Seven HR Decisions I Wouldn’t Hand to AI

AI is the most confident colleague you will ever have. It has never once said, “Let me check on that and get back to you”. Ask it anything and you get a polished answer in about four seconds.

That confidence is part of the problem. It’s especially risky when AI moves from helping with HR work to influencing HR decisions. AI in HR decision-making can be useful, but there’s a big difference between asking a tool to help with the work and asking it to make the call.

The Oregon Bureau of Labor and Industries recently gave Google’s AI a pop quiz using questions employers ask them all the time. About half the answers came back wrong or incomplete. That’s a coin flip. You wouldn’t let a coin flip decide the proper way to handle someone’s final paycheck.

I’m not anti-AI. I use it, and so does our team for various tasks. It’s great for first drafts and for getting you past a blank page, but it gets risky when an employer stops using it to help write things and starts using it to decide things.

Here are seven places employers are tempted to hand AI the wheel, and why I’d take it back.

“Screen these 200 resumes and give me the top 10.”

This one feels like a gift. Two hundred applicants, one hiring manager, and zero free time.

The catch is that AI learns from patterns, and some of those patterns are ones you don’t want. Amazon famously scrapped an experimental recruiting tool after it taught itself to downgrade resumes that included the word “women’s.” Your tool may not be that obvious. For example, it may place too much weight on where someone went to school, treat any employment gap negatively, or consider factors such as location that may unintentionally disadvantage certain candidates and introduce bias into the process.

Discrimination laws still apply when software makes the first cut. “The tool did it” is not a defense anyone wants to test. If you use AI to sort, a human still needs to look at who got left out and why. A real person still needs to make the decision and be accountable for the outcome.

“Who should we cut in the layoff?”

You upload a spreadsheet with names, salaries, tenure, and last year’s productivity numbers, and ask AI who should go. It will happily give you a list. It may even explain its reasoning.

What it doesn’t know: the person with the productivity dip was out on protected leave for three months. Or the “highest cost, lowest output” group skews older, because your highest-paid people are often the ones who’ve been there longest.

You need to be able to explain every name on that list. AI won’t be sitting next to you when someone asks, and if you don’t have a clear, objective reason for those decisions, the reason can appear discriminatory.

“Write a termination letter for an employee who keeps missing work.”

AI will write you a clean, professional termination letter. It might also cheerfully tell the employee their final paycheck will arrive on their next regular payday.

In Oregon, when you let someone go, their final pay is due by the end of the next business day. Miss it, and penalty wages can stack up to 30 days of that person’s pay. Have employees in other states? It gets even more complicated.

Then there’s the reason itself. “Keeps missing work” needs a closer look. If some of those absences were protected sick time, they can’t be held against the employee. AI has no idea which absences were which. Your records do.

AI might suggest asking the candidate what they make now and offering a bump. That’s normal advice in a lot of places.

However, in Oregon and some other states, you can’t ask about salary history before you make an offer, and you can’t screen applicants based on it. Oregon’s pay equity law also requires that pay differences between people doing comparable work rest on specific factors such as seniority, merit, education, training, experience, or location. “They negotiated harder” isn’t on the list of reasons you can pay someone more.

AI could help pull market ranges, but where this person lands and how that compares to the people already doing the job requires a human who knows your team.

“Summarize these witness statements and tell me what happened.”

Investigations are slow and uncomfortable. I get the appeal of pasting everything into a chatbot and asking for the verdict.

A few problems. First, you may have just put confidential employee information into a tool that wasn’t approved or designed to handle it. It doesn’t notice who hesitated, who changed their story, or who got oddly specific about a detail nobody asked about. And AI summaries like to smooth things out. Sometimes they “summarize” a quote nobody actually said.

Investigation findings need to come from a person who can stand behind every line and use their people knowledge and expertise to put the pieces together.

“Can I deny this accommodation request?”

You type in the request, ask whether it’s reasonable, and get back a tidy paragraph about how it’s an undue hardship. Some tools will even draft the denial for you.

Accommodation requests require a discussion with the employee. You talk with the employee about the limitation, what might help, and what’s workable for the business. That back-and-forth is usually called the interactive process, and a chatbot can’t have it for you.

“Write performance reviews for my team of 12.”

This is the one I understand most. Twelve reviews are due Friday, and AI writes in such a lovely, encouraging tone.

That’s the issue. AI tends to be relentlessly positive and a little generic. So now Sam’s review says he “consistently exceeds expectations.” Three months later, you want to let Sam go for performance reasons, but your own paperwork disagrees.

The flip side is just as risky. If AI pulls in attendance and notes Sam “missed nine days this year,” and some of those were protected sick time, Oregon law says you can’t count them against him. That includes his review.

Use AI to clean up your wording. The examples and ratings have to come from you.

So what is AI good for?

Plenty. Drafting a first pass of a policy you’ll review. Turning messy notes into a clear email. Brainstorming interview questions. Explaining a concept in plain language so you know what to ask.

My rule of thumb: AI can help you write. It shouldn’t decide anything. If the outcome affects someone’s job, pay, or leave, a person with real context should make the call.

Where to go from here

None of this is legal advice. Think of it as an HR person waving a friendly caution flag.

If you’re a Xenium client, your HR Business Partner is the person to run these by before you hit send. That’s what we’re here for. If you’re not a client yet, we’d be glad to talk.

And if you want to hear how leaders are actually using AI, and where they’re drawing the line, join us at the 2026 Leadership & Economic Summit on November 12 at the Sentinel Hotel in Portland. We’re co-hosting with Columbia Bank, and the lineup includes an executive panel on AI. Register here.