The candidates you most want to hire have inboxes that look like everyone else’s outreach. Three sentences of flattery, a pitch, a calendar link, a ghost. The reason their reply rate to recruiters is in the single digits is not that they are too busy. It is that the messages all look the same, and they have stopped reading them.
LinkedIn’s Future of Recruiting 2025 shows 37% of recruiting professionals using or experimenting with generative AI, with outreach as a top use case. The volume of templated messages going out has gone up sharply. Reply rates, in most teams’ data, have gone the other way. AI is unusually good at writing outreach that performs better than that. It is also unusually good at writing outreach that performs worse, because the easy thing it produces is exactly the thing candidates are now sick of.
This is a workflow for the first version, not the second.
Variable-depth personalisation, not “personalisation”
Most “personalised” outreach is a template with one variable plugged in: the candidate’s first name and a sentence about their current employer. Strong candidates clock this in two seconds.
Real personalisation has variable depth, by which I mean: different candidates get different amounts and types of personal context based on what is publicly visible about them. A senior engineer with three open-source repos and a conference talk gets a message that names a specific PR. A product lead with a thoughtful Substack gets a message that references a specific post. A passive candidate with a sparse profile but a strong work history gets a message that names a specific past project from their resume. The point is not to flatter. The point is to demonstrate that you actually looked.
Open ChatGPT or Claude, paste the candidate’s profile, and run a depth-detection prompt before you draft:
I am about to write outreach to this candidate for a [ROLE] at [COMPANY]. Here is their public profile and any artefacts I have linked: [PASTE]. Tell me which of these signals are strongest for personalising a first message: a specific past project, a public artefact (talk, blog, repo), a career pattern (recent move, sabbatical, founding), or a domain niche they are clearly known for. Identify the one strongest signal and explain in one sentence why a recruiter referencing it would feel different from a templated message. If none of the signals are strong, say so and recommend I send a tighter, less-personalised note rather than fake personalisation.
Then draft the message using only that one signal, plus a tight pitch and one specific question. A 90-word note built around a real artefact outperforms a 200-word note built around fake intimacy.
A sequence prompt that doesn’t sound like a sequence
Tools like Gem, hireEZ, and integrations inside Greenhouse and Lever all support sequenced outreach. Most teams design sequences as one persuasive note, then two reminder bumps. The bumps are where reply rates die, because the bumps are pure pressure.
Better sequences have a different angle in each touch. Try this draft prompt:
Design a 4-touch outreach sequence for a [ROLE] at [COMPANY]. The candidate is a [SHORT DESCRIPTION OF CANDIDATE TYPE]. The role’s biggest pull factors are [LIST 2 TO 3]. The likely objections are [LIST 2]. Each touch should have a different angle and tone, not just a reminder. Touch 1: opening, anchored on [ONE SPECIFIC SIGNAL]. Touch 2: content-led, sharing [SOMETHING USEFUL TO THE CANDIDATE EVEN IF THEY DON’T REPLY], such as a relevant article, a public team blog post, or a question worth thinking about. Touch 3: addresses one likely objection directly. Touch 4: a short, polite “closing the loop” message that gives them a clean way to decline without feeling rude. For each touch, give me the subject line, the message body under 120 words, and the days-since-previous timing.
The middle two touches are the ones that change reply rates. A candidate who ignored the first message will sometimes engage with a “here’s a thing I thought you’d find interesting” follow-up, especially if the thing is genuinely useful and not a thinly disguised pitch.
The second-touch trap
Here is the failure mode that quietly destroys outreach metrics. Most recruiters write a thoughtful first message, then automate three identical-feeling bumps. The candidate who was on the fence after message one gets pushed off by the third reminder. Reply rate on touch one was 8%. Cumulative reply rate after the sequence is 9.2%. The bumps barely added anything, and they cost goodwill from the candidates who decided the brand felt pushy.
The fix is to count “no reply” as data, not as a prompt to send the same message again. After touch two, ask:
I sent the following two messages to a [CANDIDATE TYPE] for [ROLE]: [PASTE BOTH]. They have not replied. Without changing the underlying pitch, what is the most likely reason a strong candidate would have read these and decided not to reply? Suggest a third message that addresses that specific reason without being defensive or pushy. Keep it under 80 words.
Sometimes the answer is that the role is not actually a fit and the right move is to step back. Sometimes it is that the comp range was unclear, or the message felt too senior or too junior for the candidate’s level. AI is good at naming the most likely objection if you ask it the question rather than asking it for “another follow-up”.
Reply-handling, where the time really hides
The hidden time-sink in outreach is not the first message. It is the dozen replies you get that need a thoughtful, customised response: scheduling, salary questions, “I’m not looking but I have a friend”, “I’d want to know more about the team”, polite declines that deserve a polite “stay in touch” without sounding canned.
A reply-handling prompt that scales without flattening:
Below is a reply from a candidate to my outreach for [ROLE] at [COMPANY]. [PASTE REPLY]. Draft a response that: directly addresses their main point, matches the warmth and length of their message (don’t be longer or chillier than they were), gives them one clear next step, and leaves the door open for a future conversation if this role isn’t right. Do not use the words “circle back”, “touch base”, or “combined effect”. Keep it under 100 words.
The “match their tone” instruction is the one that matters. Strong candidates write short, direct replies. A 200-word warm reply to a 30-word direct message reads as desperate. AI gets this right when you tell it to match length, and gets it wrong by default.
The failure mode to watch for
The AI tell. Outreach written without enough context produces phrases like “I came across your profile and was impressed”, “your background really stood out”, and lists of generic adjectives (“strategic, innovative, results-driven”). Candidates know what these mean now. Before you send, run the message through a quick check:
Read this outreach message as a senior candidate who receives 5 cold pitches a week. Which sentence, if any, sounds like it was generated by AI rather than written specifically for me? Flag the exact phrase, explain why it reads as templated, and suggest a sharper alternative.
Two passes through that filter, and the messages start landing in the inbox of people who have stopped reading the others.
The honest version of “easy”
The work is not exactly easy. It is faster, and the messages are better, and the time you used to spend on the first 30 messages now happens in the first 8. The other 22 are time you reinvest in the candidates who replied. That is what “easy” actually means here, and it is the version of AI-assisted outreach that survives contact with reply data.