AI Agents for HR: Onboarding, Policies and Questions

AI agents for HR are most useful on three jobs: turning an onboarding checklist into tracked work, answering employees’ policy questions from the handbook, and drafting announcements. In each, the agent reads what it is given and drafts, and a person approves anything that changes a policy, reaches all staff, or touches one employee’s situation. The data rule matters as much as the jobs: keep medical records, investigations, performance files and pay out of the agent’s reach. In the US, medical information about employees must be kept in separate, confidential files under the ADA, and an agent that can read the whole HR folder has quietly broken that separation.

This page starts where hiring ends. Screening and interviewing are covered in AI agents for recruiting; the jobs for a whole small business are in AI agents for small businesses.

Three jobs worth handing over

1. Onboarding checklists

The checklist itself, with the federal paperwork, the access list and a filled example, is in the onboarding checklist template. The agent’s part is keeping it moving: it copies the template for each new hire, fills in the role-specific lines from the job description, drafts the welcome email and the first-week schedule, and each morning lists what is still open. A person does the parts that need judgment or authority: examining identity documents, granting access, and meeting the new hire.

2. Policy questions from the handbook

Employees ask the same questions every week: how many vacation days carry over, how to claim a work expense, what the remote-work rules are. An agent can answer them well if it is held to three rules:

• It answers only from the current handbook and approved policy documents, and quotes the section it relied on, so the employee can check.
• When the handbook does not answer the question, it says so and hands it to HR. It never fills the gap with what policies usually say.
• Anything about one person’s own situation, such as leave for a medical reason, a complaint, pay or a disability, goes straight to a person without an answer.

The questions the handbook could not answer are the useful by-product. Collect them each month; they are the list of what the handbook should say and does not.

Keep exactly one current version of the handbook where the agent reads it, and retire old copies from that folder the day a new one is approved. An agent answering from last year’s leave policy is worse than no agent, because its answer arrives with a section number and sounds authoritative. When a policy changes, a person checks a handful of the agent’s answers on that topic before it goes back to answering on its own.

3. Drafting announcements

Open enrollment reminders, office closures, a new expense tool, a policy change. The agent drafts from the approved facts, in plain words, with the date and what the employee has to do in the first two lines. A short brief keeps it honest:

Announcement: Change to the expense policy
Audience:     All US employees
Facts:        Only those in the approved policy text (attached)
Must say:     What changes, from which date, what to do now
Must not:     Interpret the policy; mention any individual case
Voice:        Plain words, no jargon; American English
Review:       HR lead approves; legal checks any policy wording

The “Must not: interpret the policy” line matters. An announcement that explains a policy in different words has quietly issued a second policy, and employees will quote whichever one suits them.

Keep medical and personnel data out

This is a plain-language summary, not legal advice. State privacy and employment laws add to it; talk to employment counsel about your own records.

The ADA regulation at 29 CFR 1630.14↗ (opens in a new tab) says information about an employee’s medical condition or history, obtained through a medical examination or inquiry, must be collected and maintained on separate forms and in separate medical files and treated as a confidential medical record. The exceptions are narrow: supervisors and managers may be told about necessary work restrictions and accommodations, first aid and safety staff may be told when emergency treatment might be needed, and government officials investigating compliance get relevant information on request. An AI tool is none of those.

Other rules follow the same pattern. The FMLA regulation at 29 CFR 825.500↗ (opens in a new tab) requires medical certifications and histories created for FMLA purposes to be kept as confidential medical records in files separate from the usual personnel files. GINA’s regulation, 29 CFR 1635.9, does the same for genetic information, including family medical history.

HIPAA is often assumed to cover all of this, and usually it does not. The HIPAA definition of protected health information in 45 CFR 160.103↗ (opens in a new tab) excludes information “in employment records held by a covered entity in its role as employer.” HIPAA’s covered entities are health plans, health care clearinghouses and certain health care providers, and a group health plan can be one of them, so records held for the plan may be covered even when your HR file is not. The exclusion does not make medical information in HR files free to share; the ADA, the FMLA and GINA still apply.

In practice, that means:

• Give the agent the handbook and the policy folder, not the HR system. A policy assistant has no reason to read anyone’s file.
• Keep medical, investigation, performance and pay records in systems the agent cannot reach, and check the connector’s permissions rather than trusting the prompt.
• Tell employees not to paste medical details into a question, and have the agent stop and hand over when they do.
• Check whether the AI tool stores or trains on what it is given, and choose a business plan or setting that says it does not.

What a human must approve

• Every policy, and every change to one, before the agent answers from it.
• Every all-staff announcement before it is sent.
• Anything about one employee: leave, accommodations, complaints, discipline, pay, performance and termination.
• Any answer the handbook does not clearly support.
• Access: which accounts a new hire gets, and removing them when someone leaves.
• What the agent can read, reviewed whenever a folder or system is connected.

Start with every item approved one at a time. Routine handbook answers that go out unchanged for weeks may be allowed to send on their own; anything about one person never is. Worked examples of where the person steps in are in human-in-the-loop AI examples.

A board for the hand-offs

The work that waits for HR is easy to lose in an inbox. On fenbs, each item that needs a person is a task: “approve the expense-policy announcement”, “five handbook questions with no answer this month”, “laptop and accounts for the new analyst”. An AI assistant files them with the details and links in the note and a priority from 1 to 10, and a person moves them through To Do, Next Up, In Progress and Completed. Add many turns a pasted checklist into tasks in one step, which is how the onboarding template becomes a week of work on the board.

Keep the board free of personal data: name the new hire by their start date and role, and link to the HR system for anything more. Standing rules, such as “the assistant never answers a question about an individual employee’s leave or health”, go on the Decisions and rules page, which every connected AI assistant reads first; facts nobody chose, such as where the current handbook lives, go in AI context notes. History records who changed each task, person or assistant. fenbs is not an HR system: it has no due dates and no assignee you can set, so deadlines go in the title or note and a person watches them.