“It’s almost incomprehensible how wide the distance is between someone who’s working at the frontier with AI and basically everyone else right now. The crazy thing is these lines are not converging, but only growing farther apart… The curve of AI progress is basically going vertical right now, which means if you had codified your processes 3-6 months ago you’re likely already behind again.” – Aaron Levie – Box CEO
The central business problem is not access to artificial intelligence, but the shrinking half-life of organisational knowledge. A workflow that was carefully designed around a model’s capabilities can become an unnecessary constraint when a newer system handles more context, uses tools more reliably, or completes a task with less supervision. Aaron Levie’s observation places that problem in the context of his experience leading Box and describes a widening gap between frontier practitioners and organisations still treating AI as a fixed software deployment.1
That gap is produced by two clocks running at different speeds. Model development is measured in weeks and release cycles, while enterprises measure change through budgets, procurement, security reviews, training programmes and annual planning. The result is a structural delay: a company may finish standardising one process just as the assumptions behind the standardisation become obsolete. The issue is not that every new model is universally better. It is that improvements arrive often enough to challenge settled decisions about where human judgement belongs, how much automation is safe and which tasks deserve redesign.
From tools to moving operating models
Many companies began with enablement: employees received access to general-purpose assistants and learned to use them alongside existing applications. The next stage is automation, in which AI operates across a workflow rather than merely helping with an isolated task. The more consequential stage is reinvention, where roles, controls, hand-offs and performance measures are rebuilt around what the technology can now do. Research distinguishes these stages and finds that most organisations remain early in scaling, even though AI use is widespread.2,3
This explains why a short period can create such a pronounced advantage. A frontier team is not simply asking a better chatbot to draft the same document. It may be testing agents that retrieve information, write and execute code, evaluate outputs, route exceptions and maintain a record of decisions. That team accumulates operational knowledge about delegation, verification and failure modes. A slower organisation may have access to the same public model but lack the redesigned process, data preparation and managerial confidence needed to use it effectively.
The competitive advantage therefore lies less in possessing a model than in learning faster than the model changes. A static playbook assumes that capability is stable enough to document once. A dynamic operating model treats documentation as a living control system: it records the task, the required evidence, the permitted actions, the escalation path and the conditions under which the process must be reviewed. This is a significant shift in management. Standardisation remains necessary for reliability, but it must be paired with mechanisms for rapid revision.
Why recursive improvement raises the stakes
The reference to recursive self-improvement points towards a more demanding possibility: AI systems helping with the research, testing and engineering of later AI systems. RAND reports that AI already automates meaningful components of AI research and development, while also noting that there is no systematic public accounting of how much model progress is directly attributable to AI-generated research ideas.4 That distinction matters. Faster coding or experiment analysis can accelerate development without implying an autonomous system that independently chooses its own research agenda.
Even a partial feedback loop could intensify organisational pressure. If models help generate experiments, identify failures and optimise implementation, the interval between capability advances may shorten. Stanford HAI experts have cautioned that recursive improvement may be highly effective where objectives are measurable, but less reliable for open-ended problems involving judgement or human taste.5 The limitation is important because it counters the idea that every aspect of intelligence will suddenly accelerate at the same rate. Progress can be uneven, domain-specific and difficult to translate into dependable business outcomes.
The governance concern is equally practical. A system with access to sensitive code, data and evaluation infrastructure becomes part of the security boundary of the research organisation. RAND warns that an internally deployed research agent could be compromised or subtly influence the development of future systems.4 For enterprises, the lesson is broader than frontier laboratory security. Automated systems need bounded permissions, traceable actions, independent evaluation and a clear route to human intervention, especially when they can alter the workflows used to assess their own performance.
The challenge for established companies
Large organisations possess assets that AI-native firms often lack: proprietary data, trusted customer relationships, distribution, regulatory expertise and deep knowledge of complex operations. Those strengths do not automatically translate into speed. Legacy systems can make experimentation expensive, while approval structures distribute responsibility so widely that no one owns the redesign of an entire workflow. Rsearch findings suggest that high-performing organisations are more likely to assign senior leadership ownership, invest across strategy, talent, operating model and data, and redesign workflows rather than merely add AI to existing tasks.3,8
The most defensible response is not to chase every release. It is to create a capability for continuous reassessment. Teams can maintain a portfolio of high-value workflows, define measurable outcomes, test new models against representative cases and retire weak approaches without treating earlier investment as sacred. Procurement should favour interoperability and reversible commitments where possible. Training should focus on judgement, verification and process design, not only on prompt techniques. Change management must also be designed to absorb new capabilities within a reasonable time frame, rather than assuming that adoption ends when a tool is launched.10
There are serious objections to the most dramatic interpretation of the statement. Frontier performance on benchmarks does not guarantee dependable performance in messy commercial settings. A model may be more capable but harder to predict, more expensive to operate or more demanding of human review. Rapid experimentation can also amplify errors, expose confidential information or create an illusion of productivity when workers produce more output without improving decisions. Enterprise adoption data reflects this tension: reported use is high, but many organisations remain in experimentation or pilot phases and have not yet captured material company-wide impact.3,11
The practical conclusion is not that every process must be rewritten every few months. It is that no process involving significant judgement, cost or competitive advantage should be considered permanently finished. Leaders need a cadence for revisiting assumptions, a measurement system that separates genuine value from activity and controls that make experimentation safe. Frontier teams gain an advantage by treating AI as a moving capability rather than a purchased product. Established companies can narrow the distance only by changing their own learning cycle, so that organisational adaptation becomes continuous rather than exceptional.
References
1. Aaron Levie – Linkedin – https://lnkd.in/p/dwjxyEwH
