Most of the conversation about AI at work is running a year ahead of the conversation about work itself. We talk endlessly about what the agents and copilots can do. We spend far less time asking whether the jobs and reporting lines they are landing in still make any sense.

That mismatch is the thing I keep snagging on. New tools ship every few weeks. Companies, when they manage it at all, reorganize over years. So the question worth a leader’s time right now isn’t which tool to buy. It’s what needs to change underneath it.

I’ve spent the last several years running AI strategy inside large industrial operations, and honestly, the same short list of questions comes up in nearly every room.

• Which human skills are worth more now?
• Are we redesigning jobs or just shaving tasks off them?
• What does a team look like in five years?
• And when a machine helps make the call, whose name is on it?

Nobody has clean answers yet, me included. What follows is just where I’ve landed so far, and I reserve the right to be wrong about a fair amount of it.

The skills that appreciate

Every wave of automation before this one moved the value of human skills around rather than deleting it, and I expect the same here. What is different this time is the specific gap the technology leaves. AI is remarkably good at answering the question you hand it, and completely unable to tell you that you have handed it the wrong one. That gap is where the human work is going.

A few capabilities are clearly worth more as the routine work gets absorbed:

1. Framing the problem, not just solving it

The rare skill now is stating the problem right in the first place. This means taking a messy business situation and turning it into something a model can actually chew on, and having some sense of what a good answer would look like even before you go asking the question.

On a production line, an AI will cheerfully flag every anomaly it can find. Working out which of those anomalies anyone should care about is still a person’s job, and frankly it’s the harder half.

2. Judgment you can be held to

A model can hand you a recommendation, but it cannot be the one accountable for it. As more decisions get generated for us, the human job quietly shifts to owning them. That is, deciding when the output is trustworthy, when to overrule it, and standing behind whatever happens next. That last part does not automate.

3. Knowing good from confidently wrong and people skills

When analysis gets cheap, and everything sounds plausible, being able to spot the answer that is fluent and wrong becomes valuable. And sitting right next to that are the things AI does not touch at all.

For example:

• Getting a nervous team to trust a new way of working.
• Talking a skeptical executive into it.
• Holding people steady through a change they did not ask for.

In my experience, most AI efforts do not stall at the model stage. They stall on the humans around it, which means the people who are good with those humans matter more now.

Rethinking jobs, not just automating tasks

Here is the pattern I see most often: A team finds a task the AI can do, automates it, and leaves the job description exactly as it was. The work gets a little faster, and nothing else changes.

It is the corporate version of paving the cow path: you make the crooked route smoother instead of asking whether the route should exist. You bank a small efficiency and walk right past the bigger prize.

A job, when you look at it honestly, is just a bundle of tasks that once made sense to hand to a single person. AI pulls that bundle apart. Some of the tasks drop to almost no cost, a few of them suddenly matter more, and what is left over often adds up to a different job than the one you started with.

If you only automate task by task, you end up polishing a bundle that has lost its reason to exist. A couple of things have helped me think about it differently:

1. Build the role around what the person is best at, not the scraps

The easy, lazy version of redesign is “let AI take the boring parts and people keep whatever’s left.” I’d flip it around. Start from what the person is genuinely best placed to own.

Usually that’s judgment and accountability, sometimes it’s the relationships, often it’s the ugly edge cases nobody wrote a procedure for. Build the job around that, and let AI carry the support work underneath.

2. Expect the work to move up a level

Give someone a capable AI partner and a lot of jobs shift upward. The analyst starts reviewing analysis. The writer starts editing. The operator starts supervising a process that mostly runs itself.

If the job description still reads the way it did before any of this, the redesign has not actually happened yet. And it is rarely one job in isolation. When AI eats the handoffs between roles, the lines between those roles usually need to move too, and you will miss that if you only ever look at one seat at a time.

The workforce in 5 years

Five-year AI predictions are mostly a way to be wrong in front of an audience, so take this loosely.

My one real conviction is that the popular framing is off. People ask about the “balance between human employees and digital workers,” as if the answer is a ratio, so many humans to so many bots. I do not think the ratio is the interesting part.

The interesting part is the wiring: what gets delegated, who is watching it, and where the accountability finally lands.

My rough expectation is that agents stop being a novelty and just become a layer of the workflow. Today’s copilots help you a turn at a time. The direction of travel is toward agents that run a multi-step process on their own while a person watches by exception, checking the odd cases instead of every step.

That is a real shift, but it is more a maturing of something we already do than some clean break with the past. If that plays out, the org chart sprouts a supervision layer. More people spend their day directing, checking, and answering for a set of agents rather than doing the task by hand.

The valuable person becomes the one who can actually run an AI-heavy process: stand it up, keep an eye on it, notice when it drifts, and own the result when someone asks.

And adoption is going to be far patchier than the headlines let on. This is the part I think most forecasts miss. Anywhere a decision carries real liability, or has to be defended after the fact, things move slowly on purpose, and rightly so.

The gap between what you can demo and what you can run at scale in a controlled setting is enormous. Sooner or later the bottleneck stops being the technology at all. It becomes trust and plumbing. Verification, liability, change management, the tangle of legacy systems, and plain human willingness to take a hand off the wheel.

None of that moves on the release calendar. It moves on people-time, which is slower and messier.

Who owns workforce transformation?

Ask who should own this and the reflex is to name one owner: IT, HR, or the business. Every single-owner answer is wrong.

Worse, the usual way of carving it up, where IT builds the tech, HR grows the talent, and the business owns the outcome, is basically a description of the fragmentation you were hoping to avoid.

Hand it to any one of them, and you can call the failure in advance. IT on its own builds slick tools nobody touches. HR on its own runs training bolted onto nothing real. The business on its own buys point solutions with no talent plan or governance holding them up.

So the real question is not who owns it, but how you share ownership without it splintering.

The version that has worked for me: The business owns the outcome. One leader owns the redesigned process and its number, whether that is P&L or a quality metric, and carries it. The functions own the capabilities, the technology, and the people. That way, there is still one person to hold responsible even though the delivery is spread across teams.

The bit that almost always goes missing is the connective tissue. These programs come apart because IT, HR, and the business each polish their own corner and nobody owns the joints between them.

Owning that coherence, making sure the tool, the retraining, and the process change are all aimed at the same target, is exactly the job of an AI strategy function. But it only works with a sponsor senior enough that none of the three can quietly bow out, and only if there is still one accountable owner who can actually make a call.

Shared ownership with no decision-maker is just fragmentation wearing a nicer label.

The thread running through all of it

Line these questions up and the same thing keeps showing through. AI is excellent at execution and cannot be responsible for anything, so the human contribution keeps sliding upward, into framing, judgment, checking the work, and owning the result.

The real challenge for a leader is to build the roles, teams, and guardrails around that upward shift, not around whichever tasks the AI happened to swallow first.

The companies that come out ahead this decade won’t be the ones with the longest list of deployed tools. They’ll be the ones that treated all of this as a design problem, made a deliberate call about what stays human and why, and put enough trust and accountability in place that people could hand real work to a machine without losing track of who answers for it.

Done right, that discipline doesn’t slow the progress down. It’s the thing that keeps the progress from quietly falling over a year later.