Step-by-step guide AI Job-Search Fast-Track: experienced professionals (3–15 years) →
In a survey of 590 US job seekers, 80% said they use AI tools and 94% said they tailor their CV for each application (Clutch, 21 July 2026). Yet 67% of the 1,000 US job seekers Resume Genius surveyed worry their CV will be rejected before a person sees it, and half say they don't understand how applicant tracking systems work (Resume Genius, 8 April 2026). So many people are using AI against a machine they can't see, and their CVs increasingly read alike. This article explains what screening software actually checks, why ChatGPT and Claude tend to inflate your claims, and a routine that tailors a CV in about fifteen minutes without inventing anything. The examples come from a test we ran for our new guide, on a fictional CV.
What happens after you click Apply
Applicant tracking systems (ATS) differ, but an application through a company careers site usually passes four stages. The one most likely to reject you automatically is the first, and it doesn't read your CV.
- Knockout questions. The form asks about right to work, location, licences or notice period. In Greenhouse, for example, an employer can set an Auto-Reject rule on these answers. The rule reads what you tick, not what you wrote.
- Parsing. The system copies your file into fields such as jobs, dates and skills. A layout it can't follow sends your experience to the wrong field, or nowhere.
- Matching. More and more often, software compares each requirement in the ad with what your CV shows.
- Search and review. Recruiters search and filter, typically by job title and skills. Then a person reads and decides.
For stage three, the vendors' own documentation is the best guide. Workday's HiredScore release note of August 2026 describes an opt-in upgrade in which a language model checks each qualification separately against your parsed CV. Applicants who meet all the basic qualifications are graded A or B, and the recruiter sees which requirements were met and which weren't (Workday, August 2026). Greenhouse's Talent Matching sorts applicants into strong, good, partial and limited matches against criteria the employer weights, highlights matched keywords and, in Greenhouse's words, "does not automatically advance or reject candidates" (Greenhouse, updated 5 October 2026).
| Stage | What it reads | What you control |
|---|---|---|
| Knockout questions | Your answers on the form | Truthful, complete answers that match your CV |
| Parsing | Your file's text, in reading order | One column, standard headings, real text |
| Matching | Your CV against this employer's requirements | Visible proof for each must-have, in the ad's words |
| Search and review | Titles and skills, then the whole page | Specific results a person can check |
Notice what's missing: a universal "ATS score". Each system compares your evidence with one employer's list for one job. The useful question isn't how to fool a machine, but whether your CV shows proof for each requirement, in the employer's words.
Why AI makes your claims bigger
Ask an assistant to match your CV to an ad and it will, including where your experience doesn't. Language models write the most plausible sentence, and the most plausible CV line for a job ad is one that sounds like the ad. The result is rarely an invented number. It's a verb or a scope that has grown: "supported" becomes "led", a spreadsheet becomes "workforce management".
We saw this first-hand. For the guide we created a fictional candidate, Robin Vermeer, a customer-operations lead with nine years' experience and a CV of duties without numbers, plus two invented ads: a senior operations role at a London home insurer and a customer operations role at an Amsterdam payroll-software company. We ran every prompt in ChatGPT (free, signed out), with Claude as a cross-check. The AI invented no numbers. It did inflate:
- Scope. Interviewing Robin for achievements, under an explicit "never invent" rule, it wrote "managing a 33-person operation". Robin ran 24 agents and 3 team leads; the other 9 agents were outsourced and run by the partner.
- Verbs. It said Robin "recruited" eight temporary agents. Robin trained them; someone else hired them.
- The ad's own words. Tailoring for the London ad, ChatGPT added "identified automation opportunities" (the ad's phrase, not Robin's evidence), managing the outsourced team, and "capacity planning" as a skill.
None of these is a lie you'd spot skimming, and every one is something an interviewer would ask about. The same Resume Genius survey found 36% of job seekers had considered listing skills they don't have yet. When every CV sounds tailored, wording stops being a signal. What stands out is specific proof that holds up when somebody checks it.
The routine: one evening, then about fifteen minutes per job
The method splits the work so the AI only ever rearranges facts you've already confirmed.
Once: an evidence bank and a master CV. The evidence bank is a private list of your achievements, one entry each: the situation, what you did, the result with a number, and how you could prove it. Let the assistant interview you for it, one role at a time, with the instruction never to invent facts or numbers. Then read every line and shrink anything bigger than the truth. Add a "Don't claim" line for what ads often ask for but you haven't done: a sector, a tool, a qualification. In Robin's test that line turned three possible stretches into honest gaps. Build a one-column master CV from the bank. Robin's bullets with a number went from 0 of 15 to 14 of 15, and a standard text extractor found five problems in the old two-column template (the name read after the sidebar, a date drifting away from its job) and none in the one-column version. Allow one and a half to two hours.
Before you paste anything: replace your name, email, phone number, address and LinkedIn link with placeholders such as [NAME], [EMAIL], [PHONE] and [LINKEDIN]. The assistant doesn't need them. Check the tool's data settings too, and turn confidential figures into percentages: "cut refunds by 23%" proves the same skill.
Then, for each application:
- Decode the ad and map your proof (about 4 minutes). In one prompt, ask for the must-haves, nice-to-haves and likely knockout questions, then a table: the ad's exact words, your best proof from the evidence bank, and Strong, Partial or Gap. Tell it not to stretch your experience and to stay under 400 words. Then check every "Strong", because the map matches words, not meaning. In Robin's test it counted onboarding new staff as proof of onboarding customers, and rated a 33-agent team Strong for a "high-volume contact centre" at a 140-person employer. If two or more must-haves are honest gaps, look for a better-matched ad.
- Tailor (about 3 minutes). In the same chat, ask for a reordered CV: bullets that prove a must-have first, the ad's terms only where they describe something you really did, and nothing that isn't in the evidence bank. Ask it to list any new fact you mentioned, so it goes into the bank before it goes into a CV.
- Truth check, in a fresh chat (about 4 minutes). The step most AI CV advice leaves out; see below.
- Export and test (about 3 minutes). Paste the checked text into your master layout and save as PDF or Word, whichever the form asks for. Then select all in the file, copy, and paste into a plain-text editor. It shows what one common extractor reads, not exactly what a given ATS sees, but it catches the usual scrambles.
The truth check
Open a new chat so the assistant isn't marking its own homework, ideally in a different assistant from the one that tailored. Paste the tailored CV and your evidence bank with an instruction like this:
Act as a sceptical hiring manager. Check this tailored CV line by line against my evidence bank, including the profile and every word in the Skills section. Flag only claims the evidence bank doesn't mention, and claims worded more strongly than the evidence ("led" where I supported, "managed" where I oversaw). Reply only with a table: the line, the problem, an honest fix. Under 250 words.
Robin's London CV went in with five real problems. ChatGPT, checking the CV it had written, caught three. Claude, checking the same CV in a fresh chat, caught all five. That's one run each, so treat it as a pointer, not a ranking. Two lessons held either way: the checker also flags lines that are fine, so you make the final call, and a fact you mention in the chat but never add to the evidence bank can slip through. "Willing to relocate" was true, but it got past one checker for exactly that reason.
Finish with your own read. Scan for the five verbs that grow on their own: led, managed, owned, drove, identified. If you didn't, change it.
What changed for Robin
| London ad (fictional test) | Result | Note |
|---|---|---|
| Ad terms Robin can truthfully claim, present in the CV | 5 of 18 in the master CV; 11 of 18 tailored with ChatGPT, 14 of 18 with Claude | Only terms for work Robin really did |
| Overstatements in the tailored CV | 5 found, 0 left after fixes | Claude's check caught 5 of 5, ChatGPT's 3 of 5 |
| Numbers invented by the AI | 0 | Across every prompt |
| AI time per application | About 1 minute | Measured |
| Your time per application | About 13 minutes | Estimated, not timed |
These are results from one fictional CV and two invented ads, so they show how the routine behaves, not what it will do for you. The "ad terms" count is a rough proxy for what keyword searches and matching tools look for. We call the routine fifteen minutes rather than thirteen because your share is an estimate, not a stopwatch reading.
No method can guarantee interviews or offers. What this one does is make sure the people and systems reading your application see what you've actually done, in the words they're looking for, with nothing you'd struggle to defend.
Where the full method lives
We've written the routine up as a guide: The AI Job-Search Fast-Track, experienced professionals edition. It's 23 pages for people with 3 to 15 years of work behind them: Robin's worked example in full, eleven tested prompts, chapters on the cover note, the application form and interview prep, EU and Dutch CV notes, and a checklist. Editions for career switchers and graduates are on the way, and our other practical AI guides are in the guides catalogue.
If you'd rather work from this article alone, the steps above are the core of it: build the evidence bank once, map before you tailor, and never send a CV that hasn't been truth-checked.
Frequently asked questions
Can I use ChatGPT or Claude to write my CV? Yes, for making true things clear: structure, wording and using the employer's terms for work you really did. What an assistant can't supply is your evidence, meaning the numbers, systems and scope that prove the work. Use it to arrange and edit facts you have already confirmed, and run a truth check before you send anything.
Do applicant tracking systems reject CVs automatically? Less often than people fear. The automatic rejections that do exist usually come from knockout questions on the application form, such as right to work, location or a required licence. Matching tools such as Workday HiredScore and Greenhouse Talent Matching compare your CV with one employer's criteria for one job, so there is no single score to chase.
How long does it take to tailor a CV with AI? In our test on a fictional CV, the AI's share took about a minute and the whole routine took an estimated 13 minutes per application, including reading, correcting and the truth check. Building the evidence bank and master CV first is a one-off job of about one and a half to two hours.
Is it safe to paste my CV into ChatGPT or Claude? Remove what the assistant doesn't need first. Replace your name, email, phone number, address and LinkedIn link with placeholders such as [NAME] and [EMAIL], check the tool's data settings, and use a percentage instead of any figure that is confidential.