What organisations must change in leadership, culture and operating model as artificial intelligence keeps getting smarter

The constraint was never the technology

AI Adoption is close to universal and returns are close to absent. In 2025, 42% of companies abandoned most of their AI initiatives, up from 17% the year before.

Around 88% of organisations now use AI in at least one business function; roughly 6% are capturing significant enterprise-wide value. RAND puts the failure rate of AI projects above 80%, about twice that of conventional IT projects.

The comfortable reading is that these are teething numbers. They are not. Researchers working from very different data have found the same causes, and they are not the models: unclear definitions of success, weak data foundations, fading executive sponsorship, and organisational culture. BCG’s repeated finding is that roughly 10% of AI value comes from the algorithms, 20% from technology and data, and the remaining 70% from people, process and behaviour change. McKinsey’s high performers are nearly three times as likely to redesign workflows rather than bolt agents onto existing ones, and far more likely to have defined where human validation is required — 65%  against 23%.

So the binding constraint is not capability. It never was.

If the constraint is organisational, the response must be too. Three shifts matter more than the rest.

Ten miles southwest of Mendota, in California’s San Joaquin Valley, there is a power pole with three signs nailed to it.

• The top sign reads 1925
• The middle one reads 1955
• The lowest sign sits at about the height of a man’s boots, and it reads 1977

In the photograph, the hydrologist Joseph Poland stands at the base of the pole with one hand resting on that bottom marker. Over fifty-two years, the ground beneath him had dropped nine metres.

Nothing had been built there. There was
no earthquake, no landslide, no mine. Farmers had simply drawn water from beneath the valley faster than the valley could replace it.

Here is the part that hydrologists find hardest to explain to anyone else.
The valley did not sink because the
water left. It sank because the fine clay layers between the water-bearing sands, deprived of the pressure that had held them apart, collapsed in on themselves. And compacted clay does not spring back. The aquifer beneath Mendota can never again hold what it once held. The water can return. The capacity to store it cannot.

What has any of this to do with AI?

“A wealth of information creates a poverty of attention.”Herbert Simon, 1971

1. Leadership

From analysis to judgement

Analysis and judgement are not the same faculty, and the difference is about to become expensive. Analysis processes information accurately, efficiently and at scale. Machines now do this better than we do, and the gap widens each quarter. Judgement is something else: it integrates information with context, values, relationships, history and consequence, and it carries accountability.

The emerging risk is not that AI replaces leaders. It is that leaders quietly stop doing the deepest cognitive work the role has always demanded. Researchers at Wharton (Shaw and Nave) named the pattern in January 2026: cognitive surrender, the adoption of machine outputs with minimal scrutiny, overriding both intuition and deliberation. The risk is not automation. The risk is abdication.

There is a second-order problem, and it sits at the top of organisational level (but it might not be what you’re guessing). WalkMe’s 2026 survey of 3,750 executives and employees across 14 countries found that 9% of employees trust AI for complex, business-critical decisions — against 61% of executives. The people authorising the deployment are the most confident
about it; the people who will live with a poor one are the most sceptical. That is not a communications problem. The executives furthest from the work have
the least accurate picture of how it is going.

Capability without judgement is velocity in the wrong direction. Judgement without capability is a well-argued irrelevance. Judgement with capability is the required combination – now, and for the coming years.

2. Culture

From adoption to intellectual accountability

Most organisations are measuring the wrong thing. They are measuring adoption — seats, licences, prompts, hours saved. Almost none are measuring whether anyone disagreed with the machine and turned out to be right.

Research on deskilling makes the stakes plain: the capabilities leaders themselves rank as most critical to long-term performance — judgement and decision-making, problem framing, creative thinking — are precisely the ones most exposed to atrophy. Because AI produces acceptable ideas efficiently, independent ideation withers; because it is superb at pattern recognition, it creates the impression of rigour where (often) none has occurred. You will have heard the name for the residue: workslop, output that looks credible but does not hold up.

A culture that rewards speed of output over quality of reasoning will get exactly what it rewards, but may not notice for one to three years. The cultural work is therefore narrower and harder than the usual change program. It means making it safe, and then normal, to say that the machine is wrong. It means asking what the model missed before asking what it found. It means protecting a small number of consequential decisions from AI entirely, not out of nostalgia, but as deliberate practice — the organisational equivalent of a pilot hand-flying the aircraft so the skill survives the autopilot.

Just to be clear, this is not a critique of AI. The machine is often right, and is getting better all the time. But not interrogating it can prove fatal.

Reward the override and you keep the capability.
Reward only the output and you risk losing everything.

3. Operating model

Redesign the work, do not decorate it

An agent attached to one step of a process built around sequential human handoffs still runs at the speed of sequential human handoffs. This is Theory of Constraints compounded and it is why so much money has bought so little.

The organisations getting returns are redesigning end-to-end workflows around what agents can actually do, then setting autonomy deliberately rather than uniformly.

A workable discipline is three tiers:

• agents that assist,
• agents that act with human approval, and
• agents that act within policy.

The tier is set by the risk of the workflow, not the ambition of the vendor. Gartner’s 2026 survey found 17% of organisations had deployed agents and more than 60% expected to within two years, the fastest adoption curve among the emerging technologies it tracks. It also expects more than 40% of agentic projects to be cancelled by 2027. Both are true. The difference between them is design.

Which brings us to the structural change that will define this decade, and takes us back to the valley.

The drawdown

Organisations are flattening. Gartner projects that through 2026, one in five will use AI to eliminate more than half of their middle management positions. Manager headcount has already fallen 6.1% between May 2022 and May 2025 (Live Data Technologies),  over which time we have only seen the initial take-up of AI. At the other end, the entry-level share of advertised roles has dropped from over 44% to 38.6% in three years (ZipRecruiter), while Harvard researchers tracking 62 million workers across 285,000 firms found junior employment at firms adopting generative AI falling around 8% relative to firms that had not — driven by slower hiring rather than redundancy, while senior employment kept rising.

Each decision is defensible in isolation. Together, they could be disastrous.

Entry-level roles were never only about output. They were structured learning environments. The junior analyst who builds the model by hand acquires an intuition for when a model is wrong, and acquires it in no other way. Middle management was where people learned to translate strategy into execution, and where an organisation noticed early that the plan and the reality had drifted apart. Remove both and you are pumping hard from a reservoir of judgement accumulated over thirty years, while paving over the ground where the rain used to soak in.

Call it judgement debt. It is borrowed quietly, it appears on no line of the accounts, and it falls due long after the executives who incurred it have moved on. An organisation that dismantles its middle in 2026 is dismantling the apparatus that produces its senior leaders in 2032. And some of what compacts does not spring back: the people who would have grown into those roles have gone elsewhere, and a good number of those who might have stepped up have concluded the job is not worth wanting.

Not everyone is drawing down. IBM announced it would triple entry-level hiring in the United States this year, reasoning that junior developers relieved of routine coding spend their time with customers instead. The early years are where judgement is forged. It is a bet on the recharge zone.

Where to start

None of this requires waiting for clarity about where the technology lands (and you certainly shouldn’t wait for that ever-moving horizon). Here are six steps you could take this quarter:

Your first step should be assessing where you are on AI maturity, where you want to get to, and how quickly you are aiming to do so. You can use Bendelta’s AI Maturity Model. [NB Our next white paper will explain its foundations and use in detail.]

3H (Head-Heart-Hands) AI Maturity Model (© Bendelta, 2026)

Then the next 5 steps are:

• Measure the override rate, not the adoption rate. How often did someone disagree with the machine, and how often were they right? What can you learn from that?
• Set autonomy tiers by workflow risk.
  Assist, act with approval, act within policy.
• Ring-fence a small number of consequential decisions as AI-free. Treat them as practice rather than principle.
• Redesign some junior roles instead of deleting them. Make sure they are designed to build judgement.
• Put at least 50% (but ideally more like 70%) of the AI budget into people. The evidence suggests that this is what separates the 6% getting real economic value from AI from the other 94%.

The markers on the pole

Poland’s photograph was taken in 1977 and became one of the most reproduced images in hydrology. It changed very little. The valley kept pumping, because each additional season was rational for the farmer who did it, and the cost fell on everyone, later, in a currency nobody was measuring.

The organisations that thrive as machine intelligence compounds will not be the ones with the best models. Those will be available to everybody, cheaply, within months of one another. They will be the ones that kept a deep reserve of human judgement, knew when to draw on it, and built the solutions that grew and recharged judgement.

Every organisation is now nailing signs to a pole, based on AI. Only those organisations judiciously investing equally in leadership, culture and operating models will like the numbers they see on their sign in three years’ time.

Sources: McKinsey, The State of AI (2025); RAND (2024); S&P Global Market Intelligence (2025); BCG, 10-20-70 and BCG Henderson Institute on deskilling (2026); Gartner 2026 CIO and Technology Executive Survey; Hosseini and Lichtinger, ‘Generative AI as Seniority-Biased Technological Change’ (Harvard); ZipRecruiter 2026 Graduate Report; Live Data Technologies; WalkMe, State of Digital Adoption 2026; Shaw and Nave (Wharton, 2026); Bloomberg; USGS.