What is Automated Employment Decision Tool?

Why It Matters

The category exists to stop an argument about terminology. A tool that ranks candidates is making an employment decision, whatever the vendor calls it, and the question that follows is not whether it counts as AI but whether it replaces or materially influences a person’s judgement. If it does, the duties attach.

New York City put the first working definition into law in 2021 and began enforcing it in 2023, requiring a bias audit, a published summary, and candidate notice before use. Other jurisdictions have since built on the same category under names like automated decision-making technology, which is why the scope question is now the first compliance question in the sector.

What Counts

The test is functional, not technical. Resume parsing that filters, an assessment that scores, an interview analyser that rates, a ranking model that orders applicants, and a chatbot that screens out unqualified candidates all sit in scope. So does a system that recommends rather than decides, if the recommendation is followed as a matter of course.

What tends to fall outside is work that never touches a decision about a person: drafting a job advertisement, scheduling an interview once someone has been chosen, or answering a benefits question from a handbook.

Where It Breaks

The first failure is the inventory. Screening capability usually arrives inside a platform the organisation already runs, switched on by a checkbox or bundled in an upgrade, so the tool is in production before anyone asks what category it belongs to. Compliance starts with an accurate list.

The second is the claim of non-use. An applicant tracking system that sorts applications by relevance is making a screening judgement, and “we do not use AI for hiring” stops being a defensible position the moment a candidate asks how the ranking was produced.

The third is trusting the vendor’s assurance. A supplier can document its own system, and it cannot assess how the tool is used inside your process, which roles it screens, or what your criteria are. Those are deployer questions.

The fourth is scope creep. A feature added later, an assessment module enabled for one requisition, a new model version behind the same interface, can each change what the system does without changing its name. The inventory has to be rechecked on a cadence rather than compiled once.

How Flytebit Handles It

We start by mapping what each tool actually does against the categories, including the capability that arrived inside a platform, and we write the classification down with the reasoning. Where a system is in scope, the obligations are built into the workflow rather than assembled afterwards: criteria in versioned configuration, evidence attached per judgement, a named reviewer, notices and retention as properties of the run, and impact ratios computed from live traffic. The industry application is on our HR & Workforce Technology page, and the control design is our AI governance and risk work.