Map the Function Before You Judge the Risk

“AI in HR” is a phrase that covers everything from a chatbot answering benefits questions to a model ranking a thousand applicants, and lumping them together makes the field impossible to reason about. Here’s a grounded map of where AI is actually doing work across the employee lifecycle in 2026 — and, just as important, where it isn’t allowed near the decision.

Sourcing and outreach: automation before the application

Finding and contacting candidates is heavily automated, but it is not consequence-free: a discovery or matching system can shape who ever sees an opportunity.

• Candidate sourcing and talent discovery. Tools search public profiles, resume databases, prior applicants, and internal talent pools to recommend people for a role.
• Job-description writing. Language models draft and rewrite postings, and flag biased or exclusionary wording (“rockstar,” “young and energetic”).
• Chat-based pre-qualification. Recruiting chatbots ask basic screening questions and book interviews, especially in high-volume hourly hiring.

Search, matching, ranking, and message drafting are distinct functions. A generated outreach email is not equivalent to a model deciding which people are discoverable.

Screening and assessment: automated, and legally load-bearing

This is where AI can materially influence who advances and where many legal and governance questions become consequential.

• Resume ranking at high-volume employers, ordering applicants for review.
• Scored assessments and AI-assisted interviews, including transcript analysis and newer two-way voice interviews.
• Knockout and matching logic on application forms.

Some of these systems can meet legal definitions such as New York City’s AEDT definition, but not every automated hiring feature does. Coverage depends on the function, jurisdiction, and how the output is weighted. Human review also needs to be real: a rubber stamp does not reveal whether the system effectively determined the outcome.

Scheduling and coordination: quietly the biggest time-saver

Schedulers negotiate interview times across calendars, send reminders, and coordinate panels. This is operationally different from ranking applicants, though accessibility, data handling, and faulty automation can still matter.

Onboarding and HR support: the chatbot era

Once someone’s hired, AI mostly shows up as an internal help desk.

• HR chatbots / knowledge assistants answer “how much PTO do I have,” “how do I change my 401(k),” “what’s the parental leave policy” — deflecting routine tickets from HR staff.
• Onboarding workflows auto-provision accounts, route paperwork, and nudge managers through checklists.
• Document drafting — offer letters, policy summaries — from templates.

The risk here is accuracy, not discrimination: a benefits chatbot that confidently states the wrong policy is a real problem, which is why the good deployments cite the source document and route uncertain questions to a person.

Performance, retention, and comp: proceed carefully

• Attrition/flight-risk prediction flags employees likely to leave. Useful signal, but acting on it clumsily — or letting it shade decisions about pay and promotion — reproduces the same proxy-bias problems as hiring models.
• Skills inference and internal mobility matches employees to internal openings and learning.
• Comp benchmarking analyzes market data to set ranges.

These touch high-stakes decisions about existing employees, so the same rule applies: model as input, human as decider, and watch for disparate impact.

What AI still doesn’t (and shouldn’t) do alone

• Final hiring and firing decisions. Both for legal exposure and because these systems inherit historical bias.
• Anything requiring genuine judgment about context — a candidate’s unusual path, a sensitive employee-relations issue, a real accommodation.
• Substituting for the bias audit and notice the law now requires when a tool does touch a decision.

The useful 2026 summary is functional: recordkeeping, drafting, search, scheduling, ranking, scoring, recommending, and deciding are not one category. Document which function a system performs, which people it affects, and how its output changes the workflow before assigning a risk label or compliance rule.