8 practical examples of AI in HR (with prompts)

Most content about AI in HR sends you to one of two dead ends: vendor hype that promises the moon or fear pieces about robots coming for your job. Neither tells you what you would actually open on your screen Monday morning.

So here are eight concrete examples of AI in HR you can picture yourself running. Each one is a real workflow rather than a vague category, with a practical tool, framework, or prompt you can use instead of another topic to research on your own.

The through-line is simple. AI already does useful HR work when it handles the mechanical parts—screening, sorting, scheduling, scoring, and drafting—and hands you better information to make the call. The examples that break are the ones that try to take you out of the decision.

None of this is future tense. In organizations that have implemented AI, HR professionals report using it frequently: 26% weekly, 20% daily, and 9% several times a day. Every example below follows the same rule: you stay responsible for the decision while the tool does the groundwork.

Examples of AI in HR

1. Screen and shortlist high-volume applicants

The everyday version of AI screening is easy to picture. Instead of skimming 200 résumés from scratch, you define the role-specific criteria that matter, ask the tool to rank candidates against them, and require a reason for every proposed score.

Rank these candidates against the following criteria for a [role] position: [must-have skill], [relevant experience], and [key responsibility]. For each candidate, give a fit score out of 10 and a one-sentence reason based only on information in the application. Do not infer or comment on age, gender, ethnicity, disability, or any other protected characteristic.

Then review both the shortlist and a sample of candidates the tool ranked lower. That second check matters because screening at scale can multiply historical bias as easily as it multiplies speed.

If you use an automated employment decision tool in New York City, Local Law 144 sets three clear obligations:

Federal antidiscrimination law still applies when an employer uses an automated system. The EEOC notes that the familiar four-fifths selection-rate test can flag potential adverse impact, but passing it does not automatically establish that a process is lawful.

Jenelle Buatti, an instructor in Ziplines’ AI for HR Professionals course, puts the operating principle plainly:

“We cannot outsource accountability.”

Best for high-volume requisitions where sheer volume is the bottleneck. Skip it—or keep the pilot very limited—if you cannot validate the tool, review its outcomes, and meet the requirements that apply where you hire.

2. Draft performance reviews in minutes

If you write 10 to 15 performance reviews in a cycle, the drudgery is often the reformatting: turning scattered notes into the structure your review template demands. A performance-review copilot can take that part off your plate.

The workflow is quick to picture. Enter your role and the reviewee’s role, pull up your standard questions, then dictate or paste anonymized notes about specific projects and behaviors. The tool maps those notes into the format your process requires. You still edit for accuracy, fairness, and voice before anything enters the performance system.

I manage a [role]. Here are my anonymized notes on their work this cycle: [paste notes]. Format these into answers to our standard review questions: [paste questions]. Keep the language specific and evidence-based. Cite the actual project or behavior behind each statement, and do not add information that is not in my notes.

That editing step is not a nicety. Hallucinations—confident statements unsupported by the input—are a documented failure mode for language models, and their frequency varies by model, task, and benchmark. The course teaches a three-step review for exactly this: check the formatting, inspect the facts and completeness, then correct or re-prompt rather than accepting draft one.

There is a second risk that has nothing to do with factual accuracy: employees may experience AI-assisted reviews as surveillance or impersonal decision-making. Be clear about what the tool does and does not do. The same discipline applies when a review feeds a calibration meeting or performance improvement plan: AI can organize your evidence, but the rating and its consequences are yours to own.

The verdict: AI writes the first draft, and you write the review that gets submitted. Read every line and make it yours before it enters the system.

3. Check and address pay-equity patterns

Pay equity is one of the highest-stakes analyses in HR, and it is one that many lists ignore entirely. That gap is worth closing because the analysis can turn a broad concern into a specific question your compensation, analytics, and legal teams can investigate.

The mechanics are more useful than a flashy headline. A pay-equity analysis accounts for legitimate drivers of compensation—such as role, level, location, tenure, and relevant performance measures—and isolates unexplained differences among comparable groups. That unexplained variation does not prove why a gap exists. It tells you where qualified analysts, HR leaders, and counsel may need to look more closely.

The older model was a periodic, point-in-time audit: pull the data once a year and address what you find. Newer tools can monitor patterns between formal audits, helping compensation teams identify where gaps may be widening. Either way, the result has to rest on a defensible method and data you can explain.

One hard guardrail: compensation data is among the most sensitive information HR touches. Exact salaries and identifiable employee records never go into a public AI tool. Use an approved environment and keep exploratory work anonymized and aggregated.

This anonymized compensation dataset includes role, level, location, tenure, performance rating, and current pay. Identify pay patterns by department and level after accounting for the listed factors. Flag areas that require further statistical review. Do not identify or make recommendations about individual employees.

Our recommendation: start with a single anonymized, point-in-time analysis to determine whether a material unexplained gap exists before investing in continuous monitoring. Confirm the problem, validate the method, and then build the case for the spend.

4. Spot attrition and engagement patterns without a data team

Can an HR generalist with a messy HRIS export and no analyst answer a real workforce question? Yes - if the question is specific and the data is handled responsibly.

Say you want to test a hypothesis: are remote employees less engaged but more productive than their on-site peers? Strip the personal identifiers from your dataset, then ask an AI assistant which of the remaining columns—engagement score, productivity score, work-arrangement type, or attrition-risk flag - you need to answer it.

This anonymized dataset includes engagement score, productivity score, work-arrangement type, and attrition-risk flag. No names or IDs are included. Compare remote, hybrid, and on-site employees on these metrics and flag any group with a materially different pattern. Frame each finding as Who (the segment), What (the metrics), and Gap (the measured difference). Do not make predictions about individuals.

In one course exercise, remote CAD technicians showed the highest productivity of the three groups but engagement scores nearly 40% lower than their hybrid and on-site peers. That turns a vague concern into a precise policy question. It does not tell you what caused the pattern or what to do about one employee.

AI enablement strategist Chiara Barden, who teaches the course’s data-analysis module, calls the framing method the “insight equation”: Who is the specific segment, What are the key metrics shown side by side, and what is the Gap or business impact?

“Quantifying what you’re saying is always stronger than just saying a broader statement.”

The ethical line matters here. AI surfaces segment-level patterns across a group. It should never output a prediction about one named person. As Barden puts it:

“AI supports HR judgment, and it analyzes patterns, not people.”

Five categories of data should never go into an external tool:

Best for testing a specific hypothesis with anonymized, aggregated data. Stop the moment you would be making a prediction or decision about one named person.

5. Personalize onboarding and clear the admin backlog

Onboarding is where repetitive admin piles up fastest, and it is also where AI and automation can give back some of the most obvious hours. Instead of handing every new hire the same generic checklist, you can generate one tailored to the person’s role, department, and location, pair it with a scoped companion for policy questions, and track who has completed what and where people are stuck.

The time savings can be substantial. IBM reports that its internal AskHR system automates more than 80 HR tasks. In 2024, it handled more than 11.5 million employee interactions and achieved a 94% containment rate for common questions. IBM also documented productivity gains in specific HR tasks between 2022 and 2024, with some improving by as much as 75%.

Even a narrow workflow adds up. In one internal course case study, connecting account provisioning across systems reduced a recurring onboarding task from about 90 minutes to roughly four minutes. The reclaimed time went back into designing the employee experience rather than clicking through forms.

To decide what is safe to automate, use the RIPE test. Score a task from 1 to 3 on four dimensions:

• Repetitive: Does it happen often in the same way?
• If-then: Can you state the rules clearly?
• Predictable: Are the inputs and outcomes consistent?
• Error-prone: Would automation reduce manual mistakes?

A score of 10 to 12 makes the task a strong automation candidate. A score of 7 to 9 is the judgment zone, where you should pilot first. Below that, keep it human.

Onboarding maps cleanly onto the Automate, Augment, Extend model: automate the checklist and provisioning, and protect the human welcome. A 30-60-90-day plan and a warm first week are where retention starts, and no dashboard replaces them. Watch time-to-productivity as the real metric, because the sooner a new hire has access and a clear plan, the sooner they can contribute.

The verdict: let automation clear the paperwork and account setup, and protect the first-week welcome as human time.

6. Draft job descriptions and policy communications

If you try only one example on this list this week, make it this one. Drafting job descriptions, interview rubrics, and policy communications is creative work with a clear right answer in your head—which is exactly what generative AI is good at speeding up while you stay the editor.

The difference between mush and a usable draft is the prompt. A weak prompt - “Write a job description for a marketing manager” - gets you generic filler. A structured one gets you something you would send. Use the CRAFT pattern:

• Context: Give background instead of a bare command.
• Role: Assign the AI a relevant role, such as an HR compliance specialist.
• Action: Break a multi-step request into ordered steps.
• Format: Specify the output structure you want.
• Tone: Match the language to the audience and purpose.

Context: We are hiring a mid-level marketing manager for a 15-person team at a B2B software company. 
Role: Act as an HR compliance specialist drafting a job description.
Action: Draft the description, then list five structured interview questions tied to the role's core responsibilities.
Format: Use Summary, Responsibilities, Requirements, and Nice-to-haves.
Tone: Professional and direct. Avoid buzzwords such as “rockstar” and “ninja.”

So instead of “write a performance improvement plan,” anonymize the scenario as a request about “a sales director,” assign the compliance-specialist role, and specify the sections you need. The output arrives usable rather than generic. Save that pattern once, and every future job description, rubric, or change memo starts with a structured prompt instead of a blank page.

Two guardrails keep this safe. Never paste employee names or other identifiers into the prompt. And run every draft through Check, Inspect, Correct before it is sent, because generative text can confidently invent policy details and regulatory citations that do not exist.

Our recommendation: start here. It is one of the lowest-risk examples on the list, it pays off quickly, and it builds the prompt discipline every other example depends on.

7. Answer routine questions with a scoped companion

Your team answers the same benefits and policy questions dozens of times a week, and candidates may wait days for an application-status update that could be delivered instantly. Both are tier-one questions, and both can be a good fit for a scoped companion.

The distinction between a companion and a generic chatbot is the whole point. A companion is narrow and guardrailed. A benefits companion might answer plan-summary, employer-match, and contribution-limit questions strictly from approved documents, with hard rules against giving legal or tax advice or personal recommendations. It should point to the source document and hand off the moment a question crosses its boundary.

You are a benefits companion. Answer only questions about plan summaries, employer match, and contribution limits, using the attached approved plan documents. Do not give legal or tax advice or personal recommendations. If a question falls outside this scope, respond: “That is outside what I can help with. Let me connect you with someone on the HR team.” 
Then stop.

The candidate-facing version handles FAQs, interview scheduling, and application-status updates instantly rather than leaving people waiting. The payoff is biggest in high-volume hiring, where one recruiter may be juggling hundreds of active candidates and status questions alone can eat a morning.

Watch the deflection rate—the share of questions resolved correctly without a person—but review the failure cases too. Those cases, not the desire for a higher number, should determine what is safe to bring into scope next.

The risk is drift. Without deliberate scope and guardrail design, the tool starts answering things it should not, and it must be kept free of personal data. Present it to employees and candidates as a support resource that routes real decisions to a person, and give it clear in-scope and out-of-scope boundaries.

Best for high-volume, policy-bounded questions with consistent answers. Skip it for anything advice-adjacent, personal, or legally consequential, and route those questions to a person.

8. Score interviews and assessments

What if you did not have to rewatch a two-hour interview to find the three moments that mattered? AI can transcribe a live or recorded interview, organize evidence against the criteria your hiring team defined upfront, surface the relevant moments, and generate a summary.

Score this interview transcript against the following rubric: [criterion 1], [criterion 2], and [criterion 3]. For each proposed score, quote or timestamp the specific evidence in the transcript. Do not infer anything about the candidate that is not directly stated. Flag any criterion for which the transcript contains insufficient evidence.

You still review the interview, but you know where to look first rather than starting from scratch. Because it is a structured interview, every candidate is scored against the same job-related rubric, which is easier to explain later than a panel’s scattered impressions.

There is real potential to make the process more consistent. There is also real risk if the model invents evidence, rewards irrelevant language patterns, or becomes the decision-maker. Validate the tool, check outcomes for adverse impact, obtain any required recording consent, protect the transcript, and keep a person responsible for the final call.

This is the cleanest illustration of the rule running through every example here: AI does the mechanical scoring and tells you where to look, and you decide who gets hired.

FAQs about AI in HR

The practical value of AI depends on the boundaries around it. These answers cover the disclosure, job-impact, data, and accountability questions that should shape your first pilot.

Do I have to tell candidates or employees that AI was used in a decision about them?

It depends on the tool, location, and use case, so check the rules that apply before deployment. New York City’s Local Law 144, for example, requires notice at least 10 business days before certain automated employment decision tools are used. Even where a specific notice rule does not apply, transparent communication can build trust. In a Pew Research Center survey, 66% of U.S. adults said they would not want to apply for a job if the employer used AI to help make hiring decisions.

Is AI going to replace my HR job?

The honest answer is that AI redistributes HR work rather than eliminating the need for HR judgment. It can take over mechanical, high-volume tasks such as sorting, scheduling, and first-draft writing. Decisions about hiring, pay, performance, and discipline still need responsible people who understand the context and consequences.

Which AI use case should I try first?

Score your options instead of guessing. Rate each one on impact, feasibility, ethics, and time to value. For automation, use the RIPE test: Repetitive, If-then, Predictable, and Error-prone. A lower-risk augmentation task such as drafting job descriptions or rubrics usually makes a better first pilot than automated decision-making.

What data is safe to paste into a public AI tool such as ChatGPT?

Strip out anything personally identifiable or confidential first. Never paste names or employee IDs, email addresses, exact salaries, dates tied to individuals, medical or accommodation information, or data from small identifiable groups into a public tool. Use anonymized, aggregated data and an approved enterprise environment that meets your organization’s privacy and security requirements.

Who is responsible if an AI-assisted decision is wrong or biased?

Legal responsibility varies by jurisdiction, contract, and use case. Operationally, your organization should act as though it retains decision ownership. Document the process, validate the tool, monitor outcomes, preserve meaningful human review, and involve legal counsel in consequential uses. A vendor recommendation is not a substitute for your organization’s judgment or obligations.