‍Who’s responsible when an AI tool screens out the wrong candidate? Most HR leaders I talk to have a quick answer. The vendor. They built it, they tested it, they gave us a bias audit and a slide that said “responsible AI” in a nice font. We’re covered.

On the latest episode of the Unmuted Podcast, I sat down with Haley Friary-Schenk, Founder and CEO of HumaniCore AI, to dig into who’s actually responsible when HR uses AI. She dismantled that answer in one sentence: “Our vendor said we were covered” is not a defence.

That’s the real story behind AI in HR right now. The debate everyone’s having is whether AI is accurate enough or fair enough. The question that actually matters is who can prove it, and who’s on the hook when they can’t.

‍You Can Buy the Tool. You Can’t Buy Out the Accountability.

Here’s the uncomfortable truth. When AI influences who gets hired, evaluated, promoted, paid, or managed out, the decision is still yours. You chose the tool. You configured it. You decided how much weight its output carries. The vendor controls the model. You control the process it sits inside.

‍Both sides have to independently demonstrate accountability for the parts they control. Neither gets to point at the other.

And they aren’t the only ones in the room. Candidates and employees need to trust the decisions being made about them. Regulators need to trust the evidence. Trust in AI isn’t a single handshake between a buyer and a seller. It’s a chain, and HR sits right in the middle of it.

We’ve already seen how this plays out in Canada. In Moffatt v. Air Canada, a customer relied on bad bereavement fare advice from the airline’s chatbot. Air Canada’s defence was, essentially, that the chatbot was responsible for its own actions. B.C.’s Civil Resolution Tribunal called that “a remarkable submission” and held that “it should be obvious to Air Canada that it is responsible for all the information on its website”.

That wasn’t a hiring case. It doesn’t need to be. Swap “chatbot” for “screening algorithm” and “website” for “hiring process”, and the logic holds. If the tool sits inside your process, its output is yours. Pointing at the technology didn’t work for Air Canada, and pointing at the vendor won’t work for you.

‍The Rules Are Arriving. Unevenly.

Regulation isn’t a someday problem anymore, but it’s also not a tidy one.

In Ontario, employers with 25 or more employees now have to disclose in publicly advertised job postings whether AI is used to screen, assess, or select applicants. That took effect January 1, 2026. New York City has required bias audits for automated hiring tools since 2023. Colorado’s AI Act has been delayed, stayed, and rewritten, with employer obligations under the replacement law now pointed at January 2027. Europe is running on a different playbook again.

If you hire across provinces in Canada, states in the USA, or other countries, there isn’t one standard to meet. There’s a patchwork. And “we’ll deal with it when it’s enforced” is a strategy that ages badly.

‍Your Tech Stack Is Where the Risk Hides

This was the part of our conversation I keep coming back to.

Most growing organizations don’t have one AI tool. They have an ATS with AI ranking, a sourcing tool with its own matching logic, an interview intelligence platform, maybe a performance tool that summarizes feedback. Each vendor has its own audit, its own definition of “fair”, and its own fine print.

Nobody is looking at how they behave together. A candidate doesn’t experience your vendors one at a time. They experience the whole funnel. If three tools each nudge the same group slightly downward, every individual audit can look clean while the outcome doesn’t.

That’s not a vendor problem. That’s an HR problem.

‍Five Questions to Ask Every AI Vendor

Haley laid out five questions every HR leader should be asking before signing, and at every renewal after:

• Who tested the tool, and did they also sell the remediation? An auditor who profits from finding problems they then fix isn’t exactly independent.
• How many people were excluded from the audit? A clean result on a trimmed dataset tells you less than it looks like it does.
• Was testing done on real applicants or a designed dataset? Lab conditions aren’t your applicant pool.
• When is the model retrained, and will you be notified in advance? An audit from 18 months ago describes a model that may no longer exist.
• Where does the vendor’s responsibility end and the employer’s begin? If the contract can’t answer this clearly, you’ve found your gap.

If a vendor gets defensive about any of these, that’s an answer too.

‍HR Tech’s SOC 2 Moment

There’s a reason no serious company buys software anymore without asking for a SOC 2 report. Security went through a period of “trust us”, then a period of breaches, then a standard for independent verification that buyers came to expect by default.

I believe AI in HR is heading for the same moment. The bar Haley described is higher than most vendor audits clear today: verification that is independent, ongoing, and reproducible, with meaningful human oversight built in. Not a one-time event. Models get retrained. Applicant pools shift. Configurations change. A certificate from launch day is a snapshot of a moving target.

The irony is that self-assessment isn’t really trust. It’s marketing. Trust is what happens when someone with nothing to sell checks the work, and keeps checking it.

Even regulators struggle here. When New York State’s Comptroller audited enforcement of NYC’s AI hiring law, the city’s review of 32 companies flagged one issue of non-compliance. The auditors reviewed the same 32 companies and found at least 17 instances of potential non-compliance. If the people enforcing the rules miss that much, “the vendor handled it” deserves a second look.

‍A Final Thought

I’ve spent enough years leading HR at growing companies to know how these decisions get made. A vendor demo goes well. Legal reviews the contract. Someone asks about bias, gets a reassuring answer, and the conversation moves on to implementation timelines.

That’s the moment Haley’s message is aimed at. The defensible evidence has to exist before the complaint, not get assembled after it. Right there, in the buying conversation, when the questions are still free.

AI will keep making HR faster. That part is settled. What isn’t settled is whether we can stand behind what it decides on our behalf.

Your vendor can tell you you’re covered. Only you can prove it.

‍The full conversation with Haley Friary-Schenk is available now on the Unmuted Podcast (or watch below), part of the WRKdefined Podcast Network, and on our show page at goscalehr.com. If you’re rethinking how your team evaluates AI in your people processes, it’s a conversation worth having, and one we have with clients at ScaleHR every week.

‍