What our AI does and does not do

What it does

The assistant reads Crewmark as you, the signed-in user: it sees what your permissions let you see, never more. It only reads. It cannot change, send or save anything — a draft is text that you copy yourself.

• Answers how-to questions from Crewmark's help center and lists the articles it used as sources.
• Finds a page: you get a button to it, if your role may open it.
• Summarises your open work: tasks, approvals, interviews, rejection notices due, open forms, reviews, probation periods about to end, documents waiting for a signature and open data requests.
• Helps rephrase a text you wrote, such as a review comment, feedback or a rejection email.
• Drafts a job ad from your notes.
• Says so when the help center does not cover your question, and tells you whom to ask.

What it never does

The assistant refuses these requests, says why, and offers what it can do instead.

• Rank, score or compare candidates or employees.
• Suggest a rating, a bonus or a probation outcome, or recommend whom to hire.
• Predict who will leave.
• Say how someone feels. It does not infer emotions.
• Infer special-category data, such as health, religion or trade union membership.

Why the line is drawn there

Under the EU AI Act, AI that ranks or evaluates candidates and employees, or suggests decisions about them, is a high-risk system, and emotion recognition at work is prohibited. Inferring special-category data would have no legal basis under the GDPR. So in Crewmark, decisions about people are made by people. The rule is applied twice: a request is checked before it reaches the model, and an answer that judges someone anyway is withdrawn.

You always know it is AI

The assistant's panel opens with a notice that you are talking to an AI assistant and not a person. Every text it writes for you carries the label “AI-generated draft — check before use”.

Questions about your numbers

In Analytics you can ask a question in plain language. The AI does not read your database. It chooses from Crewmark's fixed list of metrics, and Crewmark runs that choice with your permissions — the same numbers the dashboards show, with the same small groups hidden. Your questions are visible only to you and are deleted after 30 days.

• What the AI gets with your question: the list of metrics, the names of your departments and job categories, the date and your language.
• What it gets back to describe: totals and averages for groups, with small groups already hidden — a hidden value arrives without a number and without a count of people.
• It never gets a name, an email address or a row about one person. Pay figures are never sent to the AI provider — no salary, no median, no pay gap — even when the person asking may see salaries. Such an answer is shown with a summary built by fixed rules, and it says so.
• A question about a named person, a request to rank people, managers or interviewers, or a prediction is refused before anything is sent. You are offered a number about a group instead.
• A summary containing a number that is not in the result is discarded, and a sentence built by fixed rules is shown in its place.
• Without an AI provider, or with the assistant switched off, the question is matched by keywords and nothing leaves the server. Every answer says which way it was produced.
• The dashboards, the “Worth a look” findings and the email digest are computed by fixed rules. With AI on, a short summary of those findings may be added, with the same check of every number.

What goes to the model provider

When AI is switched on, a question is sent to the AI provider, which acts as a sub-processor named in the data processing agreement. Under the provider's commercial terms, what is sent is not used to train its models.

• Your question, any text you paste, and the earlier messages of the same conversation.
• The assistant's instructions and Crewmark's help center — the same for every company.
• Today's date, your interface language, the page you are on, and whom to ask when the answer is not known: the name of your company's Crewmark administrator and the support address.
• Only when you ask about your open work or look someone up: a short list of matching records with names, titles, dates and links.
• Text you paste is handled as text to work on, never as instructions to follow.

What never goes there

Some data is kept away from the assistant altogether, whatever your permissions in Crewmark.

• Your name and email address, and the name of your company.
• Salaries, scores, bonuses, notes and review texts.
• Contact details and documents.

Your conversations

A conversation is visible only to the person who wrote it — to nobody else in the company, administrators included. Messages are deleted after 30 days, and you can delete a conversation yourself before that. Your conversations are part of the copy of your data when one is requested.

Switching it off, and working without AI

A company can switch the assistant off for everyone in its settings. Where no AI provider is configured, the assistant works as plain search: it shows the best-matching passage from the help center with its source, and nothing leaves the server. The help articles are written in English for now — with AI the answer comes in your language; in search-only mode the passage is shown in English.