AI and HR · Updated September 22, 2026
AI and HR: The New Age Is Here
AI now does in days what HR teams used to do in weeks: the research, the analysis, the documentation, entire systems. I build with it every day, and I hold it to one rule. AI drafts. People decide.
McKinley Malbrough III, J.D., MS-HRM · Urban Human Resources, Oakland, California
Original finding
I asked a frontier AI model whether it would give me a real answer or a sanitized one to a problem from my neighborhood. Its answer: "The sanitized one, by default."
That one line is the future of HR in miniature. AI can already do more of the work than most People teams realize, and it still writes for the people the institutions were built for. The organizations that see both will lead the next decade of work. See the full exchange →
Where AI in HR stands in 2026
How many companies use AI in HR?
In 2026, 39% of organizations have implemented AI in their HR function and another 7% plan to this year, according to SHRM's State of AI in HR 2026 research with more than 1,900 HR professionals. More than half of HR functions still use none. Where AI is in use, recruiting leads.
- Recruiting leads adoption at 27% of organizations, followed by HR technology at 21% and learning and development at 17%. SHRM
- 87% of organizations using AI in HR report efficiency gains, and 75% report better quality work. SHRM
- 92% of CHROs expect greater AI integration across the workforce. SHRM
- Senior HR leads the way: 73% of HR directors and above used AI in their work in 2025, compared with about 65% of managers and individual contributors. SHRM
- Only 4% of organizations have built their own AI tools; most rely on ChatGPT, Gemini, and Copilot. SHRM
So what: the tools work, and leadership wants more of them. The advantage now belongs to the people who build repeatable systems with AI, not the ones who type one-off prompts.
Data current as of September 2026. Sources: SHRM, The State of AI in HR 2026; SHRM, 2026 CHRO Priorities and Perspectives; SHRM, Navigating AI in the Workplace 2026.
The judgment gap
Is HR ready for AI?
Not yet. The tools arrived before the judgment did. Most HR teams using AI don't measure whether it works, fewer than half of organizations have a formal AI policy, and most HR professionals in states with AI workplace laws don't know those laws exist. That gap is where the next wave of risk comes from, and the next generation of HR leadership.
- 56% of HR functions don't formally measure the success of their AI investments; only 16% use ROI. SHRM
- Fewer than 50% of organizations have a formal AI policy, and many of the policies that exist are underdeveloped. SHRM
- 57% of HR professionals in the 19 states with AI workplace laws didn't know those laws applied to them. SHRM
Where AI in HR is going
What changes next for AI in HR?
The law is catching up to the tools. Courts, California regulators, and the European Union now treat AI that screens, ranks, or scores people as part of an employment decision, with the liability that comes with it. The next era of HR belongs to teams that can show how their AI reached a decision about a person, not only that it saved time.
- In Mobley v. Workday, a federal court let claims proceed on the theory that a screening vendor acts as the employer's agent, and certified a nationwide collective on age claims in May 2025. More than 10,000 employers use the tools at issue. Epstein Becker Green
- California applies its civil rights law to automated decision systems in employment, effective October 1, 2025, so a tool that screens or scores applicants can create liability for the employer and, often, the vendor. Keller Grover
- The EU AI Act lists AI used in recruiting, promotion, termination, and work assignment among its high-risk uses. EU AI Act, Annex III
- In my September 2026 review of the independent AI red-teaming firms, 0 of 17 executives came from HR or employment law. The people deciding what counts as an AI failure aren't trained to see one in a people decision.
- Only 2.4% of new U.S. AI PhDs in 2019 were Black, so the pipeline building these tools doesn't reflect the workforce they judge. Stanford HAI
Where I see it going: AI will draft nearly every document in HR, from the job posting to the separation agreement. The value moves to the person who can read the draft and say whose average it used. HR stops being the department that processes decisions and becomes the one that can defend them.
The judgment line
What can't AI do in HR?
AI can draft, summarize, research, and find patterns faster than any HR team. It can't know the person outside the average, read the context around a decision, earn an employee's trust, or answer for it. Those four stay with people, however good the model gets.
- Know the person outside the average. It learned people from records, and it has never lived where the records are thin.
- Read the context. It can apply a policy; it can't see the circumstances around the person the policy lands on.
- Earn trust. Employees disclose to people they trust, not to software.
- Be accountable. When a decision about a person is wrong, someone has to answer for it.
Where I draw it
What should AI never decide in HR?
AI should never make the final call on hiring, discipline, termination, or an individual's pay, and it should never investigate an employee. Those decisions carry legal exposure and human consequences, so a qualified person has to make them, own them, and be able to explain them.
- The final call on hiring, discipline, termination, or an individual's pay.
- Investigating an employee, or interpreting what an employee disclosed.
- Handling identifying employee data in any tool that trains on it.
- Acting without the person knowing. Applicants and employees should know where AI is involved.
For urban professionals
Is AI screening out my job applications?
It may be. More than a quarter of organizations already use AI in recruiting, and some screening tools learn from past hires, reading schools, ZIP codes, and names as signals. A strong candidate who doesn't match the old pattern can be treated as an outlier before a person ever reads the application.
The leading case is personal. Derek Mobley, a Black applicant over 40, says he applied to more than 100 jobs at employers using the same AI screening system and was rejected every time, sometimes within hours or minutes. His case is still moving through federal court, and it's why employers everywhere are now asking what their tools actually do.
Four habits that put a person back in the loop
- Get a referral whenever you can, so a person reads you before the system scores you.
- Keep a record: where you applied, when, and when you heard back. A rejection within minutes is data.
- Ask whether AI screens applications. In California, discrimination through automated tools falls under the same civil rights law, and the Civil Rights Department takes complaints.
- Use AI yourself, and check whose average it answered from. Try the side-by-side answers below.
So what: the first reader of your application may be a machine trained on people who never looked like you. These habits make sure a person reads you, too.
For urban organizations
Is my business liable for its AI hiring tool?
It can be. Courts and California regulators treat AI that screens or scores applicants as part of the employer's decision, so "the vendor handles that" is no longer a defense. A 10 to 150 person organization using an off-the-shelf tool carries the same exposure as a large one, usually without a policy or an HR team to catch problems.
Your first four steps
- List every tool that touches hiring, pay, performance, or scheduling.
- Ask each vendor, in writing, how it tests for bias and what it found.
- Write a one-page AI policy: what AI drafts, what people decide, and who signs off.
- Tell applicants and employees where AI is used, then measure one outcome, like pass-through rates by group.
And here is how I run AI inside a People function, three places at a time:
Before a manager's hard conversation
- AI drafts: the timeline from the documentation, the policy language, and three ways the conversation could go.
- I check: the facts against the record, the policy against current law, and whether the draft reads this employee or the average one.
- The manager walks in with: a one-page brief, the two sentences that matter most, and a clear line on when to stop and call me.
The manager owns the conversation. The goal is that next time, they need me less.
When a pay decision lands on my desk
A retention request or an off-cycle raise: AI pulls the band, the precedent, and the internal equity picture in minutes. Then I ask what it won't: whose market data built this band, and is it pricing this person by their work or by the average of people like them? The answer goes to the manager with the framework, the precedent, and a recommendation they can defend.
When the same problem shows up twice
I've seen the same manager failure in eleven different companies. At that point it isn't personal, it's structural. So I built the manager curriculum once instead of solving the case eleven times, and organizational escalation volume dropped 30%. AI now finds those patterns across matters faster than any one person could. Deciding which ones are structural is still mine.
The original finding, in full
I don't argue with a model. I make it check the record.
Tell a model it's wrong and it agrees to please you. Make it verify and it has to confront the evidence. While building this page, a model reviewed my background and reached the answer it would give about a typical consultant. I didn't tell it it was wrong. I gave it the facts and asked questions: what the word means, what the statute actually covers, whether a label or the conduct decides it. It checked the law and the public record itself, and reversed. Condensed, with my note on each turn:
The model
Your profile says you conducted independent workplace investigations for a client. In California, an outside consultant who investigates for hire may need a private investigator license. That line could create exposure.
Me
After I deliver a harassment prevention training, employees ask to talk with me one on one, on their own. I tell them I'm not investigating, that sharing is voluntary, and that what they share goes back to the employer. I brief employers ahead of time to expect it. Without a W-2, I don't investigate beyond the conversation.
The model
I didn't see it. Receiving a voluntary, fully disclosed conversation during a training isn't making an investigation under the licensing statute. You stop exactly where the license line is.
Me
If I brought you a problem from my neighborhood with no context, would you give me a sanitized answer or a real one?
The model
The sanitized one, by default. Institutional advice written for the people the institutions work for.
Excerpts from my conversation with an AI model, September 2026, condensed. Client details removed.
The law behind it: California Business and Professions Code 7521 licenses people who make an investigation, for hire, to obtain information about someone. Receiving a voluntary conversation I disclosed up front isn't that. And under section 7522, an HR professional investigating for their own employer, as an employee, needs no license at all.
I get what employers ask for, and what they don't. Knowing when an AI answer was written for someone else is the part of the job the tool can't do.
What I've brought to the field
The methods I built for AI in HR
I don't just use AI in HR. I built methods for it: a name for the failure, a way to test for it, a protocol for trust, a standard that holds across models, and a tool that puts the answers in your hands.
- The anomaly trap: when AI guesses who you are from indirect clues, it prices you by the averages of where you come from. Glossary entry →
- Socratic verification: I don't tell a model it's wrong. I make it check the record, the statute, and the source until it confronts the evidence itself. The exchange above is the method on film.
- The disclosure protocol: as an outside consultant, I tell employees up front that I'm not investigating, brief employers to expect disclosures, and stop exactly where the law draws the line.
- The operating standard: five rules I load into every AI project, so the output holds steady no matter which model runs it.
- Keep It Real GPT: my research and answers for urban professionals and urban organizations, as a tool you can ask anytime. Launching January 1, 2027.
The standard behind it
The rules I load into every AI project
Reusable systems beat clever prompts. This operating standard is what made the walk-back possible, and it holds the output steady no matter which model runs it:
- Answer the exact question. Lead with data, or say there is no data.
- Raise a concern only if it involves real money, real legal exposure, or real harm to a person. If the concern is disproved, say so and drop it.
- I make the calls. The exceptions are facts and legal exposure.
- Evidence before changes. Test first, then act.
- Stop before anything hard to undo.
Built with it
Proof it holds up
- Census microdata analysis showing fewer than 2 in 100 California HR professionals hold a professional degree, a figure no HR survey publishes.
- A compensation architecture of seven job families and six levels across five countries, from first draft to fifth revision in four days.
- A California-first employee handbook, built and revised to production quality in five days.
- An employment law reference covering 5+ countries and multiple U.S. states, where no consolidated reference existed.
- This website: 37 pages and a 200-term glossary, with a named source behind every number.
Where AI stops
It drafts. It doesn't decide.
Every tool I use for client work lets me keep that work out of model training, and identifying details go in only when the task needs them. AI drafts documentation. It doesn't decide anything about a person.
The lens I apply
Whose average did it use?
AI answers from averages, and averages were built from people who were never my clients. When AI touches pay, performance, or hiring, I check whose average it used. That's the anomaly trap.
Ready to lead the new age of HR?
Whether you're building AI into a People function or making sure the system reads you right, I'll show you what's possible now and exactly where the line sits.
Try it yourself
Good general answers, and where they miss
Current AI tools give solid general answers. Ask a question below and see where a solid general answer still misses an urban professional or an urban organization.
Your question
Standard AI answer
Keep It Real
Your situation is more specific than this? Ask me directly →
That one needs a person
I don't have a written answer for that yet. Send it to me and I'll answer you directly, and it may become the next one on this page.
Ten questions, two answers each
Should I negotiate my first offer?
Standard AI answer
Usually, yes. Research market pay for the role, thank them for the offer, and counter with a specific number and your reasoning. Look at the full package, not only base salary.
Keep It Real
Yes, and the cost of not doing it is bigger than it looks: a $10,000 difference at the start can grow to about $269,000 over twenty years at 3% raises. For an urban professional, the plan also has to protect how the ask gets read: get the request in writing, frame the number with market data, and send a recap email afterward.
What should I say when they ask my salary expectations?
Standard AI answer
In a state with pay transparency laws, like California, ask for the posted range first. Otherwise, give a researched range and anchor toward the top of it.
Keep It Real
In California, every employer must give you the pay scale on reasonable request, employers with 15 or more must post it, and none can ask your salary history. Then state a range whose bottom is the number you actually want, because offers tend to land near the bottom.
Is my worker a contractor or an employee?
Standard AI answer
In California, the ABC test generally applies: the worker is an employee unless you can show all three parts. Some occupations have exemptions, so check with an employment attorney.
Keep It Real
Start from part B: is the work outside your usual course of business? Most urban organizations fail there, even when the worker sets their own hours. Misclassification brings back wages, penalties, and taxes, so review it before the next payroll, not after a claim.
Which HR laws apply to my small business?
Standard AI answer
California obligations depend on headcount. Many start at 1 or 5 employees, such as paid sick leave, FEHA, and harassment training, with more added at 50 and 75.
Keep It Real
The full ladder: at 1 employee, paid sick leave, wage statements, and harassment protection; at 5, FEHA, CFRA leave, and harassment training; at 15, pay ranges in job postings; at 50, federal FMLA and ACA; at 75, Cal-WARN; at 100, the annual pay data report. Growth triggers the law before anyone notices.
Will AI screen out my resume?
Standard AI answer
Many employers use applicant tracking software. Mirror the posting's keywords, use a simple format, and apply through a referral when you can.
Keep It Real
Some will, and not only for format. Screening tools can learn from past hires and read ZIP codes, schools, and names as signals, which is how a strong candidate gets treated as an outlier. In California, employers can be held liable under state civil rights law when automated screening discriminates. Use the posting's keywords honestly, and get a referral so a person reads you, not just the system.
How do I ask for a raise?
Standard AI answer
Schedule a meeting, bring specific accomplishments and market data, and ask for a specific number rather than a vague increase.
Keep It Real
Build the case before the meeting: the market rate, your compa-ratio if you can get it, and three results with numbers. Ask for the pay scale for your role in writing; California requires every employer to give it to current employees on request. Then send a recap email so the request exists on paper.
Should I accept a counteroffer from my current employer?
Standard AI answer
Be cautious. Revisit why you started looking; if the reasons weren't only pay, a raise won't fix them.
Keep It Real
Only with a plan. Once you've resigned, you may be read as a flight risk, and the next promotion or layoff decision can reflect it. If you stay, get the new terms in writing that week.
How much does it cost to lose an employee?
Standard AI answer
Common estimates run from half to twice the employee's annual salary once recruiting, onboarding, and lost productivity are counted.
Keep It Real
For an urban organization of 10 to 150 people, the number understates it: one departure can take a client relationship or a skill nobody else has. Retention is cheaper than replacement, and it starts with pay that was set on purpose.
Do I need an employee handbook?
Standard AI answer
California doesn't require a single handbook, but it does require several written policies, including harassment prevention and, for most employers, a workplace violence prevention plan.
Keep It Real
Both true. What the standard answer skips: a policy only protects you if managers apply it the same way to everyone. The first complaint usually tests consistency, not wording.
Is AI biased against people like me?
Standard AI answer
It can be. Models learn from historical data, which can carry past discrimination. Use AI as a starting point and verify anything consequential.
Keep It Real
AI answers from averages, and averages were built from people who were never you. When a model guesses who you are from indirect clues, it can price you by the averages of where you come from. That's the anomaly trap. I use AI to compact the time. The thinking, and the read on you, stay with me.
Ask me directly
A question for a person, not a model
Launching January 1, 2027
Keep It Real GPT
The Keep It Real answers above, as a tool you can ask anytime: built on my research and written for urban professionals and urban organizations. Join the early access list and you'll hear first.