The advice on this subject splits into two useless halves. One says never use AI on your resume, which ignores that it is genuinely good at several parts of the job. The other says paste in the job description and let it write the thing, which produces a document that sounds like everybody else's and that you then have to defend in an interview.
The useful version is a division of labour. There are tasks in writing a resume where a model outperforms most people, and tasks where it cannot help by construction. Knowing which is which is the whole skill.
What a model is genuinely good at here
- Compression. Cutting a 40-word bullet to 20 without losing the substance. This is tedious, mechanical, and models are better at it than most people are at their own writing.
- Structural critique. "Which of these bullets describe duties rather than outcomes?" is a question it answers well, because the distinction is a pattern in the text.
- Vocabulary translation. Career changers need to say the same work in a different field's words. A model has read both fields and is good at the mapping.
- Interrogation. Ask it to list the questions an interviewer would ask about a bullet. This is genuinely useful, and it is a use most people never try.
- Proofreading for the errors spellcheck misses: a correctly spelled wrong word, an inconsistent date format, a title that changed between versions.
What it cannot do, by construction
It does not know what you did. Everything specific and therefore everything persuasive — the number, the system you owned, the decision you argued for and lost, the thing that broke and what you learned — exists only in your memory. A model asked to supply that will invent something plausible, because that is what it does when the information is absent.
It also cannot judge what matters for this particular job at this particular company. It can echo the job description back to you. It cannot know that the team is rebuilding on a deadline, or that the hiring manager has been burned by someone who could not communicate with finance.
The rule that follows
Supply the substance yourself; delegate the shaping. If a sentence in your resume contains a fact you did not put there, delete it. That single rule prevents almost every way this goes wrong.
A workflow that works
Step one, write badly by yourself. Get everything down without worrying about phrasing: what the job was, what you built, what changed, what the numbers were. Ugly and complete is the correct output. This is the step people want to skip and the one that determines whether the rest is worth anything.
Step two, ask for a critique, not a rewrite. "Which of these bullets describe responsibilities rather than results? Which claims have no evidence attached? What would an interviewer ask about each?" You get a list of problems and keep authorship.
Step three, fix the problems yourself. Where the critique says a bullet lacks an outcome, you are the only one who knows the outcome. This is where the resume actually gets better.
Step four, delegate the polish. Now hand it back for compression and consistency. "Cut each of these to under 25 words. Do not add any information that is not already present." That last clause matters and is worth including every time.
Step five, read every line as yourself. Any sentence you would not say out loud in an interview gets rewritten in your own words, even if the model's version is better prose. Slightly worse and defensible beats slightly better and borrowed.
What "voice" actually means on a resume
Voice on a resume is not personality or style; there is very little room for either in a bulleted document. It is something narrower and more useful: the particular things you choose to mention, and the level of detail you go into about them.
Two engineers with identical jobs write different resumes because one thought the migration was the interesting part and the other thought the incident review process was. That choice is voice, it is evidence of judgement, and it is precisely what gets flattened when a model decides what to emphasise. A generated resume reads as though nobody in particular held the job, because in a sense nobody did.
This is why the register everyone recognises — leveraged, spearheaded, drove alignment — is a problem beyond being tired. It signals that no choices were made.
Prompts that produce something useful
The difference between a helpful response and a generic one is almost entirely in whether you asked for judgement or asked for text. These four are worth keeping.
- "Here are my bullets for this role. For each one, say whether it describes a responsibility or a result, and if it is a responsibility, tell me what information is missing to turn it into a result. Do not rewrite them."
- "Here is a bullet and the job description. What are the three questions an interviewer would ask about this bullet?" Their answers show you what the line implies that you have not substantiated.
- "Cut each of these to under 25 words. Do not add any information that is not already present, and do not change any number." The constraints are the whole prompt.
- "I did this work in teaching. I am applying into product management. What does each of these activities get called in that field?" This is the translation task, and it is the one models are best at.
Notice that three of the four ask for analysis rather than prose. That is the pattern: the model is most useful as a reader of your writing, not a producer of it.
Where it helps most: the career change
If there is one situation where these tools earn their place, it is changing fields. The core difficulty of a career change is that you have the relevant experience and describe it in the wrong vocabulary, so a reader in the new field cannot see it. You cannot easily fix this yourself, because the words you have are the only ones you know for the work.
This is a translation problem, and translation is the thing a model trained on both fields is genuinely good at. Describe what you actually did in your own words, ask what the target field calls each of those things, then check every suggestion against the truth and discard anything that overstates. The result is your experience, in a language the reader speaks. That is a real advantage, and it is a different activity from asking a machine to write you a resume.
The check that catches everything
Read your finished resume aloud and, for each bullet, say the next two sentences you would give an interviewer who asked about it. Out loud, not in your head.
Where you can do that fluently, the line is yours and it is true. Where you stall, one of two things is wrong: either the line claims more than happened, or it is written in words that are not yours and you are translating on the fly. Both are fixable now and both are expensive in the room.
The tools that are not chatbots
Most of this article is about language models, because that is what people mean by AI tools. The less discussed category is more reliably useful: things that measure rather than generate.
A parser check tells you what an applicant tracking system extracts from your file, which is information you cannot get any other way. A match breakdown tells you which of your components is weak against a specific posting. A tracker tells you which stage of your funnel is failing. None of these writes anything for you, and all of them tell you something true that you did not know, which is a better trade than fluent prose you have to verify.
Is any of this cheating?
No, and the framing is not useful. Nobody thinks less of you for using a spellchecker or asking a friend to read a draft. The line is not tool use; it is truth. A resume improved by a model that describes work you really did, in claims you can defend, is an honest document. A resume that describes a stronger candidate than the person who will arrive at the interview is a problem regardless of what produced it, and it was a problem long before any of this was automated.
Measure it instead of guessing
The component breakdown against a real posting: which bullets read as duties, which terms are missing, and what a parser makes of your file. It reports; you write.
Try it free