A person should own every decision about a person

I think AI has a real place in people operations.

Recruiters shouldn’t have to start a new search from a blank page every time. Managers shouldn’t need to dig through six documents to find the onboarding steps they were supposed to complete. People teams shouldn’t spend a good part of the week rewriting the same information for different audiences.

There is a lot of work in those processes that technology can genuinely help with.

But there’s a line I don’t want us to cross simply because we can make the workflow faster.

A person should own every decision about a person.

It’s a principle I keep coming back to when thinking about AI in hiring and employee support.

Assistance is not accountability

Imagine an AI assistant prepares a candidate summary.

That can be valuable. It might organize interview notes, pull together evidence against job requirements, and make it easier for the recruiter to prepare for a conversation.

But a summary is only as good as the information it received and the assumptions behind it.

A person still needs to decide what matters, check what is missing, and understand whether the assessment is fair.

Or imagine an employee asks an HR assistant about a leave policy.

If the answer is grounded in the right, current policy for their region, great. That’s useful access to information.

If the assistant makes up an entitlement, retrieves an outdated policy, or misses a sensitive detail, that’s not just a factual mistake. Someone may make a life decision based on it.

The person accountable for the service doesn’t disappear because a bot delivered the answer.

So where do the gates go?

I don’t believe every sentence AI generates requires a committee to approve it.

A meeting summary and an employment decision aren’t the same thing. A draft agenda and a candidate decline aren’t the same thing.

The more consequential the outcome, the more deliberate the review should be.

A practical approach is to ask:

Who can be affected by this output?

What information is the system allowed to use?

Can the mistake be reversed?

Would the person receiving it understand that AI was involved?

Who can stop the workflow when something looks wrong?

And who do we contact when the model doesn’t know?

Those questions are less fun than making a slick demonstration. They’re also the questions that make the system usable beyond the demonstration.

The one message a candidate rereads

Most hiring emails are read once. A decline gets read three times, forwarded to a partner, and sometimes screenshotted. It’s the message a candidate remembers when someone asks them, a year later, what it was like to apply.

A model can write a good decline. What it can’t know is what isn’t in the system. That this person was referred by someone on the team. That they came close last year. That someone promised them feedback on the phone.

Those details live in someone’s head, not in the ATS. A person reading the draft for ten seconds catches them. An automation never will. That’s why, in the recruiting workflows I design, a decline never sends itself.

The human needs enough information to actually review

There’s a trap here too.

It’s easy to add an ‘Approve’ button and say there’s human oversight.

But if the reviewer is receiving a polished recommendation with no supporting evidence, no indication of uncertainty, and no practical way to challenge it, we haven’t designed very good oversight.

We have designed a button.

Good review means the person can see the source, understand the rationale, notice what’s missing, edit the output, and take responsibility for what happens next.

Sometimes the best choice is to keep the process fully human. That can be an informed product decision, not a failure to innovate.

The point isn’t to make HR slower

I want good tools. I want people teams to have more capacity. I want candidates to hear back, managers to find answers, and employees to get support without waiting on someone’s inbox.

But I’d rather build a slightly slower system that people can trust than a very fast one that creates work for everybody who has to clean up after it.

AI can do a lot of the preparation.

People still need to own the call.

That’s how I want to build.