AI for HR Professionals: Seven Workflows and the Employee Data That Stays Out

HR teams write a great deal: policies, handbooks, manager guides, benefits explanations, onboarding plans, review templates, survey summaries. Much of that writing is general and reusable, which makes it a good fit for AI drafting. HR teams also hold some of the most sensitive data in any company, so the dividing line between “general” and “about a person” is the line that matters most.

This guide covers seven workflows that stay on the safe side of that line, then the tasks and data that should stay out of AI tools. Laws were checked on November 9, 2026.

Seven workflows

1. First drafts of policies

A remote work policy, an expense policy or a social media policy starts from a common structure. AI can produce that structure quickly. The company’s facts and the applicable state law then shape the final version.

Draft a [POLICY NAME] for a company with [NUMBER] employees in [STATES].
Company facts: [e.g. hybrid, three office days, home office stipend of X]
Include sections for: purpose, scope, eligibility, the rules, how exceptions are approved, and who to contact.
Mark every point that may vary by state law with [CHECK STATE LAW], and do not state what the law requires.
Plain language, under 700 words.

The bracketed flags are the important part. Leave rules, final pay timing and similar items vary by state, and a chat tool may state one state’s rule as if it applied everywhere. The draft goes to employment counsel or a qualified HR adviser before it’s published.

2. Manager guides and conversation scripts

Managers often need a short guide for a specific moment: a first one-on-one, giving critical feedback, handling a request for flexible hours. These are general, so no employee data is needed.

Write a one-page guide for first-time managers on [SITUATION].
Include: what to prepare, a suggested opening line, three questions to ask, what not to say, and when to involve HR.
Keep the tone practical. Do not give legal conclusions; say "talk to HR" where one would be needed.

When the situation might involve a disability, a medical condition, a complaint or possible discipline, the “involve HR” step should come first.

3. Benefits explanations

Open enrollment produces the same questions every year. AI can turn plan summaries, which are already distributed to employees, into a plain-language FAQ.

Using only the plan summary below, write an employee FAQ with 10 questions and short answers about [PLAN TYPE].
Use plain language and the plan's exact figures. If the summary doesn't answer a common question, say "check with HR" instead of guessing.
[PASTE PLAN SUMMARY TEXT]

4. Onboarding plans

A 30-60-90 day plan for a role is a structure plus role-specific goals.

Build a 30-60-90 day onboarding plan for a [ROLE] joining a [TEAM SIZE] team.
Include for each period: learning goals, people to meet (by role), first deliverables, and a check-in question for the manager.
Format as a table.

5. Survey comment themes

Engagement surveys produce free-text comments that are slow to read in bulk. AI can group them into themes, as long as the comments are stripped of names and the groups are big enough that individuals can’t be recognized.

Here are anonymized comments from an employee survey (no names, teams or locations).
[PASTE COMMENTS]

Group them into 5 to 8 themes. For each theme give: a short label, how many comments fit it, two representative paraphrases (not direct quotes), and one question leadership could ask to learn more.
Do not guess who wrote any comment.

Small teams are a real re-identification risk. If a department has four people, its comments can point straight to individuals even without names. A sensible safeguard is to set a minimum group size and only run comments from groups above it.

6. Review language checks

Performance reviews written in a hurry drift into vague or personality-based language. A tool can flag those patterns without making any judgment about the employee.

Review this performance feedback (employee name removed). Flag:
- statements about personality rather than observable work;
- vague praise or criticism without an example;
- anything referring to age, health, family, religion, national origin or other personal characteristics.
For each flag, suggest a question the manager could answer to make it specific. Do not rate the employee or change the substance.
[PASTE FEEDBACK]

This helps reviewers write clearer feedback. It isn’t a performance assessment, and the rating stays with the manager.

7. Job descriptions and leveling

Job descriptions and career-level frameworks share structure across roles.

Draft a leveling framework for [JOB FAMILY] with [NUMBER] levels.
For each level give: scope of work, typical decisions made without sign-off, collaboration expected, and two observable examples of performance at that level.
Describe the work only. Do not use years of experience as the sole marker for any level.

The final line keeps the framework about what people do rather than how long they’ve been working. The recruiting-side detail, including ad wording and structured interviews, is covered in how recruiters use AI.

Choosing where the drafting happens

The same prompt carries different risk depending on the account it runs in. Two vendor examples, checked November 9, 2026, show what to look for:

  • OpenAI’s data controls page says personal ChatGPT Free, Plus and Pro accounts share data for model training by default, with an opt-out. Content in ChatGPT Business and Enterprise workspaces isn’t used for training by default.
  • Microsoft’s enterprise data protection page says Copilot Chat used with an organizational account applies the organization’s retention policies and sensitivity labels, and that prompts and responses aren’t used to train foundation models.

For HR, the practical rule is simple. Work drafting happens only in the company-managed account, never in someone’s personal chatbot login, even for “general” documents. Personal accounts drift into personal data quickly. A policy draft becomes a question about one employee’s leave, then a pasted email from that employee.

Tasks that stay with people

TaskWhy it stays out of general AI tools
Workplace investigationsWitness statements and allegations are highly sensitive, and conclusions need a documented human process
Accommodation requestsMedical information has confidentiality rules of its own
Termination and discipline decisionsConsistency and documentation matter, and some state laws now regulate AI’s role in these decisions
Pay decisions for named employeesCompensation data by person is confidential and fuels pay-equity risk if mishandled
Form I-9 and immigration documentsIdentity documents and status information

Medical information is the clearest example. The EEOC’s enforcement guidance on disability-related inquiries says employers must treat medical information from employees as a confidential medical record, shared only in limited circumstances with supervisors, first aid and safety staff, and government officials investigating ADA compliance. An AI vendor isn’t on that list.

When HR tools become regulated decision tools

Drafting help is one thing. Software that scores, ranks or recommends employees for promotion, discipline or discharge is another.

  • New York City. The rules on automated employment decision tools cover promotion as well as hiring. The city’s DCWP FAQ says employers need a bias audit within the past year, a published summary of results, and notice to NYC-resident candidates and employees 10 business days before use. For promotion candidates, notice can go in a written policy or procedure.
  • Illinois. Amendments to the Illinois Human Rights Act took effect January 1, 2026. The Illinois Human Rights Commission says they bar AI use with a discriminatory effect and require notice when AI is used to influence or facilitate covered employment decisions. The detailed notice rules were still proposed as of April 2026.
  • California. The Civil Rights Council’s automated-decision system regulations took effect October 1, 2025.
  • Federal law. Title VII, the ADA and the ADEA apply to employment decisions however they are made. The EEOC took its AI-specific technical assistance offline in January 2025, but the statutes are unchanged.

HR teams that buy performance, analytics or talent software should ask whether any feature produces scores or recommendations about individual employees. Those features need the same review as a hiring tool.

Don’t paste this in

Medical or disability information, accommodation requests, investigation notes and witness statements, SSNs, I-9 documents, bank details, or pay and performance data for named employees. The EEOC’s ADA guidance requires medical information to be kept as a confidential medical record. Tools that score or rank employees for hiring or promotion in New York City need a bias audit and notice under Local Law 144.

A one-page HR AI policy

  1. Approved tools. List each tool and its account type.
  2. Allowed inputs. General drafting, public documents, anonymized and aggregated data above a minimum group size.
  3. Banned inputs. Everything in the box above.
  4. Decision tools. Any feature that scores or ranks people goes to HR leadership and counsel before use.
  5. Review. Policies and employee-facing documents are checked by a qualified person before publication.
  6. Records. Keep a list of where AI is used in employment decisions. Illinois’s proposed rules would require AI-use records, and the list is useful for audits and notices anyway.

Teams evaluating recruiting software specifically can compare vendors’ bias-audit statements in the best AI for recruiters.



Previous Post
How Recruiters Use AI: Sourcing, Screening and the Rules That Follow