Here's a question. An employee is leaving a job in California. Their manager wants to know if the employee gets their accrued PTO paid out on their last day.

The answer?

Yes. California treats accrued vacation as earned wages, and earned wages get paid at separation.

As most of you know, there's more to this answer. We'll come back to it, I promise.

Ten years ago that "answer" was worth something. It separated a seasoned HR professional from someone early in their career who would have to "get back to you" with an answer. People built entire careers on being the one who knew the "answer." Myself included.

But now, with AI, anyone can get that answer in seconds.

So when AI can provide the knowledge that was once hard-earned, what's left? Is there something underneath what we actually do that goes beyond providing an answer? And what does it mean to be a domain "expert" in a post-AI world, where knowledge is no longer worth what it was before?

These are the questions I get to explore. I'll share what I'm learning, but more importantly, what you can do to grow your career in a post-AI world.

Let's get back to my PTO question, because the answer I gave was incomplete. But it serves to make an important point.

Yes, California pays out accrued PTO…

None of that was in the original question. The person asking could not have known to ask about any of these caveats, nor should they. They came to us seeking knowledge. But instead they got something else…expertise.

This is the difference.

Knowledge answers the question asked. Expertise notices the question is wrong.

Now another example, with a bit more at stake.

A manager asks whether they can let someone go who has been having performance issues. The employment is at will. The knowledge-based answer is, technically, yes.

But that technically correct, knowledge-based answer can get you sued if the person filed a complaint three weeks ago and nobody mentioned it. Because now the question is not whether you can terminate. It is whether you can terminate without it reading as retaliation. Different question. Different answer. Vastly different consequences.

In this scenario, if the manager asked an AI and it said yes, it would have technically answered the question. But it would have no way to know about the complaint, and no way of knowing that's what mattered in this situation.

Now, this is not a story about AI being wrong. The output was right.

Also, models are getting better and cheaper seemingly every other week, so this is a situation that time and fine-tuning will certainly address. But there is a broader phenomenon I'm noticing.

In practice, most leaders will NOT trust AI for an answer in this situation. They would much rather speak to somebody with fifteen years of employee relations who reads that question and notices what was not said.

This "noticing" is the asset. This is the value that expertise brings that pure knowledge does not. But it largely goes invisible. More on why that is shortly.

If AI can give a correct answer to the wrong question, and that answer looks fine, the real question becomes: how easily can you tell?

Imagine a spectrum. On one end sits work where a wrong step is obvious almost immediately. On the other, work where you may never fully know.

Software code sits at the obvious end. If the code is wrong, it does not run. You find out in seconds.

Most HR people do not write code, but we deal with the same thing every day in a different form: data that has to be right before anything else can move. Running payroll is a good example. There are lots of ways to know when it is wrong.

Code and payroll are two sides of the same coin. A wrong step has feedback loops that announce themselves, and there are systems built to catch it.

Now put a hiring decision at the other end. Or a leveling call. Or a decision about who is ready for promotion.

These are the hardest to verify. There are no agreed-upon, built-in rules to run them against. Each of us brings our whole life's worth of context to the situation, maybe some concurrence, and makes a judgment call. And when feedback does arrive, months later (maybe years), it comes tangled with everything else that changed in the meantime. If someone you promoted struggles a year in, was it the promotion, the new manager they got in a reorg, or the market? You cannot cleanly isolate it. There is no flag in the system, and I don't know if there ever will be.

This is where the difference between knowledge and expertise stops being a philosophical point and becomes a career one.

AI tools are getting very good at the obvious end of the spectrum, because that is where a wrong answer can be caught and corrected quickly. AI learns based on feedback loops. It will get there.

Expertise stays needed where verification is hardest. And that is where most of what a People function does actually lives.

It puts HR in a more critical position for its expertise over the long term, not a less critical one.

A recent Harvard study, run inside a leading management consulting firm, tested this directly. Researchers gave 758 consultants a set of realistic work tasks and randomly gave some of them access to AI. These were individual-contributor level consultants, not twenty-year veterans, which matters for reading the results.

The researchers started from a hypothesis. There is a boundary, which they called the jagged frontier, between the tasks AI handles well versus the tasks it does not.

Eighteen tasks were designed to sit inside the frontier of what AI does well. Coming up with new product ideas. Writing marketing copy. Drafting a persuasive memo. The kind of work a capable generalist does well with a good brief.

One task was designed to sit outside of what AI does well. Consultants were given financial data on a company and interview transcripts with its people, and asked what the business should do.

But there was a twist. The numbers pointed one way and the interviews, read carefully, pointed another. Getting it right meant noticing that the obvious reading of the data was wrong.

Inside the frontier, AI helped a lot. Consultants using it were more than 30% better on quality and about 25% faster.

Outside it, the result flipped. Consultants without AI got the answer right ~85% of the time. Consultants with AI got it right 60% of the time in one group and 70% in another. On average, about 20 points worse.

Then the finding that matters most for this essay. The consultants using AI produced answers that were consistently rated as more coherent and more persuasive.

This rating held regardless of whether the answer was right or wrong.

This is the dangerous part. The wrong answers did not get rarer. They got harder to spot, because they showed up better organized than the right ones. The very qualities that make an answer feel trustworthy, that it is clean and confident and well argued, stopped being evidence that it was true.

Here is the part I think people get wrong about their own position.

If you feel like you have value because you know the answer, that value is gone.

I do not say that to be harsh. I say it because the people I see most exposed are senior and well-regarded, and they have spent years seeing the value they bring in exactly that way.

Coming back to this. What did not go anywhere is the power of noticing. And noticing is a strange asset. Why? Because noticing always came free with the answer. It came bundled, so it never had a separate line item. Now the answering has been unbundled from the noticing, and the noticing is still priced at zero and goes unnamed.

Two things follow from that.

If you work where the feedback is fast and obvious, where a wrong answer surfaces quickly and the work rests mostly on knowledge, the tools being built will inevitably catch up to you. That is good in one sense. They will help you do the job faster. But your judgment will not develop as much, because it does not have to.

If you work where the feedback is slow and ambiguous, you are in a better position to lean on your expertise and build on it. That is where the noticing is worth the most, and where it will stay worth the most.

For people entering the workforce: get really good at directing AI agents. It is a safe assumption that most junior talent will. They have been using AI through high school and college, and I would not bet against them. What we do not have an answer for yet is whether that produces the judgment they need to build domain expertise. That is a separate question, and it is being asked in a world that incentivizes speed. A senior HR leader put it to me plainly: if everyone can just ask AI, what is learning and development for? She did not have a good answer. Neither do I.

This has changed how I teach. I still teach people how to use the tools. But tools alone do not translate into the transformation organizations are looking for. What I have really started to do is teach people how to think: how to critically spot and articulate the problem, turn it into something AI can consume, and go through the cycles with it to build whatever solves it. The spotting, the noticing, understanding what is not being said. That is the whole discipline.

Take the decisions your function makes and sort them by one question.

How easily can we tell if we got this wrong?

Where the answer comes fast, let people move quickly. AI tools can be used to catch what is wrong, and they are good at it there.

Where the answer takes months and years and arrives tangled with everything else, that is where expertise is needed now more than ever. Have someone who knows the domain, not just the process, read the important calls. And treat a polished answer as a starting point, not as proof.

Then try one exercise with your team. Take a real question somebody sent you. Ask your team what the answer is. Then ask what would have to be true for that answer to be wrong.

The first question is the knowledge test, and you will find out quickly that most people have it.

The second question is the work.

If you manage people: how do you tell whether someone on your team is building judgment or just getting faster? Hit reply, I read all of them.