Map HR AI use cases

HR is not one task. It is a chain that can include workforce planning, writing a role description, advertising, sourcing, application intake, screening, interviewing, selection, onboarding, learning, performance support, employee service, internal mobility, and offboarding. A model used at one stage can influence every later stage. If a generated job description contains an unnecessary requirement, screening may exclude capable people before an interview occurs.

Map each workflow before adding AI:

1. What outcome is the organization trying to achieve?
2. Which inputs are authoritative, and who maintains them?
3. Which steps are administrative, advisory, or decisional?
4. Who is affected by an error?
5. Can a person inspect, correct, and appeal the result?
6. Which laws, policies, contracts, or consultation duties apply?

Do not treat “HR AI” as a single risk category. Formatting an approved policy into a FAQ is not equivalent to ranking employees for redundancy. Consequence, data sensitivity, scale, and reversibility matter.

A practical map uses four groups.

Drafting and transformation includes rewriting a job advertisement in plain language, translating an approved onboarding guide, summarizing survey comments, or turning interview notes into a structured draft. These uses can be valuable, but outputs still require review for accuracy, tone, omission, and confidentiality.

Retrieval and service support includes helping employees find a leave policy or giving recruiters the current interview template. The system should cite the governing source, show its effective date, and route uncertain or personal questions to a qualified person. A fluent answer is not a policy ruling.

Analysis and recommendation support includes identifying evidence related to published criteria, spotting inconsistent interview coverage, or aggregating workforce trends. These uses need stronger validation because the output can shape judgment even when labeled “advice.”

Decisions and actions includes rejecting a candidate, changing pay, assigning a performance rating, initiating discipline, or selecting someone for promotion. These are consequential actions. A model should not receive authority merely because a human can theoretically override it. The human gate must be informed, timely, and real.

Good early uses usually have bounded inputs, inspectable outputs, reversible consequences, and a responsible reviewer. Examples include:

• drafting alternative wording from an approved role brief;
• creating interview questions tied to job-related criteria;
• checking whether an interview pack covers every competency;
• summarizing non-sensitive training feedback with source references;
• preparing a first draft of a routine internal announcement;
• classifying employee queries for routing without deciding entitlement.

Avoid beginning with emotion inference, personality prediction from video or voice, hidden monitoring, health inference, or opaque ranking. These methods may lack a sound connection to job performance and can create serious legal, scientific, and dignity concerns.

Before a pilot, create a short card:

• Purpose: the legitimate HR outcome.
• Users: who operates and reviews the tool.
• Affected people: candidates, employees, contractors, or managers.
• Inputs: exact data fields and their sources.
• Output: draft, evidence extract, recommendation, or action.
• Prohibited use: what the output must never determine.
• Human gate: who checks what, and when.
• Failure response: how work continues safely if the tool is wrong or unavailable.
• Retention and access: where prompts, outputs, and logs go.
• Measures: quality, fairness, privacy, time, and user experience.

“Improve recruiting” is not a sufficient purpose. “Create a recruiter-reviewed evidence worksheet against five published essential criteria” is testable. It defines an output without silently authorizing rejection.

Time saved matters, but it can hide transferred work and unequal errors. Measure whether reviewers correct outputs, whether qualified candidates are missed, whether recommendations differ across relevant groups, whether employees understand how the system is used, and whether people can obtain a correction. Include false positives and false negatives. An assistant that flags every application may achieve high recall while saving no work.

Compare the pilot with a realistic baseline. If human reviewers already disagree, document that disagreement rather than presenting it as AI failure alone. The goal is a better sociotechnical process: criteria, people, tools, records, and recourse working together.