What Can an AI Agent Do in HR for a Small California Employer, and What Should Stay With a Person?

Let an agent chase paperwork, send reminders and draft schedules for a person to approve. Keep the Form I-9 document check, any shift cut that touches pay, and every decision about who is hired, disciplined or let go with a person. In California you answer for what the agent does, so start with one workflow, read-only, and keep a log.

I am Mac Malbrough. I hold a J.D. and an MS-HRM, I am not an attorney, and this guide is for the owner or office manager of a company with 10 to 150 employees who keeps hearing that “AI agents” will run HR. Most of the coverage is written for a chief human resources officer at a company with thousands of people. This is the version for a company where one person does payroll, onboarding and scheduling before lunch.

What does agentic AI mean in HR?

An agentic AI system takes several steps toward a goal on its own, using the tools you connect to it. A chatbot drafts an offer letter when you ask. An agent notices that a new hire has not signed the handbook acknowledgment, sends a reminder, checks the signature platform two days later, and tells you which of the five new hires still owe paperwork.

The difference is the follow-through. The agent can read your HR system, write to it, send messages and trigger the next step without a person pressing a button each time. That is also where the risk sits. A bad draft can be deleted. A bad action may already have changed a schedule, a pay record or an email thread by the time anyone looks.

How much of the agentic AI pitch is real?

Less than the vendor demos suggest. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. It also said that of the thousands of vendors using the word “agentic,” only about 130 offer genuine agentic products. The rest are relabeling ordinary automation, which Gartner calls agent washing. In the summary I read, National CIO Review added that Gartner expects at least 15% of day-to-day work decisions to be made autonomously by agents by 2028, and that a January 2025 Gartner survey found only 19% of organizations had made significant investments in the technology.

Read those numbers two ways. The technology is moving into HR software whether you shop for it or not, so you will be offered it. And most early projects disappoint, so a small employer has no reason to be first. Before you buy, ask the vendor three plain questions:

• What can the agent change on its own, and what does it only suggest?
• Where is every action logged, and can I export the log?
• Can I turn each action off separately, and what happens to work already in progress?

If the answers are vague, you are probably looking at automation with a new label. That is fine for a reminder tool. It is not a reason to hand over decisions.

Which onboarding tasks can an agent handle?

Agents suit onboarding because it is a checklist with deadlines, and most of the work is chasing people. Reasonable jobs to hand over include:

• Sending the offer packet and reminding the new hire about the federal W-4, the California DE 4 and direct deposit forms.
• Collecting the signed handbook acknowledgment and the signed wage theft prevention notice, which California requires you to give non-exempt employees at hire.
• Scheduling day-one equipment, system access and the first-week meetings with the manager.
• Telling you each morning which new hires are missing which documents.

The agent should not complete the Form I-9 document check. The employee completes Section 1 by the first day of work. Section 2 is due within three business days of the start date, and the employer’s representative must examine the original documents in person. DHS allows an alternative live video procedure only for employers enrolled in E-Verify in good standing. The agent can book the appointment and chase the employee to bring documents. A person looks at the documents and signs. The USCIS I-9 Central page has the current rules.

Why should an agent not trim shifts to avoid overtime?

Some vendors pitch an agent that adjusts shift distribution to prevent overtime. In California that goal collides with several pay rules, and an agent optimizing for “no overtime” does not know about them unless someone wrote them in.

• Reporting time pay: an employee who reports to work and is sent home early, or given less than half the usual or scheduled day, is generally owed half the usual or scheduled day at the regular rate, with a floor of two hours and a ceiling of four. A last-minute shift cut can create that bill. Zaller’s summary walks through the cases, and the DIR page is here.
• Split shifts: a schedule broken up by unpaid gaps you set can trigger an extra hour of pay at minimum wage, as the DIR split shift page explains.
• Daily overtime: California counts hours over 8 in a day, not only over 40 in a week, under Labor Code 510. An agent that only watches weekly totals will miss it.
• Meal and rest periods: a compressed schedule that pushes people past the fifth hour without a break creates premium pay under the rules in my meal and rest break guide.

California has no statewide predictive scheduling law, according to the Zaller article, but some cities do, and I walk through the Bay Area ones in my local ordinance guide. The safe design is an agent that proposes a schedule, flags the pay consequences of each change, and waits for a manager to approve. It should never publish or edit a worked schedule on its own.

Does California treat an agent’s decision as your decision?

I read it that way, and you should plan on it. The Civil Rights Council’s regulations on automated-decision systems took effect October 1, 2025 and reach employers with five or more employees. They apply to tools that make or help make employment decisions, and they require you to keep the relevant data and records for four years. I explain who is covered in my guide to AI hiring tools and the Civil Rights Council rules.

An agent that ranks applicants, recommends discipline or decides who gets the good shifts is making or shaping an employment decision. If its pattern of choices hurts a protected group, the vendor’s name on the software will not move the problem off your desk. The practical rule is simple: the more an action affects someone’s job, pay or schedule, the more a named person should review it before it takes effect.

How do you pilot an agent without betting the company?

Pick one workflow with a clear finish line and low stakes, and run it in stages. New-hire paperwork chasing is a good first choice. Overtime scheduling is a bad one.

1. Read-only for two weeks. The agent reports what it would do. You compare that against what your office manager actually did.
2. Act with approval. The agent drafts messages and changes, and a person clicks approve on each one.
3. Act alone on the narrow, reversible tasks. Reminders qualify. Pay, schedule and discipline changes do not.
4. Review the log monthly. Look for actions nobody expected, and note who reviewed them.

Limit what the agent can see, too. Give it the onboarding file, not the whole HR system with medical leave notes and complaint files. Record in your handbook that employees’ work may be handled by AI tools and who to contact with questions. My piece on AI notetakers in HR meetings covers one tool you should keep out of the most sensitive conversations.

What should you do this week?

List every HR task one person in your company repeats each week, and mark each as “chasing,” “calculating” or “deciding.” Agents may help with the first, a person should check the second, and the third stays with a person. Then pick the one chasing task that costs the most time and write the three vendor questions above next to it.

If new hires are the bottleneck, my 30-60-90 onboarding kit gives you the checklist an agent would follow, so you can run it by hand first and automate it second. If you want a second set of eyes on the whole HR system, the HR compliance check-up is where I start.

For the longer view on how the HR role itself changes, read what happens to HR’s job when AI coordinates more of the work.

Start smaller with one drafting task and a one-page AI use policy.

To find where your team stands, see the four stages of AI use on an HR team.

Related reading: ten AI pilots for a small HR team, ranked by risk.

Sources

• Gartner, press release of June 25, 2025, and National CIO Review summary, read October 31, 2026
• Zaller Law Group, Scheduling Restrictions on California Employers, August 18, 2023, read October 31, 2026
• DIR, reporting time pay and split shift pay
• Labor Code section 510
• USCIS, I-9 Central
• California Civil Rights Council, automated-decision system regulations, June 30, 2025

McKinley holds a J.D. but is not a licensed attorney. Articles here are general information, not legal advice. For your specific situation, talk with an employment attorney.

Data current as of October 2026. Sources are linked where each figure appears.