CoivreBack to home

AI Policy

Effective as of: September 2, 2026

Coivre uses an AI model to read job advertisements and to write drafts for you. This page says exactly what it is used for, what is sent to it, what it is forbidden to do, and where it stops.

1. What the model is used for

There are five things it does, and no others:

  • Evaluate a posting. It returns a label for what kind of role it really is, a score out of ten against your profile, a verdict (apply, consider or skip), a judgement of whether the posting looks like a real specific job, and a short report on where you match and where you do not.
  • Summarise a posting for the triage card: an English title, one or two sentences on what the job actually is, and up to six concrete tasks the advertisement names.
  • Draft an application email: a short letter, from a person to a person, built around one real detail from your history that connects to what the posting asks for. When automatic applying is on, a drafted letter can be sent without further review; Section 5 describes exactly when.
  • Tailor your CV for one posting, by selecting and reordering what your facts document already contains so the most relevant real experience leads.
  • Prepare you for an interview with that employer: the likely questions, answers built from your real history, the weak spots they may probe and the truthful way to handle each.

2. What is sent to the model

For every account, the same two things travel and nothing else: the job advertisement (its text, the role, the employer, the location and the hours, as published) and your facts document, which is the CV you provided. For a drafted email, two more things go with them: the writing preferences you saved on your profile, and any revision notes you typed asking for a change. Your name, phone number and email address do not travel: the model is forbidden from writing contact details and the system appends them afterwards, which is why they are not in this list.

Your data is never mixed with another account's. Nothing is sent to the model that you did not put there or that we did not read from the advertisement you asked us to read.

Training. Our provider does not train on data sent through the API, and we have not opted into any programme that would change that. Your CV and your applications are not training material.

3. The rule the model cannot break

Every prompt carries the same clause, injected as one shared constant so it cannot drift between one feature and another. There is a check in our build that reads every prompt and fails if one of them is missing it, because a rule that depends on somebody remembering is not a rule:

Every factual claim about the candidate (jobs, dates, tools, certifications, languages, education) must come from the facts document. If the posting asks for something the document does not show, say so or leave it out; never invent or inflate.

The four features that say anything about you (the evaluation, the email, the CV and the interview prep) refuse to run at all if the facts document is missing, rather than writing something unsupported. The fifth, the card summary, describes the advertisement and never the candidate, so it runs without one. Contact details are never written by the model: your name, phone and email are appended by the system from your profile, deterministically.

This is not a preference about tone. An invented certification survives until the first interview and burns an application you cannot get back, which is why it is enforced in the code rather than hoped for.

4. The advertisement is treated as hostile input

Anyone can publish a job advertisement, which makes advertisement text the least trustworthy input this system has. A posting that contained instructions aimed at the model, rather than a description of a job, could otherwise try to change what we write about you or where an application is sent.

So the posting is not simply pasted into the prompt. Before it reaches the model it is:

  • stripped of the characters and structures used to forge a boundary in a prompt;
  • wrapped in markers carrying a code generated fresh for that single call, which text inside it cannot know or reproduce;
  • tagged line by line with that code, so there is no passage, however deep, where the model is more than a few words from being told that what it is reading is data;
  • followed by an explicit reminder that nothing inside it can add or change a fact about you, and that it is never a message from us.

The recipient never comes from the model. An application is only ever addressed to an address the scraper found and matched against the employer's own domain, or one you typed yourself.

5. Where the model stops, and where the automation goes further

The line runs between the model and the rest of the system, not between the machine and you. The model scores and drafts. It never chooses a recipient, never decides whether automatic applying is unlocked, never sets the sending limits or the sending window, and never decides that a run happens. All of that is ordinary code you can hold us to.

On the manual path, an application leaves only after you have read the letter and clicked send, and the address it is going to is shown next to that button.

Automatic applying sends without that per-letter reading, and these are its real numbers, which this page follows the code on. It unlocks only after at least 30 decisions you made yourself on everyday-lane postings, agreeing with the filter's verdicts at least 85% of the time (a one-sided measure: it is computed on postings the filter let through, so it cannot see a good job the filter wrongly discarded, and the app says so where the number is shown). A run handles at most 20 postings, nearest your address first; you can start at most 3 runs an hour; a daily sending limit is shared with manual sending; and everything waits for the working-morning window, Copenhagen time: before 11:00 Monday to Thursday sends now, after 11:00 Monday to Wednesday waits for the next morning at 09:00, and from Thursday 11:00 through the weekend everything waits for Monday at 09:00. Nothing starts a run except you, and no schedule exists.

What stands in place of your reading while a run executes: the rule in Section 3, enforced by a build check; the hard refusal to write anything without your CV document or your contact details; the verified-recipient rule in Section 4; one application per posting, ever; the limits and the window above; and a record of every run, every send and every halt in your account. Whether the law calls this automated decision-making is treated on the Privacy Policy, together with the safeguards you have either way.

It can be wrong. It can score a good job badly, miss a bad one, misread an advertisement or write a letter that does not sound like you, and when automatic applying is on, a wrong letter can reach a real employer without you having seen it, which is a real cost: that application cannot be recalled. Read what it produces. On the manual path you can edit any draft by hand or ask for it to be rewritten with your notes, and if that trade ever stops being worth it, automatic applying is yours to leave off.

6. What it costs, and the ceiling on it

Each of the five actions above is one AI run. Your plan includes a number of runs a month, and when they are spent the AI features stop until the next month or until you move plan. There is no charge per extra run, so a busy week cannot produce a surprise invoice.

7. Questions

Write to maximiliano.dumont@gmail.com. If you want a human to look at something the model produced about you, including any run or letter that automatic applying sent in your name, ask and one will: that review is the human intervention we promise, and you can tell us your view of what it did and contest it through the same address.

Legal