Most enterprises have deployed AI. The ones building durable advantage are doing something the others are not: turning operational deployment into a compounding learning system. The window to convert access into advantage is narrower than it appears.

The data on enterprise AI ROI can no longer be ignored. MIT's Project NANDA tracked 100 AI initiatives across major corporations and found 95 percent generated zero measurable P&L impact. RAND's separate analysis puts the broad AI programme failure rate at 80.3 percent. These numbers are not just measuring technology failure, they are measuring a strategic error: the widespread confusion of AI access with AI advantage.

Most enterprises have arrived at the AI-gate. But very few have built an AI advantage. A helpful framework to draw a clear distinction between those two positions is the 10:20:70 rule, establishes that model selection accounts for roughly 10 percent of an AI initiative's outcome, data and infrastructure for 20 percent, and organisational redesign for the remaining 70 percent. The path to building a clear capability change begins with harnessing of tools, redefining of process and refactoring of teams at scale, not a procurement process. As most large organisations access the same foundation models at broadly similar costs, what separates the narrow cohort generating meaningful enterprise-wide returns is how deeply the technology is embedded systematically, running at operational scale, connected to proprietary data and workflows, generating learning with each iteration across their operating model.  

While the prioritisation of specific implementations across verticals vary by industry, the principle is consistent across prior major technology replatforming. The rest of the pack are merely managing a portfolio of pilots, and making the mistake measuring their success by the wrong standards.

This is the Access Trap: the strategic misalignment of treating AI access as equivalent to AI advantage. From our observation, modern enterprise AI programmes are almost universally measured by coverage and number of projects including tools deployed, use cases identified, pilots approved. These merely metrics track what an organisation can currently do with the tools. Advantage accrues to organisations that turn deployment into a learning system, and those two activities look alike from the outside while producing entirely different results.

Lessons from the industry

The semiconductor industry spent six decades establishing this structural distinction that is widely apparent today. When Taiwan Semiconductor Manufacturing Company (TSMC) began expanding its foundry capacity in the 1990s, it was one of several capable silicon chip manufacturers. Today it controls roughly 70 precent of the global foundry market and produces roughly 90 percent of the world’s most advanced chips, and its position reflects not exclusive technology but what large-scale production and its learning loop generates. In the first quarter of 2026, TSMC reported record revenue of $35.7 billion, and some estimates show that its technology lead over nearest competitors at five to ten years.

TSMC did not acquire this position through exclusive technology. It built it through what every chip it produced generated: yield data that improves the manufacturing process; which lowered cost per chip; lower cost expands production volume; higher volume generates more data. As Chris Miller documents in Chip War, scale and manufacturing sophistication became inseparable over five decades of iteration — TSMC is simultaneously the world’s largest and most technically advanced foundry for the same reason. The mechanism now operating in enterprise AI follows the same logic: deployment at operational scale generates proprietary learning that no form of investment in access alone can replicate.

The same mechanism is also observed in JPMorgan Chase which began building AI infrastructure in 2017 — COIN, its contract intelligence system, processed 360,000 hours of legal work in its first year, and LOXM, its AI trading system, launched the same year to optimise equity order execution. That foundation is what LLM Suite is running on today. By late 2025, the bank had extended LLM Suite to more than 250,000 employees and was moving into agentic workflows across 450 depployed use cases.  

Derek Waldron, the bank’s chief analytics officer, distinguishes the two investment types directly: obtaining a model via an API is one decision; deploying it across 250,000 employees with connectivity to proprietary internal data is a fundamentally different one. JPMorgan began building this infrastructure in 2017. Eight years of operational deployment at that scale accumulates learning that a competitor switching on the latest API today cannot and will not acquire.

Organisations that remain in the pilot phase are falling behind a compounding curve, and the gap widens faster than it appears from the inside. Case in point: Huawei's trajectory illustrates the ceiling a late entrant encounters. Despite billions invested in domestic chip development, Huawei's Ascend 910C delivers roughly 60 percent of an NVIDIA H100's performance, and projections from January 2026 suggest the gap will be seventeen times wider by 2027 as the NVIDIA frontier keeps advancing.  

The ceiling that laggards will encounter is a not a technology but an institutional capability ceiling: the proprietary data, refined workflows, and institutional knowledge that accumulates only through sustained operational deployment, and for which no accelerated pilot programme or tool purchase can be substituted.

Commoditisation will not close this gap

While some believe that the commoditisation of AI, where open-source model improvements and reduced inference costs, will make the echnology widely accessible and bridge the performance gap between proprietary and open-source solutions, this is unlikely to translate in enterprise context because the true advantage lies in how deeply and systematically AI is integrated and continuously improved within organisations. With greater access, more organisations can leverage advanced AI capabilities without prohibitive costs, potentially accelerating innovation across industries. However, the gap in change management and shifting culture will continue to persist, as those with proprietary integration and learning systems continue to outpace others in sustained performance and competitive advantage.  

But having access and advantage are different things, and the history of cloud computing establishes the distinction precisely. Amazon Web Services made cloud infrastructure a commodity from 2006 onward. Amazon had spent nearly a decade running at cloud scale before any competitor could purchase any equivalent infrastructure. The operational capability built in that decade across supply chain intelligence, recommendation systems, logistics optimisation compounded through the foundational infrastructure rather than commoditising with it. Late entrants gained parity on tooling. And until today, few have closed the capability gap.

Building the advantage

The executives closing this distance have made one fundamental shift in how they govern AI: they treat deployment as a continuous process of learning capture rather than a project with a completion date. Progress is measured by whether each system generates signals that feeds back into improved performance, not by how many tools are licensed or how many pilots are running. Every AI system in production is either improving through use or falling behind a frontier that keeps advancing. The advantage it holds compounds with every transaction processed, every query refined, every workflow rebuilt around the model.

This means three things in practice. First, change the measurement of success: replace pilot count and tool coverage with specific measures of business outcomes (e.g., productivity improvements -> reduced need for headcount -> output per employee) for functional teams who are leveraging AI. Second, change the definition of done: deployment is not a completion event; each AI system in production needs an owner accountable for its learning trajectory, not just its keeping uptime. Third, change the budget model: treat AI deployment not as a capital project with a final cost but as an operating commitment that deepens over time. Organisations that aim to build durable advantage recognise that such as transformation requires continuous effort and investment, and it was never about the access.  

Enjoyed the read? Let’s take it further.

Connect to unlock exclusive insights, smart AI tools, and real connections that spark action.

Schedule a chat to unlock the full experience