prompts.

Losing that reply costs little, because the role was not the work that person wanted.

What survives the assumption that a model wrote the message?

One checkable sentence about the recipient's own work survives that assumption. A model drafts fluency, a compliment about the company and a paragraph on why the role is interesting, and a practiced reader discounts all three on sight. What no draft supplies is the reading behind a specific claim: the talk they gave, the pull request they merged, the postmortem they published.

Spend the message on that sentence and cut the rest.

One measurement of what happens when a signal becomes free to produce came from the other chair.

Freelancer.com launched a one-click cover-letter writer in April 2023. Jingyi Cui, Gabriel Dias and Justin Ye compared 5,499,707 bids around that launch and found the correlation between how closely a cover letter matched the job post and receiving a callback fell by 51% ².

The correlation with winning the job fell by 79% ². Weight moved onto a signal the tool could not produce: the correlation between callbacks and the platform's own past-work review score rose 5%.

The figures are before-and-after correlations on a freelance marketplace rather than causal estimates in corporate hiring, the paper has not been through review, and the people writing were bidders rather than recruiters.

Applying the finding to a sourcing message is an argument, not a measurement.

Timestamps in the same study measured how long anybody spent on a draft. Over 75% of the AI-generated letters were submitted within one minute of clicking generate, and about 5% at least five minutes later ².

Within a single worker, 3.5 more minutes of editing went with a higher chance of winning the job, an association the authors state plainly is not causal.

The candidate side of this exchange worked it out first. What to put in a cold message so it gets answered names the tell from that chair: a template reads as a mass send before the first line ends.

The tell applies to a recruiter's outreach without a single change.

Say what the process will ask for, in the first message

Say how many conversations there are, whether an exercise exists, roughly how long it takes, whether AI is allowed inside it, and who reads the result. Five short clauses. Somebody deciding whether to spend an hour on a first call is deciding against unknown future hours.

The outreach is the only place where those hours can be made known early enough to matter.

The cheapest refusal is the earliest one. Naming the assessment stage in the post moves drop-off earlier reports the same mechanic on the inbound side. A person who would refuse a two-hour assignment should refuse it in the first message, not in week three with a panel already scheduled.

Promise only an exercise somebody can grade. An AI-open exercise is expensive on the reviewer's side, and what an AI-open exercise costs your engineers to grade is the bill that arrives after the outreach goes out.

A stage announced in the first message and quietly dropped in week two costs more than never naming it.

A published case against writing any of this out comes from a trial on job adverts.

In 2022 the Behavioural Insights Team put 10,468 working-age participants in front of versions of a single advert. Willingness to apply was significantly greater for the vague high-level-skills version than for the version that spelled out required behaviours, and the specific version failed its own check: readers did not rate it any clearer ³.

The report calls that analysis secondary and says the manipulation failed, so it is not evidence that detail repels people.

The result supports something narrower. Adding detail did not by itself make the advert read as clearer, while inside the same trial, the clearer a participant found an advert, the more willing they were to apply ³.

Specific and clear are two different achievements.

A sentence about the work has to be readable by the person who does the work rather than by the person who wrote the requisition.

Read the draft in the send queue once more, with one question: which sentence could not have been sent to anybody else? If the answer is none, the message will be answered by the people who answer everything.

Cut the paragraph about the mission. Keep the sentence about the evaluation set.

See a sample report >