Every major technology wave produces the same organisational response. Cost is reduced. Headcount falls. Margin improves. The operating model is not rebuilt. Then the assumptions underneath the saving shift, and the organisation discovers it spent the window on the wrong thing.
There is a pattern in enterprise transformation that repeats with such consistency that it should, by now, be familiar. A new technology arrives with genuine capability. Organisations identify the horizontal efficiencies it enables. They restructure around those efficiencies, reduce the cost base, pass some of the saving to customers as a competitive signal, retain the rest as margin, and declare the transformation complete. The operating model underneath is not fundamentally rebuilt. The workforce capacity removed is not replaced with a different kind of capability. And when the cost assumptions underlying the saving eventually shift, as they always do, the organisation finds itself with a thinner structure, a compressed margin, and nothing new on the revenue side to absorb the pressure.
This is not an AI story. AI is the current chapter. But the pattern predates it by three decades, and understanding the pattern is more useful than analysing the current chapter in isolation.
In the 1990s, enterprise resource planning systems promised to consolidate fragmented processes into a single, integrated platform. The productivity case was real. Industry research, including work published by Aberdeen Group, found that organisations implementing ERP cited cost savings as the primary business goal in 46 percent of cases, with reduced process time, lower IT costs, and decreased inventory levels among the top reported returns.
The savings materialised in many cases. What did not follow was a fundamental redesign of how the organisation created value. ERP standardised what existed. It did not transform it. The operating model after implementation largely resembled the operating model before it: faster and cheaper to run, but structurally similar. The technology absorbed horizontal complexity without prompting a serious rethink of where the enterprise was competing or how it intended to grow.
The pattern was visible even at the time. Organisations that extracted the efficiency dividend and stopped there found themselves, five years later, running leaner versions of the same model, without the structural innovation that the ERP window could have funded.
The early 2000s brought a different instrument for the same logic. Offshoring, the movement of service-sector work to lower-cost labour markets, accelerated rapidly after 2002. Research published by the Center for American Progress found that during the 2000s, U.S. companies reduced their domestic workforce by approximately 2.9 million positions while employing around 2.4 million workers internationally. By 2004, Forrester Research estimated that 400,000 white-collar service jobs had moved offshore, with displacement running at twelve to fifteen thousand positions per month, a figure the Council on Foreign Relations reported at the time.
The economic logic was straightforward. The opening of China and India created access to large pools of labour at a fraction of domestic cost. For functions that were process-intensive and rules-based, including back-office operations, customer service, software testing, and data processing, the arbitrage was substantial and defensible.
What the offshoring wave also produced, and what was less examined at the time, was a hollowing out of institutional knowledge at the layer that was moved. The tacit understanding of processes, the informal networks of expertise, the accumulated judgement of people who had done the work for years: these did not transfer cleanly to a lower-cost location. They dissipated. And when labour costs in offshore markets began to rise, as they did from the mid-2010s onward, organisations discovered that the arbitrage had a shelf life that their operating models had not been designed to account for.
Cloud migration, which accelerated through the 2010s and into the 2020s, was sold on two arguments simultaneously: infrastructure cost reduction and strategic flexibility. The first argument proved partially true. The second was largely deferred.
The cost reduction case was genuine for organisations that managed the transition well. Multiple vendor and analyst studies suggest that organisations moving core workloads to the cloud typically achieve thirty to sixty percent reductions in total cost of ownership, with return on investment periods of twelve to eighteen months in well-managed migrations.
What the cloud story did not adequately account for was consumption behaviour at scale. Flexera's annual State of the Cloud report, which has surveyed more than 750 enterprise cloud decision-makers each year for over a decade, found that organisations wasted between 27 and 32 percent of their cloud spend every year from 2019 through 2025. The rate has not meaningfully declined. At Gartner's projected global cloud infrastructure spend of 675 billion dollars for 2025, that figure represents approximately 182 billion dollars in annual waste. A 2024 Gartner Peer Community survey found that 69 percent of organisations exceeded their cloud budgets in 2023. Managing cloud spend ranked as the top cloud challenge for six consecutive years in Flexera's research.
The pattern repeated precisely. The technology delivered real efficiency. Organisations did not rebuild their operating models to govern or grow through it. The cost discipline required to realise the saving over time was not embedded. And the strategic flexibility argument, the idea that cloud would unlock a fundamentally different way of building and delivering products, was captured by very few and deferred by most.
The AI efficiency wave of the past three years has followed the same logic with greater speed and more visible consequences.
Organisations identified horizontal efficiencies in customer service, data entry, document processing, software testing, and basic analysis, and restructured around them. A 2026 survey of enterprise executives, cited in Morgan Stanley research, found that AI adoption is producing a net four percent workforce reduction across industries, with the highest-displacement functions reporting reductions of fifteen to twenty percent. The institutional examples are not ambiguous. Bank of America's CEO Brian Moynihan stated in the company's fourth quarter 2025 earnings disclosures that AI saves the equivalent of roughly two thousand coding positions. Dell's fiscal 2026 annual report documented a ten percent workforce reduction, the third consecutive year of similar reductions, concurrent with a forty percent increase in AI infrastructure revenue. JPMorgan Chase's 2025 shareholder letter and Bloomberg disclosures reported two billion dollars invested in AI development against a matched two billion dollars recovered through headcount reduction and efficiency gains.
The savings were real. A portion was passed to customers. The remainder was retained as margin. And the operating model underneath was not fundamentally rebuilt.
What distinguishes the AI wave from its predecessors, and what makes the prospective bill larger, is that the saving was built on a cost structure that has not yet been fully priced.
OpenAI's estimated revenue for 2025 was approximately 3.7 billion dollars, while analyst estimates placed its losses at approximately five billion dollars, representing spending of roughly 1.35 dollars for every dollar earned. These are third-party estimates; OpenAI does not publish financial statements. But the underlying dynamic is structural, not specific to one company. Frontier model providers have priced AI inference aggressively to capture market share, and the current token prices that enterprises have embedded into their operating cost models are, in material part, subsidised by venture capital and hyperscaler cross-investment. That subsidy has a finite life.
The capital commitment data makes the direction clear. Hyperscalers committed over 600 billion dollars in AI infrastructure capital expenditure for 2026, a 36 percent increase over 2025. Alphabet's projected capital expenditure for 2026 is expected to approach 175 to 185 billion dollars, nearly double its 2025 level, with the majority directed at AI infrastructure. These are decade-scale investments that must be recovered through revenue. The structural economics of the AI infrastructure market make upward price normalisation a question of timing, not probability.
There is a further complication that the per-token pricing narrative obscures. While the unit price of a token has fallen, Ramp's enterprise spending data shows that the average cost per million tokens across major providers dropped from approximately ten dollars to two dollars and fifty cents in a single year, total enterprise AI spend has risen sharply. The FinOps Foundation's 2026 State of FinOps Report found that the average enterprise AI budget grew from 1.2 million dollars per year in 2024 to seven million dollars in 2026. The shift from standard language models to reasoning models, which generate thousands of internal reasoning tokens before producing a response, has altered task-level economics in ways that most enterprise cost models have not yet absorbed. The per-token price is lower. The tokens consumed per outcome are substantially higher. The total bill is rising.
Across all four waves, the deeper failure is not that organisations took the efficiency dividend. Taking it was rational. The failure is what they did not build while the window was open.
In each case, the efficiency window created time and margin that could have funded a more substantive redesign: of how value is created, how propositions are structured, how operating capability is built for the next competitive environment. In most cases, it funded cost reduction and returned to business as usual with a leaner structure.
The evidence in the current wave is unambiguous. Deloitte's 2026 State of AI in the Enterprise report, drawing on a survey of 3,235 senior leaders across 24 countries conducted in late 2025, found that two-thirds of organisations report productivity and efficiency gains from AI. Revenue growth, however, remains largely aspirational: 74 percent of organisations hope to grow revenue through AI in the future, compared to just 20 percent that are already doing so.
McKinsey's 2025 State of AI report, based on a survey of 1,993 executives conducted between June and July 2025, found that 80 percent of organisations set efficiency as the primary objective of their AI initiatives. Only 39 percent can link any EBIT impact to AI at the enterprise level, and for most of those, the impact is below five percent. The organisations seeing the largest returns were those that set growth and innovation as explicit objectives alongside efficiency. McKinsey found these organisations were nearly three times more likely to fundamentally redesign workflows rather than simply automating existing processes. Most organisations did not do this.
A 2025 Boston Consulting Group study, cited in enterprise AI research published by OpenAI, found that organisations treating AI as a value creation instrument achieved 1.7 times revenue growth and 3.6 times greater total shareholder return compared to peers over the same period. The gap between those organisations and the rest is not access to technology. Both groups have access to the same models and APIs. The gap is operating model design: whether the organisation was restructured to create new value or only to reduce existing cost.
When the cost assumptions shift, as they shifted in the offshoring wave when labour arbitrage eroded and as they shifted in the cloud wave when consumption behaviour overwhelmed unit cost reductions, organisations face a margin squeeze from two directions simultaneously. The cost of running the technology at the scale now embedded in operations increases. The workforce capacity removed to fund the saving is not easily recovered. And the tacit knowledge, institutional judgement, and domain expertise that resided in those roles does not return.
This is the operating model failure that each wave produces in its second act. The organisation that shed roles to fund the efficiency gain now needs to respond to a changed cost environment with innovation capacity it no longer has. Gartner Vice President Helen Poitevin, commenting on AI's workforce impact, made the industrial logic plain: organisations chasing value only through headcount reduction are likely to find themselves on a path of limited returns.
The governance dimension compounds the risk. In most organisations, boards approved the efficiency programmes without examining the cost assumptions beneath them in sufficient depth. In the cloud wave, 69 percent of organisations exceeded their cloud budgets in 2023 despite cost optimisation being the stated top priority for six consecutive years in Flexera's research. The saving was presented as durable. The consumption dynamics that eroded it were not examined with equivalent rigour. The AI wave is replicating this pattern at greater scale and higher speed.
The question that most boards have not yet asked, and that most management teams have not yet answered, is straightforward: what did we build with the efficiency window that will still be creating value when the cost assumptions normalise?
For most organisations, the honest answer is: not enough.
The organisations that navigate the next phase well will not be distinguished by their access to better AI. They will be distinguished by whether their operating model was designed to create value through it, not merely extract cost from it. That is an operating model question. It requires an operating model answer.
The window to build that answer before the bill arrives is narrowing. It has not yet closed.
The efficiency dividend is spent. The operating model underneath it has not been rebuilt for what comes next. And the bill for that omission has not yet arrived.
It will.