> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hr-easy.nlead.ch/llms.txt
> Use this file to discover all available pages before exploring further.

# AI transparency

> What applicants are told, how consent and opt-out work, and why there are no cut-off scores.

AI in recruitment touches people who have no relationship with the organization yet and
no leverage. The platform's position on that is written down and shipped as a
transparency notice.

## The application fit assessment

When a recruiter requests it, the content of an application is sent to a large language
model, which returns three things:

| Output                | What it is                                                                                                                                          |
| --------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- |
| Summary               | A short narrative — at most five sentences — describing how the application relates to the position, with relevant strengths and development areas. |
| Fit score             | A whole number from **0 to 10** indicating how closely the application matches the stated requirements.                                             |
| Language observations | A count of spelling and grammar issues in text the applicant wrote.                                                                                 |

The score bands are documented and shown **next to the number** in the recruiter
interface, so a "6" is never presented as a bare figure whose meaning the reader has to
assume.

## The guarantees

<AccordionGroup>
  <Accordion title="It is decision support, never a decision">
    The score is one input among many. Every hiring decision is made by a person.
  </Accordion>

  <Accordion title="There are no cut-off scores">
    No application is filtered out, hidden, or auto-rejected because of its score. Every application
    stays fully visible to the recruiting team. There is no threshold to configure, because there is
    no threshold.
  </Accordion>

  <Accordion title="Applicants can opt out — and change their mind">
    Opting out is offered when applying and can be changed at any time from the applicant portal.
  </Accordion>

  <Accordion title="Opting out is invisible to the recruiter">
    Recruiters see **"Not available"** — the same wording shown when a score is missing for any
    other reason. An opt-out cannot be read as a signal, because it cannot be distinguished from an
    absent score.
  </Accordion>

  <Accordion title="Withdrawing consent deletes the output">
    If consent is withdrawn after a summary was produced, the summary **and** the search index entry
    derived from the application are deleted.
  </Accordion>
</AccordionGroup>

That fourth point is the one that is easy to get wrong. A system that shows "applicant
declined AI processing" has made opting out costly, which makes the consent
meaningless.

## Optional per installation

Swiss HR Easy is deployed per customer, and the AI features are optional. Where a
feature is not configured, the parts of the notice describing it do not apply.

The operating organization is the **data controller** for its own instance and remains
responsible for how it configures and uses these features.

## The notice itself

A full transparency notice is maintained in the repository, written to be read by an
applicant rather than by a lawyer: what the systems do, what data they use, who
processes it, what the outputs are used for, and how to decline.

<Warning>
  The notice is currently a **draft and has not been reviewed by legal counsel**. Two sections are
  knowingly incomplete and marked as open: the retention period for AI outputs, and an independent
  bias evaluation. Do not publish it to applicants until those are resolved and a data-protection
  advisor has reviewed the whole document.
</Warning>

## The broader rule

Across the whole product, not just recruitment:

> AI rankings must always show the criteria, the data used, and the user retains the
> decision.

That is [ADR-006](/concepts/design-philosophy), and it is why the
[survey report](/guides/people/surveys#the-ai-report) states its data basis, why
[sentiment](/guides/ai/semantic-search#2-sentiment) is a signal rather than a rating,
and why [receipt OCR](/guides/finance/expenses) shows the original image next to the
extracted figures.

## Auditing

Every AI operation is audited as `AI_OPERATION` with token usage. If someone asks what
was processed, when, and on whose instruction, the answer is a query.