There's a category of job-search tool whose entire pitch is volume: connect your résumé, set it loose, and it applies to hundreds of jobs while you sleep. Somewhere in those hundreds of applications, the tool meets a screening question it can't answer from your résumé ("How many years of production Python experience do you have?") and it has two options. It can stop and ask you. Or it can guess.
The volume tools guess. Users of mass-apply services have documented the bots filling in screening answers (years of experience, salary expectations, work authorization, willingness to relocate) with values the AI invented to maximize the odds of passing the filter. The application goes out under the user's name. The user finds out, if they ever do, in the interview.
Why "it's just a draft" doesn't apply
Drafting tools have a built-in safety valve: you read the draft before you send it. Mass-apply bots remove that valve. The send is the product. That changes the failure mode entirely. A hallucinated sentence in a draft costs you an edit. A hallucinated screening answer in a submitted application costs you the credibility of everything else on the page, and recruiters do check the inconsistencies first.
It also doesn't stay hypothetical. Experience minimums are often knock-out questions backed by verification later: a background check, a skills screen, a casual "tell me about your five years with Python." There is no good moment to discover the answer was never yours.
What refusing looks like
The fix isn't "better AI" in the sense of a smarter guesser. It's an AI that is structurally unable to guess about you. When we built Pickier's drafting engine, we gave it one hard rule: every claim in a cover letter or tailored résumé must trace to a fact in your profile. When the model reaches for something it doesn't have (a tool name, a team size, a number), it can't invent it. It has to leave a visible gap and ask:
«NEEDS: which incident-management tool did you run?»
You answer in your own words, or you delete the sentence. Either way, the finished draft carries a receipt ("Drafted from 14 facts in your profile · 0 invented"), and the second number is not a goal. It's a constraint.
Fewer, better, honest
The deeper disagreement with the volume tools isn't technical, it's strategic. A recruiter can tell application #312 of a bot run from a letter written by someone who read the posting. Five applications a day that you've actually read, scored against who you actually are, with claims you can actually defend. That's not a limitation of our approach. It's the approach.
They apply blindly. We won't.