Access to frontier intelligence has stopped separating one company from another, and most executives I speak with have just accepted it. Anyone with a corporate card and an API key can reach capabilities that sat behind enormous research budgets two years ago. Scarcity has moved to the ability to make those capabilities count inside a real business, with real customers and real constraints.
Competition has settled into a different register over the past eighteen months. Firms are being pulled apart by how deeply they have rewired the way work actually moves through the organisation, from the first customer signal to the final decision. Converting raw capability into a repeatable operating habit is considerably harder than procuring it, and considerably harder for a rival to copy.
A frontier without a leader
Benchmark leadership has turned into a fragile asset, held for a quarter and surrendered by the next release cycle. The score difference between the top-ranked model and the tenth-ranked model fell from 11.9% to 5.4% in a single year, and only 0.7% separated the top two. By March 2026, the strongest American model led the nearest Chinese model by 2.7%.
Pricing has travelled the same path, and the descent has been steep enough to redraw most procurement assumptions. Inference costs for performance equivalent to GPT-3.5 dropped more than 280-fold between late 2022 and late 2024, and open-weight systems narrowed much of the remaining distance to closed ones.
Capability that keeps getting cheaper and more plentiful starts to behave like electricity, powering a great deal and explaining very little about who ends up ahead.
Adoption turned out to be the easy part
Almost nine in ten organisations now use AI in at least one business function, and roughly two-thirds have yet to scale anything across the enterprise. Around 6% qualify as genuine high performers, meaning they attribute 5% or more of EBIT to their AI work. The other 94% are spending without a number to show the board.
Some 95% of generative AI pilots produced no measurable effect on profit and loss, with the cause traced to integration and learning gaps instead of model quality. Brittle processes, missing context and workflows that forget everything between sessions do far more damage than any benchmark deficit.
Investment intent has not slowed for a moment, even among boards that struggled to show returns last year. Of the companies surveyed, 92% plan to raise AI spending over the next three years. Only 1% of leaders call their own organisation mature, meaning AI is embedded in workflows and driving business outcomes.
Context, data and hard-won memory
A frontier model arrives knowing almost everything about the world and nothing whatsoever about your business. It has never seen your pricing logic, your escalation thresholds, your churn signals or the tacit judgement your best operations lead applies on a Friday afternoon. Encoding this material into systems an agent can act on is slow, unglamorous and extremely difficult to lift from a competitor.
Proprietary workflows also compound in a way that licences never do. Each cycle produces a record of what the system got wrong and how it was corrected. The next cycle uses that record. A competitor starting now has no such record, and there is nowhere to buy one. In McKinsey’s survey of 2,000 respondents, redesigning workflows end to end correlated with realised value.
Bolting AI onto old processes
Dropping an assistant into a process built around human throughput preserves every assumption baked into the old design. Queues stay queues and handoffs stay handoffs, while approval chains keep absorbing the time the technology just gave back. High performers are nearly three times as likely as other organisations to have fundamentally redesigned individual workflows as part of their AI programmes.
Serious redesign begins with structural moves that most transformation programmes avoid until they run out of easier options. Decision rights move closer to the point of work, handoffs collapse, and information travels without a committee waiting at every junction. Teams then divide labour deliberately, sending pattern-heavy execution to machines and reserving human attention for judgement, exceptions and accountability.
Software delivery offers the clearest early evidence of what that rebuild produces once it takes hold. Organisations that rebuilt their operating models around continuous human-agent collaboration are reporting three- to fivefold productivity improvements alongside a 60% reduction in team size. These numbers came from rebuilding the delivery cycle itself, with tooling as a consequence of the redesign.
AI leaders will redesign the company
Buying capability is a procurement exercise, and plenty of organisations have already completed it with some skill. Redesigning an operating model calls for something closer to industrial engineering, applied to knowledge work and led from the top. Management layers disappear, job descriptions get rewritten, and one person now carries leverage that used to need a full team and a quarterly planning cycle. Governance has not kept that pace. McKinsey put average responsible-AI maturity at 2.3 out of 5 across some 500 organisations, up from 2.0 a year earlier. Only a third reached a mature level on strategy, governance and agentic controls.
Trust as part of the workflow
Autonomy changes the shape of the risk once agents begin acting on live systems. Concern moves beyond a system producing the wrong sentence towards a system taking the wrong action, misusing a tool, or operating outside the guardrails someone assumed were holding.
Nearly two-thirds of McKinsey’s respondents named security and risk as the leading barrier to scaling agents, ahead of regulatory uncertainty and technical limits.
A workflow that answers this logs which person or system performed each step. Human sign-off comes before the steps that are expensive to reverse. Leading organisations are markedly more likely to have defined those checkpoints in advance instead of discovering the need after an incident. Provenance and accountability then earn their keep as the reason a board signs off on wider deployment.
The moat surrounding the model
Commoditisation has further to run, and the next few model generations will arrive on similar terms for everyone. Durable advantage will accrue to the organisations that have accumulated proprietary workflows, hard-won context and institutional judgement encoded into daily operations. Those assets can take years to assemble and cannot be procured in a single quarter.
Every competitor will soon draw on comparable intelligence at comparable cost, which makes the surrounding architecture the whole contest. Advantage belongs to whoever has built the better company around the same available capability.
