The AI Ambition Gap

AI adoption is accelerating. Enterprise value is not keeping pace.

Few technologies have moved from demonstration to daily use as quickly as generative AI. Employees use it to draft, analyse, summarise, code and search. Organisations have launched copilots, internal assistants, pilots and increasingly autonomous agents. AI is no longer waiting outside the enterprise.

Yet adoption is not the same as transformation — and neither guarantees value.

McKinsey’s 2025 global survey captured the tension clearly: 88 per cent of respondents reported regular AI use in at least one business function, while only 7 per cent described AI as fully scaled across the organisation. Industry studies report a similarly wide gap between the number of pilots launched and the number producing measurable financial outcomes.

The precise failure rate varies with the population, the definition of success and the time horizon. A pilot that has not yet affected the P&L is not necessarily a failed technology. But the pattern is difficult to dismiss: companies are becoming better at acquiring AI than at converting it into a new organisational capability.

The strategic question is no longer whether an organisation is adopting AI. It is whether AI is changing the economics, speed or quality of the system around it.

The productivity paradox is not new

Every general-purpose technology creates a period in which the tool is visible before its full economic effect is.

Erik Brynjolfsson, Daniel Rock and Chad Syverson describe this through the Productivity J-Curve. New technologies require complementary investments that conventional accounting and productivity measures struggle to capture: redesigned processes, new skills, data infrastructure, management systems and business models. Those investments can absorb time and capital before the eventual gains appear.

This matters because AI can look simultaneously impressive at task level and disappointing at enterprise level.

A published field study of more than 5,000 customer-support agents found that access to an AI assistant raised productivity by about 15 per cent on average, with the largest gains among less experienced workers. That is a meaningful result. It shows how AI can diffuse part of the tacit knowledge of high performers.

But a local productivity gain does not automatically become enterprise value. Faster customer-service responses matter only if quality, resolution and customer economics improve. Faster code matters only if testing, security and deployment can absorb it. Faster analysis matters only if the organisation can make a better decision while the information is still useful.

The missing layer is often not another model. It is the organisational machinery that converts an improved task into an improved outcome.

AI capability has a jagged frontier

The ambition gap also exists because AI is neither uniformly capable nor predictably weak.

Research led by Fabrizio Dell’Acqua and including Ethan Mollick found that consultants using AI completed more tasks, worked faster and produced higher-quality results when the work sat inside the technology’s capability frontier. On a task deliberately designed outside that frontier, however, AI users were less likely to reach the correct answer than the control group.

The lesson is not that AI cannot be trusted. It is that confidence and competence do not always move together.

That makes blanket deployment a poor substitute for work design. Organisations need to know where AI performs reliably, where human judgement must remain decisive, what evidence should trigger escalation and who remains accountable when human and machine recommendations diverge.

The relevant question is not “Where can we use AI?” It is “Under which conditions does AI improve this decision — and how will we know when those conditions are absent?”

This is why prompt training alone is insufficient. The real capability is learning where delegation creates value, where collaboration works better and where the machine should remain outside the decision.

The value is created by the conversion system

Most organisations already have an AI layer. What they often lack is a conversion system between the technology and the business outcome.

Marco Iansiti’s work on AI operating architecture helps explain what this system requires. When data, algorithms and networks become embedded in how a company creates value, the change is architectural. Workflows, feedback loops and decision-making begin to operate differently. AI is no longer an application used by individual employees; it becomes part of how the firm learns and acts.

That transition demands complementary choices:

Without these choices, AI remains attached to the organisation rather than integrated into it.

Many organisations measure activity because value is harder

Licence counts, active users, prompts and pilots are useful implementation metrics. They are weak measures of strategic progress.

An organisation can increase all four while leaving customer value, cycle time, decision quality and operating cost largely unchanged. It can also generate real value in one workflow without creating an enterprise-wide productivity effect. Both realities can be true at the same time.

The measurement question must therefore begin with the business system, not the technology.

Did the redesigned process reduce time to resolution? Did it improve forecast accuracy enough to change an inventory decision? Did it allow expert capacity to move towards higher-value work? Did it increase revenue, reduce error or make a previously uneconomic service viable?

This does not mean every experiment needs an immediate P&L return. Exploration has option value. But leaders should distinguish explicitly between experiments designed to produce knowledge and deployments expected to produce economic value. When those categories blur, permanent pilot portfolios emerge.

Innovation and scale need different organisational conditions

AI strategy is often described as a race to scale. In practice, scale is useful only after the organisation has learned where the technology creates value.

Carl Benedikt Frey’s work places this tension in a wider history of innovation. Decentralised systems are often better at experimentation and discovery; coordinated systems can be better at mobilising resources and scaling what is already understood. Progress slows when either logic becomes dominant.

The same tension appears inside the enterprise. A tightly controlled rollout can standardise a tool before the organisation has discovered its most valuable uses. Uncoordinated experimentation can produce hundreds of pilots that never become shared infrastructure, redesigned workflows or measurable outcomes.

Leadership therefore needs to connect exploration with a disciplined route to scale:

The ambition gap closes when AI stops being measured as a technology rollout and starts being governed as a portfolio of organisational bets — with room to discover, evidence to decide and an operating system capable of scaling what works.

Four lenses on the AI ambition gap

A useful AI keynote should do more than demonstrate what the technology can produce. It should help the audience understand why value is difficult — and what has to change next.

The productivity J-curve

Erik Brynjolfsson

Why AI’s economic payoff depends on complementary investments in workflows, skills and organisational capital. ↗

The capability frontier

Ethan Mollick

How people learn to work with a technology whose strengths are powerful, uneven and changing rapidly. ↗

AI operating architecture

Marco Iansiti

How data, algorithms, workflows and feedback systems reshape the architecture and economics of the firm. ↗

Innovation and institutions

Carl Benedikt Frey

Why technological capability becomes progress only when institutions support both discovery and scale. ↗

These are not interchangeable AI conversations. For adjacent perspectives, explore Scurati’s Deep Dives on Digital Transformation and the Future of Work.