Every enterprise now has AI in use somewhere. Most have it in use in many places. The tools are better than they were eighteen months ago. The investment is real. And yet, for most organisations, the gap between what AI is demonstrably doing and what the board was told it would deliver has not closed.
This is not a tooling problem. It is not a training problem. It is not a problem that more roles, more governance committees, or more frameworks will address. What we tend to find, when we look at where the expected gains have not materialised, is something structural — a mismatch between how AI is being introduced and how the organisation is designed to work.
In Brief
- The bottleneck has moved: the challenge is no longer building fast enough, but validating whether what is being built is safe, secure, and worth it.
- Adapting AI to existing work structures captures automation gains at the task level — it does not produce enterprise-level value.
- McKinsey’s 2025 research shows 88% of organisations use AI, but only 39% report any measurable impact on enterprise earnings.
- Capturing value from AI requires a different way of organising work — not more roles, more management committees and working groups, or more training on existing frameworks.
- An organisation that redesigns work around AI capability operates from a different strategic position than one that inserts AI into existing work.
AI produces faster than organisations can review
For the better part of two decades, the dominant question in large-scale delivery was whether organisations could build and deploy fast enough to keep up with demand. That problem generated an enormous market in frameworks, certifications, and training programs, all oriented toward accelerating delivery throughput at scale.
That bottleneck has moved.
The challenge for organisations operating with AI at any material scale is no longer whether they can build something in the time available. It is keeping up with the need to validate whether what they are building is safe, secure, and valuable. AI can produce outputs at a speed that has decoupled delivery from the validation mechanisms built to review it. The frameworks, governance layers, and quality gates that existed to manage human-paced delivery were not designed for this. Most of them are still in place, unchanged, while the rate of output they are expected to review has accelerated dramatically.
The structural consequence is predictable. McKinsey’s 2025 State of AI research found that while 88% of organisations report regular AI use in at least one business function, just 39% report any impact on earnings before interest and tax at the enterprise level (McKinsey & Company, 2025). This gap persists despite significant investment in AI tooling and workforce capability. The explanations that point to immature technology or undertrained teams do not account for organisations that have addressed both and still find the impact does not reach the enterprise level. What does account for it is something earlier in the causal chain: the gap reflects a structural condition, where AI is running faster than the organisational model built to govern and direct it.
What makes this pattern so persistent is that the existing approach is defensible from the inside. Each individual decision — fund this use case, add this role, extend this governance layer — is rational at the decision point. The difficulty is that a collection of individually rational decisions does not add up to a different operating model. It adds up to the same operating model with AI tools embedded in it. This is why more of what the organisation was already doing — more roles, more governance layers, more training in current frameworks — tends to produce marginal returns. It is solving a throughput problem in a constraint that is no longer about throughput.
Inserting AI doesn’t change the structure
The pattern that surfaces most reliably across large enterprises is what might be called AI-into-existing-structure: the approach of identifying where AI can improve what the organisation is already doing, and inserting it there. This is not irrational. It is the lowest-risk path to demonstrating early returns. And it does produce returns — at the task level, often significant ones.
The problem is that task-level automation gains do not aggregate to enterprise-level value unless the work structure itself changes. Gartner’s research is blunt on this point: over 40% of agentic AI projects are forecast to be cancelled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025). The risk controls problem in particular is a symptom of the structural mismatch — governance designed for human-paced work cannot absorb AI-generated output at AI pace without either slowing AI to human speed, or accepting that much of what AI produces is not being properly reviewed.
The deeper issue is that organisations working this way are optimising the wrong variable. They are measuring AI by its ability to accelerate what the organisation was already doing. The research suggests this is not where the enterprise-level value lies. McKinsey’s research on the agentic organisation indicates that organisations beginning to capture material AI value are moving toward operating models built around AI’s distinct capabilities: operating continuously, sharing context at scale, and executing across workflows that would require many human handoffs (Sukharevsky et al., 2025). The organisations capturing material enterprise value are redesigning the work, not accelerating the existing version of it.
The gap between adoption and impact is structural
The research is consistent on what separates organisations capturing AI value from those that aren’t — and it is not which tools they use or how many.
Sukharevsky et al. (2025) identify that organisations beginning to capture material AI value are moving toward flatter work structures where teams share context across boundaries and hold accountability for outcomes rather than delegating upward through a hierarchy. McKinsey terms these structures “agentic networks” — the critical shift is not in the tools used but in how work is coordinated: who holds what decision, how information moves, and what governance looks like when it must operate at AI speed rather than at committee cadence.
The evidence on governance is particularly relevant. Gartner (2025) notes that governance in AI-era organisations cannot remain a periodic, paper-heavy exercise — it must become real-time, data-driven, and embedded, with humans holding final accountability. This is a structural change to how oversight works, not a marginal extension of what oversight was doing before. Adding AI governance roles to an existing governance structure does not address this. The mechanism by which governance operates must change.
McKinsey’s broader research on operating models (Krivkovich et al., 2025) identifies that even high-performing organisations carry a 30% gap between strategy’s full potential and what is actually delivered. McKinsey attributes this to shortcomings in their operating models, not their strategies. The addition of AI does not fix this gap. In most cases it widens it: AI accelerates the delivery side of the equation, increasing the volume of output the operating model must absorb, without addressing the organisational conditions that were already preventing that output from paying off.
The implication is that organisations looking to resolve their AI value gap through investment in additional roles, additional training in existing frameworks, or additional governance layers are unlikely to close it. Not because those investments are without merit, but because the gap is structural. The tools the organisation is adding are operating in a structure that was not designed to take advantage of what those tools can do.
What a different work structure looks like
The organisations beginning to capture enterprise-level AI value share a structural feature: they have changed how work is organised, not just what tools are used to do it. This is not a technology decision. It is an operating model decision.
Specifically, what tends to shift is the unit of accountability. In structures designed for human-paced delivery, accountability runs through a hierarchy: work is assigned, reviewed, escalated, approved. In structures designed to take advantage of AI, accountability moves closer to outcomes — teams hold responsibility for results rather than for executing defined tasks, and AI handles the execution of the tasks that would otherwise have consumed their capacity. The human role shifts from doing to deciding: deciding what to pursue, assessing what AI has produced, and holding accountability for whether the outcome was worth pursuing.
This restructuring is not achievable by adding AI to an existing team without changing what the team is accountable for. A team that was previously accountable for building features will use AI to build features faster. That is useful. It is not the same as a team accountable for customer outcomes that uses AI to explore a wider range of options than any human team could assess. The structure of the accountability determines what AI’s speed and scale actually produce.
The governance question follows directly. If teams hold outcome accountability, governance shifts from reviewing outputs — what was built, how long it took — to reviewing evidence: whether the outcomes the team was accountable for were achieved, and what the team learned from the attempts that did not achieve them. This is a different governance rhythm, operating on different information, with different questions at its centre. It requires AI-generated evidence to be interpretable by the governance process, which in turn requires the governance process to be designed for that purpose from the start — not retrofitted.
None of this is resolved by adding roles to an existing structure. It is resolved by asking what the structure is for — what it is designed to produce — and then designing it again with AI capability as a core assumption rather than an addition.
The organisations that will establish a durable position from AI are not the ones that deployed the most tools. They are the ones that understood, early enough to act on it, that the value was in the work structure those tools made possible — and that the time to build that structure is before the tools are running, not after.
The operating model is still answering the wrong questions
An operating model designed for AI is not a technology design. It is an answer to six questions that executives are already accountable for — accountability, risk visibility, cost trajectory, decision interpretability, control design, and capability ownership. The difference between an operating model that captures AI value and one that doesn’t comes down to how it answers each of those questions. Most organisations have answered them in ways that made sense before AI was a material part of how work gets done. The answers need to change.
| What the executive is accountable for | An operating model designed for AI | An operating model with AI added to it |
|---|---|---|
| Accountability: who carries the outcome, not just the output | Outcome accountability sits with teams who direct AI and assess what it produces | Accountability sits with roles defined by tasks — the task is done when AI completes it, regardless of whether the outcome was achieved |
| Risk visibility: would you know before it became a problem | Risk is surfaced continuously through AI-generated evidence that governance reviews in near real time | Risk is surfaced periodically through reports assembled after the fact — by the time the review happens, the exposure has already occurred |
| Cost trajectory: is this getting cheaper or more expensive as it scales | Cost per outcome falls as AI handles more execution and teams focus on higher-value decisions | Cost rises with scale because AI output requires human review that was not redesigned for AI pace — more output means more reviewers |
| Decision interpretability: can you explain a decision if called to account for it | Decisions are logged, traceable to the human who directed the AI, and reviewable against the criteria used to assess the output | Decisions are difficult to reconstruct — the AI acted, but who directed it, on what criteria, and who verified the result is often unclear |
| Control design: could a regulator, auditor, or minister understand how this is governed | Controls are embedded in how work runs — approval authorities, scope boundaries, and override mechanisms are defined before AI operates, not applied after | Controls are applied to AI output the same way they were applied to human output — periodic, manual, and not designed for the volume or pace AI produces |
| Capability ownership: does this build something the organisation owns, or does it create dependency | The organisation builds judgment about what AI should and shouldn’t do — that judgment is an organisational capability that persists | The organisation builds proficiency in using specific AI tools — when the tools change, the capability leaves with the vendor contract |
The column on the right describes most organisations operating with AI today. The column on the left describes where the enterprise-level value is. The distance between them is not a technology gap. It is an operating model gap — and it is the gap that determines whether AI investment reaches the enterprise level or stays at the task level.
What this means for senior leaders
- AI value at the enterprise level requires structural change, not just AI adoption — the organisations reporting material impact on earnings have changed how work is organised, not just what tools their people use.
- Governance designed for human-paced review cannot absorb AI-generated output at AI pace — it will either slow AI to human speed or leave most of what AI produces unreviewed.
- Task-level efficiency gains are real but do not aggregate to enterprise value without a change in what teams are accountable for — the unit of accountability must shift from executing tasks to achieving outcomes.
- Organisations that redesign work structures around AI capability now will be better placed than those that retrofit — the structural gap between early movers and later ones widens as AI capability scales.
References
Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
Krivkovich, A., Di Lodovico, A., Weddle, B., Maor, D., Mahadevan, D., & Steele, R. (2025, June 18). A new operating model for a new world. McKinsey & Company. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/a-new-operating-model-for-a-new-world
McKinsey & Company. (2025). The state of AI in 2025: Agents, innovation, and transformation. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
Sukharevsky, A., Krivkovich, A., Gast, A., Storozhev, A., Maor, D., Mahadevan, D., Hämäläinen, L., & Durth, S. (2025, September). The agentic organization: Contours of the next paradigm for the AI era. McKinsey & Company. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era
