The new operating model is the moat

AI drags organisational design back onto the CEO agenda

There is a pattern that repeats across every major technological disruption. Leaders treat it as a tool problem first, a strategy problem second, and an organisation problem last. By the time the organisation question becomes unavoidable, the decade is already lost.

Agentic AI is following the same script. Most boards have approved an AI strategy. Most executive teams have licensed the tools, funded the pilots, and appointed someone with the word "AI" in their title. Very few have sat with the harder question. What does this company look like, structurally, when a meaningful share of execution work is done by software that plans, acts, and iterates on its own?

That question belongs to the CEO and the COO. It is not an IT problem, not a transformation office problem, not a centre of excellence problem. It is an operating model problem, and it has returned to the top of the agenda after nearly two decades of being delegated downward.

A note of honesty up front. The frontier is uneven. Agents are startlingly good at some tasks and embarrassingly bad at adjacent ones. They handle high volume classification, routing, drafting, and structured reasoning with real reliability. They are still questionable on long horizons, where the context itself is ambiguous, dare I say contradictory, and where the rules of the process are fluid. The leaders who win will design for that jaggedness, not wish it away.

Execution is commoditising, but only where the work is bounded

The loudest version of the AI story says execution is collapsing in cost everywhere. That is not quite right, and the imprecision is expensive. Execution collapses in cost inside a specific shape of work: bounded outcomes. You can instruct an agent with confidence on the expected inputs, outputs, and exceptions. And where ambiguity remains, you can still instruct around it with guardrails, risk tiering, and exception management. That is what makes agentic execution performant at enterprise scale. Everything else is moving more slowly, and in some places, not moving at all.

The bounded category is larger than it sounds. Tier one support, first pass underwriting, invoice and purchase order processing, contract triage, claims intake, customer onboarding, large parts of software delivery, routine analysis, most of what we politely call knowledge work. These are outcomes where a human today converts structured or semi structured inputs into a small set of possible outputs, and where exceptions can be defined and routed. In these outcomes, agents are not a pilot. They are a structural change.

The unbounded category is where most of the breathless claims fall apart. Open ended strategy, novel investment theses, cross organisational judgement calls, messy stakeholder conflict, anything that requires stitching together context that has never been written down. Agents help here. They do not replace the work. Pretending otherwise leads to the pilot graveyard that every large company already has.

For a vivid illustration, see Anthropic's Project Vend, where Claude was given a month to run an automated office shop and broke down exactly where the work turned unbounded.

This distinction matters because it reorders the executive to do list. If you treat AI as a uniform wave, you plan for a future that is not arriving. If you treat it as a sharp edge cutting through bounded work, you plan for a future that is already here, and you plan it with precision.

The imagination gap is the real bottleneck

There is an irony sitting in the middle of this moment, and it deserves to be named. The executives most trusted to steer organisations through uncertainty, those with the deepest domain expertise and the longest track records, are also the ones most likely to be quietly blinded by their own experience.

This is not a criticism. It is a structural observation. Judgement is built from pattern recognition, and pattern recognition is built from cycles already lived. The senior leaders in every large company were shaped in a world where technology was expensive, slow, and unforgiving. Where a failed programme cost millions and stalled careers. Where the safe answer was nearly always to move cautiously, validate extensively, and build incrementally.

Those rules of thumb are now actively misleading. The cost of trying has collapsed. The cost of waiting has gone up. The asymmetry has inverted, and most experienced leaders are still pricing risk using the old curve.

The leaders who move through this transition well tend to look different on paper. Not because they are younger or more technical, although sometimes they are both. They are different because they combine real domain knowledge with real technical fluency and the willingness to make decisions at the edge of their own comprehension. Broad enough to see across functions. Deep enough to know what an agent can actually do in a specific sub-domain. Bold enough to act before the full picture is clear, because in this era the picture is never going to be fully clear before the window closes.

The pattern shows up concretely in three places. First, leaders who have built something with AI themselves, even a small thing, make materially better calls than leaders who have only been briefed on it. Second, it won't surprise me one bit if the T shaped profile, narrow and deep plus wide and curious, outperforms the traditional deep specialist at the executive level for the first time in a generation across all industries. Third, personal identity gets in the way more than anyone admits. Many executives have spent twenty years becoming the CFO, the CHRO, the CIO. Asking them to reopen what their function even is, from first principles, is not a small ask. It is slow work, and it is human work, and it will not be solved by a slide in a town hall.

This is where the most senior leaders either compound their advantage or quietly become the bottleneck.

Debt does not disappear. It moves.

A seductive story has taken hold in AI circles. That the collapse of human comprehension as a constraint, agents reading what humans cannot, agents refactoring what humans cannot hold in their heads, is wiping technical debt off the balance sheet. That is half right, and the half that is missing is the half that bites.

It is true that one specific kind of debt is shrinking. Comprehension debt. The accumulated cost of code written to be understood by the next human who may never arrive, of dashboards designed for finite cognitive bandwidth, of pipelines shaped for auditability rather than raw efficiency. An agent does not need a readable codebase. It does not need clean naming conventions. It can work with whatever representation is most efficient for the machine and rewrite it on the way through. That category of debt, the one that compounded every year for the last four decades, is genuinely getting smaller.

But debt does not disappear. It moves. And the new ledger is unfamiliar enough that most organisations are not even tracking it yet.

Evaluation debt. The quiet accumulation of agent behaviours that nobody has stress tested under the conditions that actually matter. The gap between the benchmarks the vendor ran and the edge cases in your business. Every deployed agent without a living evaluation harness is a liability with no maintenance schedule.

Provenance debt. The growing chain of decisions where nobody can cleanly answer, after the fact, what model produced what output, on what version of what data, under what policy. Regulators will ask. Auditors will ask. Customers will ask.

Permissions debt. Agents need access to systems to be useful. Most companies have not touched their identity and access architecture since they built it for humans. An agent with an employee's credentials is an audit finding waiting to happen.

Rollback debt. What happens when the agent is wrong at scale. Not one mistake. A thousand mistakes in a weekend, spread across customers, with no easy way to undo.

Governance debt. Policies written for a world in which humans made the consequential calls, now stretched over a world in which they no longer do. The gap widens quietly.

The net is that technical debt has not gone away. It has rotated, from the bottom of the stack to the top. From code that humans have to read, to behaviour that humans have to trust. Most organisations are celebrating the first and ignoring the second. That is where the next wave of expensive surprises will come from.

Not every agent is the same kind of risk

The conversation about AI risk has collapsed into a single, unhelpful shape. Either it is existential and unknowable, or it is a compliance checklist. Executives who are actually deploying agents need something between those two. A working taxonomy that tells them where to move quickly, where to move cautiously, and where to not move at all without serious institutional scaffolding.

Three tiers are enough to start. They are defined by consequence, not by technology.

The taxonomy is useful not because it solves the risk problem, but because it helps sort it. Most executive teams today are treating every AI use case with the same caution, which sounds responsible, but is paralysing. The cautious thing is to move fast in tier one, invest in scaffolding in tier two, and build real governance in tier three.

Two plausible futures for the same function

Ask any executive team today what their function looks like in three years and you will get one of two answers. Both are defended with conviction. Both cannot be fully true. And the honest position is that we do not yet know which one will dominate. It is worth holding both in view, because they lead to very different decisions about hiring, organisational design, and succession.

The most honest reading is that both theses will play out, in different functions, in different companies, at different speeds. The organisations that assume only Thesis A will discover Thesis B as a competitive surprise, usually from a newer entrant with a cleaner sheet. The organisations that lean into Thesis B too aggressively will discover that the judgement, context, and relationships they just let go were the actual product.

This is not about picking a thesis. It is about being deliberate per function and bold in designing the path accordingly. That is what an AI organisation design actually looks like. It is a set of targeted bets, held with honesty about what we do not yet know.

Six disciplines to pay attention to in the new operating stack

If execution is collapsing inside bounded outcomes, the obvious executive question is, what do we now need to pay attention to? Six things stand out, and they compound. A company that is good at one and bad at the others will not win. A company that gets most of these right will pull away.

These are the six. They are deliberately phrased as short, concrete labels rather than grand concepts, because this is where the operating model actually lives.

Five moves for CEOs and COOs

Pulling this together, there are five moves that every CEO and COO should have in motion inside the next two quarters. None of them are speculative. All of them are overdue in most companies.