What it genuinely helps with, and what it quietly ruins

One test decides every use

Can you verify the output, and is the output itself the evidence?

Where you can check the result and the result is not what you are being judged on (reformatting, summarising, drafting a first version, generating practice questions) a model saves real hours. Where the output is the evidence of your ability, using one destroys the thing you were trying to produce, and everybody downstream finds out at the technical stage.

Everything below follows from that single test.

Where it saves hours

Vocabulary translation. The same job is called “Data Analyst” in one country, “MIS Executive” in another and “Business Intelligence Analyst” in a third. Paste three adverts and ask what terms they share that your CV does not use. This is useful, because you cannot see your own vocabulary.

Gap analysis against an advert. Paste the advert and your CV and ask: ”List each stated requirement and quote the line in my CV that evidences it. Mark the ones with nothing.” You will get a short honest list of holes, which is what you actually needed. Do not then ask it to fill them.

Decoding jargon. “Experience with medallion architecture and dbt semantic layers in a Databricks environment” is four things, and knowing which two you already do under different names changes whether you apply.

Research. Company, product, market, competitors, likely metrics, as a starting point to be checked, never as a fact.

Practice material. “Give me eight SQL questions on window functions at increasing difficulty, with answers” produces a decent exercise set in seconds, and you can check every answer by running it.

Rehearsal. Covered in its own lesson.

The first draft of an outreach message, which you then rewrite entirely, because the first draft’s job is to stop the blank page, not to be sent.

Where it costs you

A cover letter written wholesale. Readers detect these at a rate that surprises people, and the tell is register rather than grammar: three-item lists, “I am particularly drawn to”, “excited about the opportunity to leverage”, a paragraph of enthusiasm containing no fact about the company. The specific damage is that a generated letter contains nothing only you could have written, which was the letter’s only job.

A portfolio project you cannot explain. Covered in its own lesson. Briefly: the interview asks why, and you will not know.

Auto-apply services. They spend your money to worsen the ratio described in the volume lesson, and applying to roles you plainly do not fit is visible in the company’s system next time.

Salary figures, hiring processes and company facts stated confidently. Models are fluent about pay bands and interview stages in your city and are frequently wrong, because that information is thin, local and changes yearly. Check against real postings and people.

”ATS score” tools. The next lesson explains why the number is invented.

The register problem, and the fix

Default model prose has a shape: balanced, enthusiastic, superlative, hedged, and structurally identical every time. It is exactly the register that a recruiter reading forty applications has learned to skip.

If you use a model in your writing at all, do the last pass yourself and cut anything that is true of any candidate. “I have a passion for using data to drive impact” survives being said by anybody. “I rebuilt our weekly stock report after finding the depot codes did not match between two systems” does not, and it is the only kind of sentence worth the paper.

A useful instruction when drafting: ”Write it in plain sentences. No adjectives about my enthusiasm. Every claim must be traceable to something in the notes I gave you.” Then read it and remove another third.

The limitation

None of this changes the funnel much. Better wording and formatting help at the margin: in one field experiment with nearly half a million jobseekers, AI writing help raised the chance of being hired by 8%. The route in matters far more: at one large US firm, referred applicants were 6% of applicants but 29% of hires (Brown, Setren and Topa). A model can make forty applications feel like productivity while the thing that would actually work (five people contacted properly) goes undone. Use it to buy time, then spend the time on the route.