Should You Still Ask for a Cover Letter If AI Writes Them All?

Should you drop the cover letter entirely?

No. Drop it from most reqs and keep it on a few. The letter is a writing sample wearing the costume of a formality, so it carries real information exactly where writing is the work: communications, brand, policy, technical writing, fundraising. Federal occupational data rates writing importance at 97 out of 100 for technical writers and far lower for most other jobs ². A blanket keep-or-kill policy gets half your postings wrong.

The fork is not seniority and it is not industry. It is whether prose is a deliverable the person will be paid to produce. A communications manager, a policy analyst and a grant writer ship writing as output, and a letter is a small honest sample of it. A staff engineer, a warehouse supervisor and a revenue-cycle specialist do not, and their letters were never evidence about the job.

On those reqs the letter was measuring effort, and effort is the part that broke. Twenty minutes of work used to separate people who wanted this job from people who wanted any job. That cost is now close to zero, the filter is gone, and what remains is a field every applicant fills and nobody reads, one contributor to the volume problem covered in what to do when applications per opening triple.

A blanket rule is what makes this expensive in both directions. Requiring letters on thirty open reqs means thirty stacks of a document that decides nothing. Dropping them everywhere means your one communications hire gets screened on a resume, which is the wrong instrument for the only role where the letter was still working.

What does an AI-written letter actually cost you?

Less than it looks, and not in the direction most people assume. In a field experiment with nearly half a million jobseekers in an online labor market, giving people algorithmic writing help on their resumes raised hires by 8%, and employers were no less satisfied with who they got ¹. Clearer writing helped employers read ability rather than disguise it. What you lost is spread: when every letter clears the bar, none of them separates anyone.

That result is worth sitting with, because it contradicts the usual complaint. The fear is that writing help strips out a signal about the person. What the experiment found was closer to the opposite: better prose made ability easier for the reader to see, and satisfaction after the hire held ¹. The letter was doing less signaling than the ritual around it suggested.

What did change is the distribution. When the median applicant writes at the level that used to mark the top of your stack, the sort you were making by feel stops sorting. The tell is no longer quality. It is specificity: whether anything in the letter could only have been written by someone who read the posting, used the product, or knows what your team argues about on a Tuesday.

Do not try to fix this by guessing which letters a model wrote. Detectors misclassify writing by people who learned English later as machine-generated, and simple prompting evades them ³, so the false positives land on the applicants you have the least reason to cut. Whether AI detectors work in hiring has a short answer, and the longer problem is that authorship was never the question: a stack of perfect-looking applications is a signal problem, not a fraud problem.

Replace the letter with one question a model answers generically

One required field on the application, capped at 100 words, tied to a decision your team actually argues about. The test for a good question is blunt: a fluent generic answer to it should be visibly generic. Ask about a trade-off inside your context (your pricing, your constraint, your users) and the non-answer sits there in plain sight next to a specific one.

Five rules make the field work, and skipping any one of them turns it back into a cover letter.

• Ask it of everyone, the same way. A question that decides who advances is a selection procedure, and applying it consistently is what makes it explainable later ⁴.
• Say that AI is allowed. A ban you cannot enforce only teaches applicants that the honest answer is the risky one.
• Write the answer key first. Three things a good answer contains, written before the first application arrives. Without it you will grade fluency by accident.
• Grade content, not prose. A specific answer in clumsy sentences beats a polished one that names nothing.
• Enforce the cap. A field with no word limit becomes a cover letter again inside a week.

The question also travels further than the letter did, because it can be reused in the phone screen verbatim: ask the applicant to say more about the answer they submitted. Someone who wrote it can extend it; someone who pasted it stalls. More options for the field, and what each one costs to read, are in what should replace the cover letter in a screen.

Keep the letter where writing is the deliverable

On communications, brand, policy, technical writing and fundraising reqs, keep the sample and change the ask. Instead of a letter about the applicant, request 250 words to a named audience under a stated constraint: a renewal at risk, a launch slipping, a policy nobody read. Say plainly that AI is allowed. Then ask the second half of the question: what the assistant got wrong, and what got cut.

The second half is the part that survives a model. An applicant who drafted with an assistant and edited hard can tell you which paragraph was a hallucinated benefit, which sentence assumed a fact about the audience that was not true, and why the version you are reading is shorter. An applicant who pasted the first output has nothing to say there, and the gap shows up in two sentences rather than two rounds.

Three things keep this fair and defensible. Give every applicant the same brief and the same constraint. Write the rubric before you read the first submission, so the standard is not set by whoever happened to submit first. And cap the time it should take, in the posting, because an open-ended writing sample quietly becomes unpaid work that filters for people who can afford to do it.

One caution on scope. A writing sample tells you about writing, and it will tell you almost nothing about how the person handles a confident wrong answer in a spreadsheet. If the posting is going to ask for AI skill on top of the writing, say what you actually require instead of putting 'AI-proficient' in the requirements and hoping the letter settles it.

What can a replacement question still not tell you?

Anything about the act itself. Whatever you put on the form comes back as a finished document, produced off-platform, with unlimited time and any tool the applicant owns. It can show you what someone chose to say about their judgment. It cannot show you the judgment: the moment a confident, wrong answer arrived and either got tested against something real or got shipped.

So treat the field as a sorting device for the first conversation, not as evidence for a decision. It earns its place by being cheap, consistent and hard to answer generically. It does not earn the weight people used to put on the letter, and stacking three of them does not add up to a work sample.

The weight belongs downstream, in one exercise run identically for every finalist and graded against something written beforehand. The arithmetic for that trade is in skipping the resume screen for a work sample. That is also where the honest limit of everything above sits: an application field can be gamed by anyone with an afternoon, and a work sample built for the occupation cannot be gamed the same way, because the interesting failures happen while the work is being done.

Several instruments cover that stage, and they are not equivalent: multiple-choice AI literacy tests, code-collaboration graders, unwatched take-homes, live AI interviews and in-house rounds all exist with different costs and different blind spots. Olive is one of them. Pick on what the role fails at, not on what is easiest to add to the pipeline this quarter.

See a sample report >