What's Left for Leaders to Decide When AI Gets Stronger

As AI gets stronger, something has to change about what a leader actually does. Not eventually. Now. Each new generation of models absorbs more of the work that used to define the job, and the model capable of taking a genuinely open brief and finding its own way through it - rather than simply executing instructions - is a recent arrival. For a digital leader specifically, the question isn't whether that shift is coming - it's what's actually left to do once most of the how, and increasingly a fair amount of the what, no longer needs them personally.

Managing AI is still managing people

The deepest lesson in management fits in one line: the more capable a person is, the more the instruction given to them has to change. That lesson doesn't stop applying because the recipient is a model instead of a person.

A first-year apprentice is handed a procedure sheet and follows it exactly, because they haven't yet built the judgement to fill in what the sheet doesn't say. A licensed tradesperson is handed the outcome the client wants and a short list of constraints and works out the method themselves. A site supervisor with twenty years in the trade is handed a problem, not a method at all, and comes back with an approach nobody on site would have specified in advance. Same trade, same tools, three completely different instructions, because the instruction was matched to the judgement already sitting on the other end of it.

Hand an execution-level instruction to someone operating at supervisor level and it reads as distrust. Hand a supervisor-level brief to the apprentice and the result is paralysis, not progress. The mismatch is the failure, not the person.

Where AI's capability actually comes from

It's tempting to treat "the model got stronger" as a single, external event - something that happens to an organisation on a release date. In practice, AI capability comes from two different places and only one of them is out of an organisation's hands.

• The first is the model itself. Every new generation genuinely reasons further, holds more context and handles more ambiguity than the one before it. That part is real and it will keep happening regardless of what any single organisation does about it.

• The second is almost entirely within an organisation's control, and most organisations under-invest in it: the context, history and pattern the model has actually been given to work from. A model with no knowledge of an organisation's standards, its past decisions, its tone, its audience and what has and hasn't worked before will operate at execution level even when it's technically capable of far more, because it has nothing else to reason with. Feed it that context deliberately and the same model starts operating closer to the tradesperson than the apprentice - not because it got smarter overnight, but because it was finally given something worth applying its intelligence to.

This is why two organisations using the identical model can get very different results from it. The gap sits somewhere else entirely: in how much of the organisation's own judgement was ever handed over for the model to build on.

So, now that it's stronger, what is the role?

Put both of these together and the territory that's exclusively the leader's keeps shrinking toward the top. As the model itself grows more capable, and as organisations get better at giving it real context to reason from, more of the how stops needing a person at all. A growing share of the what - the strategic choices that used to be the leader's clearest point of value - is moving the same way. A model given a genuinely open, well-contextualised brief can now produce a strategic option nobody on the team had thought of.

Peter Drucker's most quoted line draws a boundary between two layers: efficiency is doing things right, effectiveness is doing the right things - execution against strategy. Drucker never formally named a third layer, but his idea points straight at it: the judgement of which questions, out of everything competing for attention, genuinely deserve it, given that time and energy are finite. Extending his framework this way gives three layers to work with. How to do it: execution, the layer of doing the work well. What to do: strategy, the layer of choosing the market, the product, the path. What deserves attention at all: the layer underneath both, where judgement rather than analysis makes the call. Call it the judgement layer.

The first two layers are exactly where AI has been closing the gap fastest, for the reasons above. The judgement layer has not moved, at any point, so far.

Calculation has an answer. Choice doesn't

None of this means AI can't lay out options, build the decision tree or weigh the trade-offs. It can, and often faster and more thoroughly than a person working under deadline. What it cannot do is make the final call once the analysis is finished, because that call rests on values, appetite for risk and a view of what the organisation is actually for, not on the calculation itself. Two organisations facing the identical set of AI-generated options can, and reasonably should, choose differently, because the right answer was never purely a function of the evidence in front of them.

How to actually think about what's worth thinking about

The judgement layer is easy to agree with and hard to operationalise. Four practical tests help separate a genuine leadership decision from a well-dressed calculation that never needed the leader specifically:

• The delegation test: if handing this decision to a model or a team would visibly signal that nobody is actually leading, the decision belongs to the leader - that discomfort is the diagnostic, not a feeling to override

• The reversibility test: a decision that can be trialled, measured and reversed within weeks is a candidate for AI-assisted iteration, while a decision that commits the organisation for years, or is expensive to walk back once made public, deserves the leader's personal attention however clean the supporting analysis looks

• The residual-discomfort test: when every option has been modelled, every trade-off quantified and a genuine reluctance to commit still remains, that reluctance is the values question surfacing - a sign the decision was never actually a calculation

• The slow-burn test: compounding decisions about culture, trust and who gets developed never feel urgent, so they lose to whatever is loudest this week - they need deliberately protected attention precisely because urgency will never nominate them

None of these tests require rejecting AI's input. They require noticing which decisions AI's input was never going to resolve in the first place.

Consider a hypothetical that most digital leaders will recognise: a mid-sized retailer facing declining footfall. AI-assisted modelling lays out the options cleanly - close the worst-performing stores and reinvest in e-commerce, renegotiate leases and hold the current footprint, or split the difference by closing a handful of the weakest sites while converting others to showroom formats. Every scenario comes with defensible unit economics. The store-closure decision fails the delegation test immediately: handing it to a model would read, correctly, as management abdicating a call that defines the company for the next decade. It fails the reversibility test too: leases surrendered, staff let go and landlord relationships ended are not easily restored once the quarter closes. And it carries the exact signature of the residual-discomfort test - the modelling is as thorough as it will ever be, every trade-off is quantified and leadership still hesitates, because the real question underneath the spreadsheet is what kind of retailer the company wants to be. That is the call the AI-generated options were never going to make.

What this changes about how the role gets measured

Most performance conversations for digital leaders still centre on output and delivery: what shipped, what launched, how fast. Those remain real measures, but they increasingly measure the layers AI is now the strongest contributor to. A harder question is becoming the more revealing one for a leadership review: of everything that could have been pursued this quarter, was the right thing chosen to focus on, and does the reasoning behind that choice hold up on its own terms, independent of whatever AI-generated analysis supported it?

The question worth sitting with

A common assumption is that stronger AI pushes people down the value chain, toward whatever is left once the model has taken the rest. The more accurate version looks like the opposite: as AI absorbs execution and starts contributing genuinely well to strategy, people get pushed upward - from executor toward manager, from manager toward strategist, from strategist toward something closer to a philosopher, whether they intended to end up there or not.

The real test sits closer to home: looking at everything AI can now make plausible, did this quarter's attention go to the handful of questions that were actually worth it, or to the loudest ones?