Tailor your CV to a job description with AI
A worked example of tailoring one real CV to one real job ad with AI, what the model gets right, and the three things you have to catch yourself.
Priya has six years at a software company, most of them in customer support, and a job ad open in the next tab. The role is customer success manager at a logistics scale-up. The ad wants someone who can onboard new accounts, own renewals, watch account health in HubSpot, and report on all of it to a head of revenue.
Her CV is nearly right, which is the worst kind of nearly. It opens with a line about resolving inbound tickets at volume, then a bullet about CSAT scores, then two more about queue management and a support handbook. Down in her second job, fourth bullet, sits the sentence that actually matters: "Ran onboarding calls for new enterprise customers." She did that for two years. It is the closest thing on the page to what the ad is asking for, and it is the eleventh thing a reader meets.
The distance between what Priya has done and what her CV puts first is the whole problem. It is also the part an AI model is genuinely useful for, as long as you know which half of the job you are handing over.
What tailoring actually means
By the time Priya is finished, nothing on her CV will be new. The onboarding line moves to the top of that role. The ticket-volume bullet drops two places and loses six words. Where she wrote "customers" and meant "accounts", the word changes, because the ad says accounts and she is describing the same thing. The support handbook she is proud of goes down to a single clause. That is the entire operation: reordering, re-emphasising, and matching vocabulary to the language the employer already used.
Reading the ad closely is its own piece of work and deserves more time than this page gives it. For CV purposes the useful output is short. Write down what the ad asks for, in the ad's own words, in the order the ad puts them. Priya's list runs: onboarding, renewals, HubSpot, reporting. Four things. Everything from here is matching one page against those four.
Getting the CV in front of a model
Priya's CV is a PDF, which is how most CVs travel, and that detail decides more than you would expect. Twenty-five of the 60 models on MultiChats accept a PDF attachment. The model picker marks them with a blue PDF pill, so you can see before you attach rather than after. If the model you have selected is not one of them, the attach is refused on the spot with a message telling you to switch, so at least the failure is loud. On the free plan the PDF-capable set is narrower: four of the 21 free models, GPT-5.4 Nano and three Gemini Flash Lite entries.
The free plan allows one attachment per message, up to 10MB, three uploads a day. A CV is a long way under 10MB, so size is not the constraint. One file per message is, because Priya cannot send the CV and a PDF of the ad in the same message. She pastes the ad as text underneath the attachment, which is better anyway, since job ads are usually a web page rather than a file.
Now the part that catches people out. If Priya's CV is a scan, or came out of a design tool as a flattened image, there is no text layer to extract. On a signed-in account a scan of up to 20 pages is transcribed first and the model reads that transcript, which can misread a date or a figure. The test takes four seconds: open the PDF and try to select a line of text with your cursor. If the text does not highlight, check every date and number in the reply against the original, or paste the CV as plain text instead. Word documents, along with text and code files, are a Pro upload.
The gap list comes first
The instinct is to say "rewrite my CV for this job". Resist it for one message. What Priya sends first is closer to an audit:
Here is my CV, and the job ad below it. Do not rewrite anything yet. For each requirement in the ad, in the ad's order, tell me which line of my CV already answers it and where that line currently sits. Then list what the ad asks for that my CV does not say anywhere.
The answer is more useful than a rewrite would have been. Onboarding: answered, buried in the second role. Renewals: not answered, though she has a line about reducing repeat contacts that is adjacent and not the same thing. HubSpot: not answered, because she used Zendesk and Salesforce. Reporting: answered, but disguised as "maintained internal documentation", which is what she called the weekly queue report she built and sent to her director every Monday.
Two of those four are real gaps and no rewriting will close them. Knowing it is worth more than a polished page, because it tells Priya what she should be ready to say out loud about HubSpot. It also stops the model papering over the holes.
The rewrite request comes second, and it carries a constraint: use only facts already on the page, do not add numbers, and tell me which bullets you moved and why.
The failures a recruiter spots in four seconds
Invention is the first and the worst. Ask for a rewrite without constraining it and a model will hand back "reduced churn by 18%" with total composure. Priya never measured churn. Nobody at her company measured it in a way she could see. The number came from the shape of the sentence.
Fabricating experience on a CV is wrong, and it is the least survivable kind of wrong, because the interviewer will ask how she got to 18 percent. There is no good version of that conversation. The same goes for the softer edits that drift: a model that promotes "helped with onboarding" to "owned the onboarding programme" has changed what she is claiming, and the person across the table will ask who reported to her.
Second is the achievement verb. Spearheaded, drove, leveraged, orchestrated, championed. They arrive in batches, and a recruiter working through forty applications in a day is pattern-matching. A page of promoted verbs is a pattern. Priya's own verbs, ran and wrote and sent, survive that reading.
Third is vocabulary matching pushed too far. Echoing the ad's language once, where it is honest, helps. Echoing it in every bullet produces a CV that reads like the job ad with a name at the top, and the effect is the opposite of the one intended.
A second model, and what it cannot do for you
Priya wants a second version without losing the first, so she branches from the rewrite and picks a different model. The action is labelled Try Again on the web and Branch here in the mobile app, and that gap matters: on a phone, Try again is the regenerate action, which deletes the answer it replaces. Branching keeps both. It opens a new thread carrying the chat up to that point, so the two versions sit in her sidebar as separate chats. Nothing appears next to anything else, so this is a reading exercise rather than a diff.
It is worth doing, because the two versions disagree usefully. One keeps her weekly report as its own bullet, the other folds it into the onboarding line and wins back a sentence of space. She keeps the stronger line from each.
What the second model does not do is verify the first. Both will happily generate the same invented percentage. Both will reach for the same verbs. Agreement between two models tells you something about how models write. It tells you nothing about Priya's career. The only check that catches a fabricated number is her own memory of doing the work.
A CV is also not a neutral document. It carries her address, phone number and current employer. She can cut the contact block before pasting, since it contributes nothing to the rewrite. As for what happens to the rest, the honest version is specific: we do not train on your conversations, some providers might depending on which model you pick, and the model's info card says so when it applies.
Own the last read
The final version is one page. Onboarding opens her second role instead of sitting fourth in it, and the summary line at the top says the word, so a skim finds it. The ticket line survives, shorter. Two bullets are gone. Nothing on it is new.
The honest summary is that AI was very good at the middle of this job and no use at either end. The middle is mechanical: holding four requirements and eleven bullets in mind at once so it can see that bullet four answers requirement one, then cutting thirty words to fifteen without losing the meaning. A model does that faster and more patiently than Priya would at eleven at night.
The ends stayed hers. Deciding that the onboarding work, rather than the handbook she is proudest of, is what this employer wants to see is a judgement about her own career that no model has standing to make. And the last read, the one that catches the sentence she could not defend in a room, cannot be delegated, because the person who has to defend it is the only one who knows what she did.
Use it to tailor the application, then read it twice before sending.