Most rejection emails go unsent because they're hard to write well and risky to write badly. Here's how to sort them by stage, hand the drafting to AI, and keep the parts that need a person.
Sort candidates by how much of their time you took, then write one rejection per stage instead of one for everyone. AI is good at the wrapper and the tone. It’s poor at knowing which facts about a person are true, so feed it one true detail, tell it not to give reasons, and have a person check every sentence before it goes. Save the sent email with the decision record, and treat any message tied to a background check as a legal script, not a draft.
Say you’re a talent acquisition coordinator at a 400-person logistics company. A warehouse operations supervisor role was posted six weeks ago, and on Tuesday the hiring manager’s chosen candidate accepted the offer. In the applicant tracking system, 140 people are still marked “under review.” Four of them sat through a three-round process, and one spent a weekend on a case exercise.
You have a decline template from 2019 that begins “Thank you for your interest” and ends “we have decided to move forward with other candidates.” Sending it to all 140 feels wrong for the finalists. Writing 140 individual notes isn’t going to happen. So the list sits there through Friday, and then through the next Friday, and the silence does the rejecting.
The way out is to notice that “rejection email” names four different jobs. The person who filled in a form needs a short, clear close. The person who did a two-hour panel needs a message that shows someone remembers them. Those are different emails, and once you sort by stage, AI can draft each one quickly without you inventing 140 sentences from nothing. What it can’t do is know what’s true about any one person, which is why this guide splits the work.
Silence is the usual alternative. In a December 2024 survey by the recruiting software company Greenhouse, 61% of the 2,500 workers questioned in the US, UK and Germany said they’d been ghosted after a job interview.[1] It’s a vendor survey of self-reported experience, so read it as a sign of how common the silence feels to candidates rather than as a precise rate. The practical point is that a clear, timely no often costs you less than the gap where nothing arrives.
Below you’ll find the four-stage ladder, a prompt you can paste, the edits that keep a draft safe, a batch routine, and a five-question check.
Start by asking how much of the candidate’s time you took. That single question decides how long the message is, who should send it, and how much of it has to be specific to the person. Here’s the ladder, filled in for the warehouse supervisor role above.
| Stage reached | What they gave you | What the message has to do | Who sends it, and how long | What AI drafts |
|---|---|---|---|---|
| 1. Applied only | Ten minutes on a form and a CV | Close the loop, name the role, say nothing about their qualifications | Coordinator, from the tracking system. Three or four sentences | The wrapper, once, in your company’s voice |
| 2. Recruiter screen | A 30-minute call and some prep | Thank them for the conversation, refer to one real thing they said, state the decision | The recruiter who spoke to them. Four to six sentences | The wrapper with a slot for one detail from the call notes |
| 3. Interviewed with the team | Several hours across two or three rounds | Name the stage they reached, give one specific strength, state the decision, offer to stay in touch only if you mean it | Recruiter, with the hiring manager copied. Six to eight sentences | A draft from three facts you supply |
| 4. Finalist | Days of preparation, maybe time off work, references | A phone call first, then a short written follow-up with the decision and an offer of feedback if you’ll really give it | The hiring manager makes the call, the recruiter or manager sends the email | The follow-up email and a short call outline. Not the call |
A worked example for a warehouse operations supervisor role. Change the senders and lengths to match how your team is organised.
Different roles in the hiring process feel this differently:
The more of someone’s time you took, the more of yours the rejection has to cost.
That rule also tells you where AI saves the most. Stage 1 is almost entirely wrapper, so AI can take nearly all of it. Stage 4 is almost entirely judgment and relationship, so AI gets the smallest role there. The ladder stops you from spending human attention on the 120 short messages and withholding it from the four that matter.
The mistake I’d expect most people to make here is pasting the candidate’s CV and the interview notes into a chatbot and asking for “a kind rejection.” The output will be fluent, and it will also contain sentences about the candidate that nobody verified. AI fills silence with plausible detail, and in a rejection email a plausible detail is a liability.
So give it a short brief and keep everything else out. Before you paste anything, check your organization’s rules for candidate data. Our guide to using AI at work without leaking confidential company data covers what’s safe to put into which tool.
Here is that brief as a prompt, filled in for a stage 3 candidate.
| I’m a recruiter at a 400-person logistics company. Write a rejection email in plain, warm English. Keep it under 120 words, use no exclamation marks, and don’t open with “Unfortunately.” Candidate first name: Dana. Role: Warehouse Operations Supervisor. Stage reached: interviewed with the team, two rounds. Decision: we have offered the role to another candidate, and this is final. One true strength to mention, from my notes: “Dana’s walk-through of how she rebuilt the night-shift handover was the clearest answer our panel heard.” Do not give any reason for the decision. Do not compare Dana with other candidates. Do not mention age, health, family, availability or anything that isn’t in this message. Do not promise future roles or say we’ll keep her details on file. Do not invent any detail about her. Give me a subject line and the email. Then list every sentence that states a fact about Dana, so I can check each one against my notes. |
Swap in your own company, role, stage and detail. The last paragraph is the line that makes the draft checkable.
Give AI the stage, the decision and one true detail, and nothing else about the person.
Holding back also protects you. A model that has read the whole interview file can repeat things that shouldn’t appear in a message to the candidate, such as a note about a career gap or a remark about their accent. If it didn’t see the note, it can’t paraphrase it. For stage 1 you don’t need the person’s name in the prompt at all, since your tracking system fills it in.
For stages 2 and 3, ask for two or three variations and pick the closest to how you write. Then read it aloud. If you’d wince saying a sentence to the candidate’s face, cut it. Save it beside your other HR prompts and reuse it each hiring round.
Candidates often want to know why, and recruiters often feel they owe them something. That pull is where drafts go wrong. A model asked to be kind will reach for a reason, and kind-sounding reasons tend to be vague (“we felt the fit wasn’t quite right”), subjective (“your energy level”) or unverifiable (“someone with more experience”). Each of these is a claim about the decision that you’d have to defend later.
The fix is to decide about reasons before the AI gets involved. For stages 1 and 2, give none. For stage 3, give a reason only if it ties to a written requirement of the role and you’d give the same reason to any candidate who’d been in the same position. For stage 4, the hiring manager decides in the call what to say, and the email repeats only that. The table below shows the edits I’d make to a first draft for Dana’s email.
| A first draft said | The recruiter changes it to | Why |
|---|---|---|
| “We felt your energy in the panel was lower than we hoped.” | Cut entirely. | A subjective impression with no link to the job requirements. It’s hard to defend and likely to hurt. |
| “We chose a candidate with more experience in our systems.” | “The role needed hands-on experience with our warehouse management system, and the person we chose had more of it.” | Keep a reason only if it matches a written requirement and is true. Then say it in that plain, specific form. |
| “We’ll keep your details on file for future roles.” | Cut, unless you’ll actually do it. If so, name the team and the time frame. | A promise the AI invented. Candidates tend to remember promises. |
| “Your walk-through of how you rebuilt the night-shift handover was the clearest our panel heard.” | Keep as written. | It came from the recruiter’s own notes, so it can be checked and it’s specific to Dana. |
| “We wish you every success in your job search.” | Keep once, at the end. | Generic, but harmless, and it closes the message. |
An illustration built on the prompt above. The first-draft lines are the kind a model typically produces unprompted, not quotations from a specific tool.
AI can phrase a reason. Only someone who was part of the decision can own it.
Two legal points belong here, and neither is a reason to freeze. First, US federal rules say employers covered by them must keep application forms and other records having to do with hiring for one year from the date of the record or the personnel action, whichever is later, and longer once a charge has been filed.[2] So save what you sent, and file it with the decision record, rather than leaving it in a chat window. Second, if a background report influenced the decision, the notice isn’t something to improvise with AI. The EEOC’s guidance on background checks describes a notice with a copy of the report and a summary of rights before you act, and a further notice afterwards.[3] Use the wording your counsel or your screening vendor gives you.
Rules on automated screening tools are changing quickly in several US states and in the EU. I haven’t put specifics here because I couldn’t confirm a current, stable version of those rules from primary sources, and this isn’t legal advice. If AI scored or ranked candidates before you got to the rejection stage, ask your employment counsel whether extra notice or explanation duties apply where you hire. For the reasons behind the decision itself, our guides to writing a performance improvement plan with AI and to disciplinary write-ups show the same pattern: the draft is fast, the facts are yours.
Back to the 140 on the list. Here’s the routine that clears it in an afternoon, with the human steps marked so they don’t get skipped.
The batch routine used in this guide. Stage 1 mostly runs through steps 1, 2 and 5 in bulk.
On a batch like this one, the stage 1 group of roughly 120 gets one wrapper, sent in bulk from the tracking system with names merged in. The roughly 14 at stage 2 each get the wrapper plus a sentence from the recruiter’s call notes. The handful at stage 3 get their own drafts. Most of the time goes to the four finalists, which is where it belongs.
Now the hardest one. Say you’re a hiring manager, and the runner-up for a finance systems analyst role spent a weekend on a case study before her final round. You call her Thursday at four. She asks what she could have done differently, and you have one true answer: the person you hired had already run a migration on the same platform. You say exactly that. After the call you paste your notes into an AI tool and ask for a short follow-up email, and the draft comes back with a line you didn’t say: “We’ll definitely keep you in mind for the senior analyst opening.” You delete it, because you don’t know whether that opening exists, and she’d write to you about it in March.
That deletion is the whole argument for the routine. The call was human and the email wrapped up the call, so the only sentences that survive are the ones you’d already said out loud. If you’re unsure whether a finalist should be told now or after the chosen candidate has accepted, ask the recruiter to set the order, and don’t let the AI decide timing on its own.
The split between AI and person needs to be written down rather than assumed. This is the one I’d put in a recruiting team’s shared notes.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Drafts the four stage wrappers in the company’s voice | Which stage each candidate reached, and the decision itself | The recruiter compares the stage column in the tracking system with the template used, weekly |
| Phrases the one true detail for stages 2 to 4 | Choosing the detail and confirming it’s accurate | The recruiter ties each detail to a line in their own interview notes before sending |
| Drafts a reason if the recruiter supplies one | Whether to give a reason at all, and what it is | Recruiter and hiring manager agree the wording; it has to match a written job requirement |
| Drafts a holding reply to “What could I have done better?” | Whether to answer, and what’s said on a call | An HR business partner reviews the holding reply once; every reply goes out under a named person |
| Nothing, for notices tied to a background check | Sending the legally required notices | Counsel or the screening vendor supplies the wording; HR confirms it was used |
A starting split for a recruiting team. Adjust the reviewers to your size, and keep the last row even if the team is one person.
Before any email goes out, run this check.
| # | Question | If the answer is no |
|---|---|---|
| 1 | Does every sentence about the candidate match my own notes? | Delete the sentence, or fix it from the notes. |
| 2 | If there’s a reason, is it one I’d give any candidate at this stage, and does it match the written job requirements? | Remove the reason. Say less. |
| 3 | Is every promise in it one I’m ready to keep? | Cut the promise, or name the team and the time frame. |
| 4 | Do the length and the sender fit the stage the candidate reached? | Switch to the right stage version. |
| 5 | Is the decision final and recorded, and will the sent copy be saved with it? | Hold the email until both are true. |
Scoring: five yeses, send. Four, fix the one and send. Three or fewer, don’t send; redo the brief and the draft.
This is also where a short “when not to” list earns its place. Don’t let AI send anything on its own, don’t let it answer a candidate’s angry reply, and don’t paste in anything you’d hesitate to read out in a hearing. Our fit test for when not to use AI at work is a quick way to sanity-check edge cases that aren’t on this page.
A rejection process can feel better to the team long before it’s better for candidates. The measurement lens we use at Future Factors has three steps: use, persistence and impact. Use is whether people are doing it. Persistence is whether they’re still doing it a few months later without being reminded. Impact is whether the work is actually better. Here’s how they look for this job.
| Step | Number to pull | Who pulls it, and when | What it tells you |
|---|---|---|---|
| Use | Share of rejections sent within five working days of the decision being final, from tracking-system timestamps | Talent acquisition coordinator, monthly | Whether the list still sits there. Five working days is our heuristic, not a benchmark |
| Persistence | In a random sample of five stage 2 to 4 emails per recruiter, how many still carry a true detail that ties back to the notes, in month three | Recruiting team lead, monthly | Whether the stage ladder became a habit or drifted back to one template |
| Impact | Count of “what happened to my application?” emails and calls per closed role, and how many finalists later accept another role or apply again | Talent acquisition lead, quarterly | Whether candidates are better off. These are proxies, so read them as direction, not proof |
A Future Factors measurement sheet for a recruiting team. It needs no new software beyond the tracking system you already have.
It helps to keep the four stages of change distinct. A recruiter who’s been shown what to paste and what to hold back has capability, which is the training part. The stage ladder gives that capability somewhere to land, and that’s workflow design. A team lead who spot-checks five emails a month is reinforcement, the thing that turns it into a habit. What you see by month three, if it worked, is adoption. Without reinforcement, a team can easily drift back to the 2019 template.
A team can start this internally with the tables on this page, and many will. Hands-on training shortens the learning curve in a few specific ways: it lets recruiters practice on your own tracking-system export and your past decline emails, it gets every recruiter working to the same standard instead of one person who’s good at it, and it covers the judgment calls that a guide can only point at. If you’re also thinking about how your employer brand reads to the people you turn down, our guide to AI for employer branding picks that up from the marketing side.
Here’s a first step for this week. Export the “under review” list for one closed role, sort it into the four stages, and write the stage 1 wrapper with the prompt. Send the oldest batch first, and book ten minutes on Monday to write the stage 4 call outline for whoever the hiring manager has to ring.
For applicants and recruiter screens, a reason usually adds more risk than value, in our judgment, so a clear close is enough. For finalists, a specific strength and one truthful, job-related reason in a call can be worth the effort. Whatever you say, say it consistently to candidates in the same position, tie it to the written requirements, and check your policy with employment counsel where you hire.
Drafting help on your own wording isn’t the same as an automated tool making or scoring the hiring decision, and the rules differ by place. Where a tool screened or ranked candidates, some jurisdictions add notice or explanation duties that are still changing, so ask counsel. Either way, a named person should be accountable for every message sent.
We suggest within five working days once the decision is final, which is a heuristic rather than a benchmark. For a runner-up, sequence matters: many teams wait until the chosen candidate has accepted, and tell the runner-up the timeline so the wait doesn’t feel like silence. Decide that order once, then write it into the routine.
Yes, at stage 1 a single wrapper with the name and role merged in is fine and often kinder than a slow handcrafted note. Avoid lines like “we were impressed by your background” for people nobody read closely. The moment someone has spoken to a recruiter, move up a stage and add one true detail.
Have a two-line holding reply ready that thanks them and says a named person will respond. Let AI draft the holding line if you like, but a person decides whether to offer feedback and handles any angry reply personally. Keep the exchange with the decision record.
The Greenhouse figure comes from Greenhouse’s own December 2024 write-up of its State of Job Hunting survey (2,500 workers in the US, UK and Germany), a vendor survey of self-reported experience. The one-year retention point comes from the text of 29 CFR 1602.14, and the notice steps after a background check come from the EEOC’s guide to background checks, both read on the publishers’ own pages in September 2026. Secondary roundups of candidate-experience statistics were not used because none could be traced to a primary source. Rules on automated screening tools were deliberately left out for the same reason, and nothing here is legal advice. The example role, company and people are illustrative, not real cases.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.