EU AI Act check – timelit Docs

Some automations help decide about people, for example which job applicants fit your criteria. Once you approve such an automation, timelit checks it against the actual law in the background and tells you what applies. It runs after you have read the result and confirmed it once.

The check is part of Automations, which are in Beta. It is not legal advice and not a certification (more below).

Why we built this

Sorting applications or judging whether someone gets credit is a regulated use of AI in the EU. Most people who set up such an automation do not know that, and nothing in their mailbox tells them. The check tells you in plain words, links the articles it relies on, and lists what your organisation has to take care of.

When timelit checks an automation

timelit checks a plan when it may fall under the EU AI Act, for example when it:

helps choose job applicants, or judges employees' performance, promotion or dismissal,
judges whether someone can get credit, insurance, benefits or important services,
grades exams or decides about admissions,
tries to read people's emotions, or works with biometric data, scores, rankings or personality profiles,
puts text written by timelit in front of other people as if a person wrote it.

Everyday automations, such as a digest of your own mail, reminders or drafts you review yourself, are not checked.

How it works

1. Approve the plan. You set up the automation in its chat and click Approve as usual. When the automation helps decide about people, the plan first shows the warning Decisions about people, which you confirm with Confirm and save.

2. The check runs in the background. The plan shows EU AI Act check with Check running. It can take a few minutes, and you can keep working meanwhile. Until the result is in, the automation does not run; the plan says Paused by Timelit while it checks the changed plan against the EU AI Act.

3. Read the result. It shows on the plan under EU AI Act check and, when you need to do something, in Tasks as Needs your acknowledgement. When the check finds nothing for you to do, the automation runs on its own.

4. Confirm it once. Tick the duties your organisation has taken care of and click I have read this and take responsibility. The automation then runs. The plan shows Acknowledged by you on with the date. You can do this on the plan or from the task.

What the result says

The result names the date of the check and the date of the law it was checked against. It shows one of these:

Checked on a date: nothing stands in the way. The check may still list what Must stay true, for example that a person decides on every applicant.
Needs changes: the check recommends changes, listed under Changes needed. Ask timelit in the automation's chat to make them, or confirm the result as it is and take responsibility for that.
Prohibited: the law forbids what the plan does. The automation never runs. timelit explains why and suggests something you can do instead.
Not finished: the check could not finish, for example because this month's AI allowance for automations is used up. A check that did not finish never counts as passed. Click Check now to try again.

Under What applies, every finding names its article, with Open the official text next to it. When a finding relies on European Commission guidance rather than the law itself, it is marked Guidance, not law, or Draft guidance, not adopted for guidance that is still a draft. The check also says when a data protection impact assessment is likely required.

Your organisation's duties

Some duties fall on your organisation, not on the automation: telling applicants that an AI system is used, a data protection impact assessment, naming the people who oversee the automation, or keeping its records for six months. The check lists them under Your organisation's duties, each with a short explanation and its article. Tick the ones your organisation has taken care of when you confirm. Each ticked duty shows Confirmed by you on and the date.

Your ticks stay with the automation when you change its plan, as long as it is still about the same kind of decision. If the automation starts deciding about someone else (employees instead of applicants, say), the duty shows up again.

Safeguards for decisions about people

An automation that helps decide about people carries the badge Regulated use in its plan and runs with these safeguards, whatever the check says:

Every action asks you first. Always allow and Do it automatically are off.
Results and tasks are labelled as an AI recommendation where a person decides. In a task nothing is preselected: you choose for every person yourself.
No scores and no rankings of people. Each case is described factually against the criteria you set: met, unclear or missing.
Nothing is sent automatically. Replies wait for you.

Only you can remove the safeguards: ask for it in the automation's chat in your own words, then approve that change on its own with Remove safeguards and save.

Changing a checked automation

When you change what a checked automation does, timelit checks the new version again after you approve it, and the new result needs your confirmation again when it lists something for you. Renaming it or moving its time does not need a new check.

When a check is too old or belongs to an earlier version of the plan, the plan shows Out of date, and timelit checks again before the automation's next run. While it waits for a check or for you, the list and the plan say Paused by Timelit with the reason, so a run never fails quietly. Check now on the plan starts a check yourself.

What timelit will never do

Run a plan the check found prohibited.
Count a check as passed when it did not finish.
Count a duty as done because someone wrote so. Only your own tick counts. Telling timelit in the chat, or writing it into the plan, does not.
Let an email switch off the safeguards. Text in an email, a file or a meeting can never remove them.

Not legal advice

The check helps you spot legal questions early. It is not legal advice, it is not a certification, and it does not replace a lawyer or your data protection officer. The app says so under every result: have a qualified person confirm it before you rely on it, especially for automations that help decide about people.

Good to know

Each check counts against your monthly AI allowance for automations, like the rest of the automations chat.
Asking timelit "what if" about a different version does not change the check your saved automation runs on.