Beyond the AI Hype: what has really changed, and what leaders should do now

For fifty years, artificial intelligence has moved through the same cycle: promise, excitement, disappointment, renewed investment. Twice that cycle ended in a funding winter. So when the excitement returns, the sensible question for any leader is not "are the demos impressive?" but "what, specifically, is different this time?"

Our study, Beyond the AI Hype, works through that question from the evidence. Its proposition is that today's acceleration is not the product of one breakthrough. It comes from five forces arriving together: data, compute, algorithms and architectures, infrastructure, and capital.

Why the last two booms ended in winter

The first AI winter ran from 1974 to 1980, after machine translation and general problem solving were oversold against the hardware and data of the time. The second, roughly 1987 to 1993, followed the collapse of the expert-systems industry, when the specialised hardware it ran on was undercut by ordinary workstations and large hand-built rule bases became too costly to maintain.

Both winters had the same shape. What failed was the match between the claims and the enabling conditions: the data, compute and infrastructure the ideas needed did not yet exist. That is the test the study applies to AI today.

The breakthrough was not the algorithm

In 2012 a deep network called AlexNet cut the error rate on the ImageNet benchmark from 26.2% to 15.3% in a single year. It is often told as the story of a new algorithm. It was not. Convolutional networks had existed since the late 1980s, and backpropagation, the method used to train them, dates from 1986.

What changed was four conditions arriving at once: millions of labelled images, consumer graphics processors that suited the arithmetic, better training practice, and storage cheap enough to keep internet-scale data. The study's first major finding follows from this: ideas can stay dormant until the enabling conditions catch up.

For your organisation, the lesson is direct. A capability that works in a vendor demo is an idea. A capability that works on your data, in your workflow, under your governance, is a deployment. The distance between the two is where most of the work is.

Progress is recombination, not a staircase

The popular picture of AI is a ladder: rules, then machine learning, then deep learning, then generative AI, then agents. The evidence shows something else. In 2016 DeepMind's AlphaGo beat Lee Sedol 4 to 1 at Go by combining three older traditions: neural networks, reinforcement learning and tree search, run on compute none of them had when they were invented. The study calls the accurate picture "a braided river, not a staircase". Progress usually comes from recombining ideas under new conditions, not from one technology replacing another.

What scale buys, and what it does not

Since 2010, the training compute behind notable models has grown roughly 4.1 times a year. Scale does improve benchmark performance and the breadth of what models can do. What it does not reliably improve is hallucination, calibration and consistency. The study calls this the most commercially important finding in that section: loss curves fall smoothly, but dependability does not follow them.

Adoption is not the same as value

The headline numbers look contradictory. Stanford's AI Index reports 78% of organisations using AI in at least one function in 2024, up from 55% a year earlier. MIT's NANDA initiative reported in 2025 that roughly 95% of about 300 enterprise AI initiatives showed no measurable profit-and-loss impact. Both can be true: one measures use anywhere in the organisation, the other measures attributable financial return on pilots over short horizons.

Task-level results are mixed too. In one controlled study, developers finished a defined coding task roughly twice as fast with GitHub Copilot. In a 2025 randomised trial by METR, experienced developers were about 19% slower with AI tools on their own mature codebases, while expecting to be 24% faster and believing afterwards they had been 20% faster. Feeling faster is not evidence of being faster.

The reason capability runs ahead of value is that the constraint has moved. Workflow design, process maturity, data quality, integration, governance, employee adoption, reliability on the specific task and measurement discipline now decide the outcome, and every one of them sits inside the buyer's organisation, not in the vendor's model.

So is this time different?

Partly. Hyperscaler capital spending passed $350 billion in 2025, internet-scale data and compute now exist, and current systems reached hundreds of millions of users within months. But reliability is unsolved, models still lack persistent memory and continual learning, and the old pattern of claims running ahead of conditions is visible in agent marketing today. The study's verdict: this is neither another hype cycle nor a completed revolution. It is a structural shift in its awkward middle.

What this means for 2026 and 2027

The study answers the questions leaders actually face:

• Scale what is measured, experiment where it is not. Nothing should move from pilot to scale without a financial or operational metric agreed before the pilot starts.
• Buy the model, build the system. The lasting asset is the redesigned workflow and the data feeding it, not the model.
• Redesign workflows, do not just automate tasks. Capability dropped into an unchanged process inherits its defects at machine speed.
• Copilots for breadth, agents for narrow work. Use agents on well-instrumented, internal processes with human review before irreversible actions, and widen their autonomy only as reliability is proven.
• Set up governance now, sized to risk, and measure outcomes, not adoption.

The order of decisions the study recommends is simple: what can AI now do, where can it be embedded, where can it operate safely, and where does it produce measurable value. Only then ask where to invest at scale.

Read the full study. Beyond the AI Hype: AI Strategy 2027/2028 covers the full history, the scaling evidence, the frontier models analysis and an evidence table that grades each major claim. Read the study