AI doesn't fail. Organizations do.

Executive summary

• Across industries, most organizations still struggle to translate AI into measurable impact at scale.
• The bottleneck rarely comes from the models themselves. It emerges in the "last mile": the difficulty of embedding AI into the unique operational context of the enterprise: workflows, incentives, governance, decision processes, and the way work actually gets done.
• This creates a growing "AI value gap." A small minority of organizations already capture disproportionate returns from AI, while most remain stuck in pilots, isolated copilots, or productivity gains that fail to compound into enterprise performance.
• As foundation models commoditize, competitive advantage shifts back toward what already makes companies unique: strategic vision, differentiation, operational coherence, and execution. AI can dramatically accelerate these strengths. It can also amplify fragmentation, inconsistency, and strategic sameness.
• This is why the center of gravity in the AI market is progressively moving toward orchestration, integration, and operationalization: the layer where generic intelligence is transformed into durable business outcomes at enterprise scale.

‍

The current wave of AI does not fundamentally change the structural weaknesses of large organizations so much as expose and amplify them. More powerful models and faster insights increase the value of good decisions, but they also magnify the consequences of organizational incoherence. The difference is that AI accelerates both value creation and value destruction simultaneously.

What used to remain marginal or manageable – siloed workflows, fragmented governance, inconsistent decision-making, shadow IT, weak data foundations, contradictory incentives – becomes significantly riskier once scaled through AI systems and agents.

In that sense, many of the challenges companies now face with AI echo the same patterns observed during previous waves of digital, data, and analytics transformation: fragmented adoption, proliferation of pilots, and persistent difficulty translating technological capability into measurable business outcomes.

‍

AI does not automatically democratize performance

For the past three years, AI has been discussed largely through the lens of capability. Every few months, a new model redefines expectations around reasoning, coding, image generation, or autonomous agents. Benchmarks improve continuously, costs fall rapidly, and what seemed impossible twelve months ago quickly becomes standard.

At the same time, a quieter conversation has emerged inside large organizations. Beyond the excitement surrounding agents, copilots, and enterprise deployments, many executives still struggle to identify where sustainable business value is actually being created. In more than three years since ChatGPT's release, only 27% of executives say AI has met their ROI expectations.

Recent PwC research illustrates the scale of the imbalance. Among more than 1,200 large companies surveyed, the top 20% capture roughly 74% of AI-driven returns, while the majority continue to generate limited or uneven impact. The most "AI-fit" organizations produce AI-driven financial performance that is 7.2 times higher than the rest of the market.

Far from democratizing performance, AI may ultimately widen the gap between organizations capable of operationalizing intelligence at scale and those that are not.

As described in our first note, this pattern should feel familiar. Over the last two decades, digital transformation followed a remarkably similar trajectory. Most companies invested heavily in new technologies, but relatively few managed to convert those investments into durable competitive advantage. The history of enterprise technology is filled with sophisticated systems that never meaningfully changed how organizations operated: CRM systems that improved reporting more than customer relationships, analytics programs that multiplied dashboards without improving decisions, or ERP transformations that standardized processes without improving agility or performance readability. AI risks reproducing the same dynamic, although at a much larger scale.

This pattern echoes earlier general-purpose technologies such as electrification. Early factories initially used electricity to improve lighting or replace steam engines while preserving existing organizational structures. As Stanford researcher Paul A. David showed in a 1990 paper, productivity gains only accelerated once factories redesigned workflows, layouts, and operating models around the technology itself.

‍

Productivity is not the destination

Part of the problem comes from the way AI is framed in public discourse. The conversation remains dominated by models, agents, and productivity gains. The underlying assumption is that more intelligence will naturally produce more value. In practice, the relationship is far less linear.

Recent research on human-AI collaboration provides a useful illustration. In a large-scale experiment conducted by researchers from MIT and Johns Hopkins, human-AI teams produced roughly 50% more outputs per worker than human-only teams. Yet the same study also highlighted important limitations: quality gains were uneven across tasks, outputs became significantly more homogeneous, and AI systems consistently struggled in areas requiring contextual or multimodal judgment. The researchers described this phenomenon as a "jagged frontier" of AI capability, where performance varies dramatically depending on the nature of the work being performed.

Companies compete through complex combinations of workflows, decisions, customer relationships, and operating models whose interactions ultimately determine performance. Productivity gains in one area can easily create inefficiencies elsewhere if they are not integrated coherently into the broader organization.

More fundamentally, even in a scenario where AI dramatically improves productivity, an important question remains unresolved: what happens once productivity itself plateaus? Enterprises have already spent two decades stretching operating models and reducing costs. The next frontier of value creation is unlikely to come from endlessly compressing labor costs further, but from reinventing products, customer experiences, decision systems, and sources of growth.

This is why, based on our 20-year experience, the companies generating the strongest returns from AI are not primarily using it to reduce costs. They tend to share three characteristics: a clear focus on a small number of critical business problems, the ability to industrialize successful use cases across workflows and geographies, and a strong emphasis on adoption, governance, and operational execution.

Leading organizations are significantly more likely to use AI to reinvent business models, redesign value chains, and identify emerging value pools rather than simply optimizing existing operations. Jensen Huang recently made a similar point when warning executives against reducing AI to short-term ROI calculations and narrow productivity metrics. Companies focused exclusively on immediate returns risk missing the broader strategic transformation AI may unlock over time.

‍

"The last mile is the entire problem"

Models don't create value, systems do. This distinction helps explain why so many AI initiatives stall after the pilot phase. Most organizations still approach AI as a technological layer added onto existing structures rather than as a redesign of how workflows, decisions, and operating systems actually function. Models generate insights. Copilots accelerate tasks. Dashboards multiply recommendations. But the surrounding organization – the enterprise context – often remains largely unchanged.

This is where the notion of the "last mile" becomes useful, although perhaps not in the way it is usually understood. The expression often suggests that the difficult part comes at the end of the process, once the technology itself is already functioning. In reality, the technical and organizational dimensions are intertwined from the beginning. George Sivulka recently captured this well in an essay on vertical software and AI for a16z. "The last mile is the entire problem," he wrote, arguing that enterprise value does not come from generic intelligence itself but from understanding a specific organization well enough to encode how work is actually performed inside it.

This operational context is what most AI systems still struggle to capture reliably at scale. As a result, many organizations generate increasingly sophisticated outputs without fully connecting them to the realities of how the business actually functions. Insights become disconnected from execution. Recommendations conflict with incentives or brand positioning. Automation layers multiply while operational coherence weakens.

The difficulty is that the enterprise context is unique by definition. Workflows differ across industries, business units, geographies, and even individual teams. Insurance companies do not operate like consumer brands. Retail organizations plan around seasonal demand and promotional calendars. Supply chains react to weather patterns, commodity prices, or geopolitical events. Product launches, business reviews, pricing cycles, regulatory constraints, and incentive structures continuously reshape how decisions are made and executed inside the organization.

The challenge therefore is less about deploying more intelligence than about embedding AI into the contextual fabric of the enterprise deeply enough for systems to produce outcomes that remain relevant, actionable, and scalable in the real world.

This shift is accelerated by three simultaneous dynamics: the commoditization of foundation models, the saturation of pilots and non-industrialized prototypes, and growing economic pressure for measurable ROI rather than promises. The competition intensifies in the most strategic layer of the stack: the integration and orchestration of AI inside business processes.

As AI agents become democratized across organizations, these tensions intensify even further. Enterprises now face not only fragmented architectures and inconsistent governance, but also shadow AI systems, contradictory automations, security vulnerabilities, and new forms of organizational incoherence created by decentralized AI adoption.

‍

The best use of AI is to amplify what already makes a company unique

General-purpose models are becoming increasingly accessible. Over time, access to intelligence itself will become less scarce. But this does not mean competitive advantage disappears. In many ways, the opposite may happen. As models commoditize, differentiation shifts back toward what has always made companies unique in the first place: their brand equity, customer relationships, products, operating culture, distribution advantages, and strategic vision.

Amazon was already exceptional at e-commerce and operational efficiency before AI. Apple was already differentiated through product design and ecosystem integration. Before the AI hype, Chanel built its strength on customer intimacy and desirability. L'Oréal's "Beauty Tech" transformation started long before generative AI became mainstream, driven by the conviction that beauty companies would increasingly compete through technology, data, and customer understanding. In each case, AI becomes powerful when it reinforces an existing strategic direction and accelerates an organization's ability to execute against it at scale.

This is why one-size-fits-all AI systems rarely create exceptional value. What increasingly distinguishes the leading organizations is their ability to align AI with the strategic foundations that already make them unique, then operationalize that advantage at scale.

At small scale, many of these issues remain invisible. Pilots operate in controlled environments with limited complexity and manageable constraints. But once AI systems are deployed across large organizations, small inconsistencies become amplified: fragmented taxonomies, conflicting workflows, weak governance, local process variations, or disconnected incentives suddenly scale alongside the technology itself.

As Alex Karp put it in a 2024 letter: "Models with trillions of parameters may be able to flawlessly mimic Goethe, but without more, add little value to the enterprise." The real differentiator lies in the ability to align intelligence with a company's strategic intent and distinctive way of creating value (in other words: it's the context, stupid!).

This evolution is already reshaping the AI market itself. Foundation models are progressively becoming utility infrastructure, while value shifts toward deployment, integration, orchestration, and operational execution at scale. OpenAI, Anthropic, and Mistral are all moving in this direction, investing not only in models, but also in workflow integration and embedded engineering capabilities designed to operationalize AI inside enterprise environments. These layers are not secondary implementation problems. They are the bridge between technical capability and business value.

‍

To read more about this subject, explore more articles here and here.