How to use AI in HR: a practical starting guide for people who already know HR

You are not behind. According to SHRM's State of AI in HR 2026 report, 39% of organizations have adopted AI in HR, while 54% have not adopted it and do not plan to do so this year.

That leaves room to learn deliberately. The advantage HR professionals bring to AI is judgment: knowing when an output is wrong, when a policy statement needs verification, and when a decision needs a person in the room.

This guide lays out a practical sequence. Start with low-risk drafting, build a review habit, protect employee data, turn one useful task into a workflow, write a short policy, and test it with a small group. Save decision-adjacent work for last, after the guardrails are already in place.

Step 1: Start with low-risk drafting and summarizing

Pick one task you will do this week anyway, and let AI take the first pass.

A good starting task clears three checks:

• The prompt contains no personal or confidential data.
• The output is a draft, not a decision.
• A person reviews it before anyone relies on it.

Where to start with AI in HR

Job descriptions are a proven place to start. In SHRM's 2025 Talent Trends research, writing job descriptions was the most common recruiting use of AI among adopting organizations, and 89% of HR professionals using AI in recruiting said it saved time or increased efficiency.

A useful way to sort potential tasks is to ask whether AI will automate, augment, or extend the work. Scheduling and reminders are candidates for automation. Drafts, summaries, and rubrics are augmentation: AI prepares material and a person improves it. Cross-referencing structured skills data across a team can extend the analysis HR has time to perform.

Try this prompt on a real document today:

The same pattern works for anonymized exit-interview themes, offer-letter templates, and behavioral interview questions. One useful workflow is a better starting point than ten disconnected experiments.

[[cta-big]]

Step 2: Review every output like a fact-checker

Treat every AI output as a first draft. Check it against the source material before it reaches an employee, candidate, manager, or decision-maker.

That review matters because AI can return confident answers that are incomplete, outdated, or influenced by the way information is ordered. A 2025 study of ChatGPT as a resume screener found a strong preference for the first resume presented when candidates appeared equally qualified. The researchers ran 2,800 resumes through the model and found that prompt changes reduced one form of position bias but introduced another.

Historical screening systems show how biased patterns can enter automated recommendations. Amazon abandoned an experimental recruiting tool after finding that it penalized resumes containing terms associated with women, according to Reuters' original reporting.

Use a simple review loop: check, inspect, correct.

• Check: Does anything look inconsistent or implausible?
• Inspect: Is the output accurate, complete, and appropriate for the audience?
• Correct: Edit the draft or send it back with more precise instructions.

Before an output touches a person:

• Verify legal and compliance claims against current official guidance or counsel.
• Scan for coded or exclusionary language such as “digital native” or “recent graduate.”
• Reorder inputs and rerun any comparison to test for position bias.
• Check the tone against your organization's actual voice.
• Confirm that the prompt contained no prohibited data.

If the model seems too agreeable, ask for a skeptical review. For example: “Review this announcement from the perspective of an employee who opposes the change. Identify unclear claims, missing context, and likely objections.”

The risk is not theoretical. Mobley v. Workday alleges that automated screening discriminated against Black, disabled, and older applicants; the litigation was still active in 2026, so it should be described as an allegation rather than a finding. Court records show the case and its discrimination claims remain ongoing.

Step 3: Keep employee and candidate data safe

Do not put identifiable employee, candidate, or confidential company data into a tool your organization has not approved.

Data privacy is already a leading HR concern. In a survey of more than 500 HR professionals, 63% named data privacy and security as a top concern.

What to keep out of unapproved AI tools

Anonymization requires more than deleting names. A team of three people may still be identifiable from job titles, location, dates, or demographic details. If the remaining context points to one person, treat the information as identifiable.

Understand the data settings of the specific tool and plan you use. In personal ChatGPT workspaces, model-improvement sharing is enabled by default but can be turned off. OpenAI says content from ChatGPT Business, Enterprise, Edu, and its API is not used to train models by default. Other vendors have different terms, retention periods, and controls.

For real HR data, use only an approved business or enterprise environment under your organization's security, retention, access, and contractual requirements. Turning off a training setting in a personal account does not make that account an approved HR system.

Step 4: Turn one task into a repeatable workflow

Once a task proves useful, document how it works.

The gap between a good experiment and a durable workflow is usually not the model. It is whether the team has written down the tool, inputs, review steps, and quality standard. SHRM found that only 16% of HR professionals use their own ROI metric for AI investments, while 56% do not formally measure AI investment success.

Document four things:

1. The approved tool and account type.
2. The information that must never be entered.
3. The person responsible for reviewing the output.
4. The definition of a usable result.

Job-description drafting with a second-pass review is a practical example:

c

A person checks both outputs before the job is posted. That review step is part of the workflow, not an optional final polish.

If your approved tool supports persistent projects or controlled knowledge bases, you can store approved examples and instructions once. Apply the same access and retention rules you would use for any other HR system.

Step 5: Write a one-page AI use policy before you scale

Write an interim policy before AI use spreads beyond a few controlled tasks.

According to SHRM's 2026 research, 49% of organizations using or preparing to pilot AI had a workforce AI policy. Among organizations with a policy, only one quarter described it as clear and future-proof; 54% said it was too restrictive and tool-specific, while 23% said it was too broad.

A useful one-page policy covers:

• Approved tools and permitted use cases.
• Prohibited data and inputs.
• Required human review.
• Named ownership of the final decision.
• A route for questions, corrections, and escalation.

As the policy matures, add data classifications, vendor review, bias-testing requirements, employee and candidate notice, recordkeeping, accommodations, an appeal process, and legal review for new high-risk uses.

Review the policy every three to six months. Write it around data, risk, and decision rights so it survives a change in vendors.

Step 6: Roll it out as a small experiment

Test the workflow with two or three colleagues before expanding it.

Give a short demonstration, share the approved prompt and review checklist, and collect feedback weekly for the first month. Keep the questions simple: Was the output useful? What was wrong or missing? How much time did review take? Would you use the workflow again?

Change management has a measurable effect. SHRM reported that organizations following change-management best practices were 2.6 times more likely to report successful outcomes.

Protect skill as well as data. A small 2025 MIT preprint on AI-assisted essay writing found lower engagement and recall in some LLM-assisted conditions, but the researchers cautioned against generalizing beyond the study. The useful lesson for HR is narrower: keep people actively involved in reasoning, review, and correction rather than asking them to accept generated work passively. Read the MIT Media Lab study and its limitations.

Track results by task. If job-description drafts consistently pass review but policy summaries do not, improve or stop the weaker workflow instead of judging the entire pilot as one success or failure.

Step 7: Approach employment decisions carefully

Save hiring, promotion, performance, compensation, and discipline for last. These uses can affect a person's livelihood and carry the greatest legal and bias risk.

Use this boundary: AI may organize relevant evidence, but a named person remains responsible for the decision. High-risk uses need documented criteria, human review, accommodation and correction paths, appropriate notice, vendor due diligence, and legal approval.

The legal landscape changes quickly. As of September 2026:

• New York City's Local Law 144 requires covered automated employment decision tools to undergo a bias audit within one year of use, publish audit information, and provide required notices.
• Illinois' amended Human Rights Act has prohibited discriminatory use of AI in employment since January 1, 2026. Illinois Human Rights Commission materials summarize the amendment.
• Colorado replaced its earlier framework with an automated decision-making law that takes effect January 1, 2027. The Colorado Attorney General is developing implementing rules.
• The EU AI Act treats certain employment and worker-management systems as high-risk. The main Annex III high-risk obligations apply from December 2, 2027 under the current implementation timeline. The European Commission explains the scope and dates.
• Federal disability law still matters when employers use AI. The EEOC warns that algorithmic tools can screen out qualified people with disabilities and that employers may need to offer a reasonable accommodation or alternative method. Read the EEOC's current guidance.

Do not rely on a general article for legal conclusions. Confirm the rules for every jurisdiction and use case with current official guidance and counsel.

AI uses to avoid in HR decisions

FAQs about using AI in HR

Will AI replace HR professionals?

AI can reduce administrative and drafting work, but HR still owns context, empathy, conflict resolution, and accountable decisions. The useful question is not whether AI can produce an output. It is whether the task should be automated, augmented, or kept entirely human.

Is it safe to put employee data into ChatGPT?

Only use real HR data in a tool and account your organization has approved for that purpose. Remove direct and indirect identifiers, follow your data-classification rules, and confirm contractual controls for training, retention, access, and deletion. A personal account with training disabled is not automatically an approved HR system.

How do I write better prompts for HR tasks?

Give the tool context, a role, an action, a required format, and a tone. Add source boundaries and constraints: what information it may use, what it must not infer, and what it should flag for review. Then verify the result against the source material.

Do we have to tell candidates when we use AI?

Disclosure depends on the tool, use, and jurisdiction. New York City already requires notice for covered automated employment decision tools, and other laws have their own effective dates and requirements. Build notice into the workflow and confirm the current rule with counsel instead of relying on a generic disclosure statement.

Where can I keep learning?

The Ziplines AI for HR Professionals course covers prompting, guardrails, workflow design, and human oversight through practical exercises. SHRM's State of AI in HR 2026 report is a useful benchmark for adoption, governance, and measurement.