$2.59T
Worldwide AI spending, 2026
Gartner, May 19 2026 · up 47% YoY
Gartner, May 19 2026 · up 47% YoY
By May 2026 the world had committed two and a half trillion dollars to artificial intelligence. Almost none of it has changed a single income statement.
Enterprise AI in 2026 is the largest, fastest, and least productive technology build-out in recorded history. This report explains the gap, and what closes it.
Two facts define the year, and they refuse to be reconciled by the usual optimism. The first is that money is no longer the obstacle. Gartner put worldwide AI spending at $2.59 trillion for 2026, a 47 percent jump in a single year and close to a trillion dollars more than 2025. The four largest hyperscalers alone raised combined capital expenditure toward $725 billion. The infrastructure is real, the chips are sold out, the data centres are rising from Virginia to Chennai.
The second fact is that almost none of this capital has produced a number a chief financial officer will sign. MIT's NANDA initiative studied 300 public deployments and found that 95 percent of enterprise generative AI pilots delivered no measurable impact on profit and loss. McKinsey's global survey reached the same place from a different road: only 39 percent of organisations report any enterprise-level EBIT impact at all, even as 88 percent use AI somewhere. A manufacturing chief operating officer put it to MIT's researchers more plainly than any analyst could: the hype says everything has changed, and in our operations nothing fundamental has shifted.
This is the defining paradox of enterprise AI in 2026. It is not a story about whether the technology works. The technology works. It is a story about the distance between buying a model and capturing its value, and about how few institutions have built the thing that closes that distance.
This report reads the year through two lenses the Global AI Forum uses to make sense of the noise. The first is CORE, a destination model. It names the four places AI creates enterprise value: Command, the mandate and the data strategy; Operations, the bottom line; Revenue, the top line; and Experience, what customers and employees feel. CORE answers the question every board is actually asking, which is not whether to use AI but where it pays.
The second lens is BEACON, a readiness instrument. It scores the six dimensions that decide whether an institution can reach the CORE destination at all: Business value, Engineering, AI capability, Compliance, Operating model, and Numbers, the data moat underneath everything. CORE is the map. BEACON is the odometer and the engine. You point at one and you measure on the other.
What follows maps the destination and the readiness across nine industries and six regions, with the named institutions that crossed the divide and the hard numbers they disclosed. The pattern is consistent enough to state in advance: the 5 percent who win are not the ones who spent the most. They are the ones who were ready.
Eight findings recur across every industry and region that follows, and they are stated here in advance because the detail only confirms them. First, the spend is real and the return is not yet, with $2.59 trillion deployed and the income statement barely moved. Second, almost all the value captured so far sits in the Operations layer, in efficiency and automation, while the Revenue and Command layers that build durable advantage remain nearly empty. Third, the model is commoditising, proven by JPMorgan swapping foundation models every eight weeks, so the moat has moved from the model to the proprietary data and the readiness to use it. Fourth, buying beats building by a factor of two, because integration, not capability, is the binding constraint. Fifth, seventy percent of the work is people and process, which is the part the failures skipped. Sixth, data insufficiency is the single largest cause of failure, carried at the highest weight in any honest readiness instrument. Seventh, the agentic wave will widen the divide rather than close it, punishing unready institutions faster and more expensively. And eighth, the capital has already noticed: the frontier labs are paying billions to enter the enterprise through private equity, because they have concluded that the bottleneck was never the model but everything around it.
More money has moved into AI than into any technology in history. Less of it has reached the income statement than almost any technology in memory. Both things are true at once.
On the nineteenth of May 2026, Gartner raised its forecast for worldwide AI spending to $2.59 trillion, a 47 percent increase over the prior year and an upward revision from the $2.52 trillion it had projected in January. To put the figure in proportion, it represents growth of nearly a trillion dollars in twelve months, inside a total information-technology budget that itself is rising 13.5 percent. There has never been a faster-growing line item in the history of corporate technology.
The composition of that spend matters more than the headline. More than 45 percent of it, over $1.16 trillion, is vendor-driven infrastructure: AI-optimised servers, network fabric, processing semiconductors, the physical plant of intelligence. Data-centre spending alone will grow 55.8 percent and pass $788 billion. This is not enterprises capturing value. This is enterprises, and the hyperscalers selling to them, building the capacity to attempt it. Gartner's own distinguished analyst John-David Lovelock said the quiet part directly: up to this point AI spending has been driven by technology companies and hyperscalers, enterprises have yet to flex their spending, and 2026 is the inflection year. He also placed AI, in the same forecast, in the Trough of Disillusionment for all of 2026.
In July 2025, MIT's Project NANDA published The GenAI Divide: State of AI in Business 2025. The methodology was not a vibe. It drew on 150 executive interviews, a survey of roughly 350 employees, and structured analysis of 300 public AI deployments. The finding became the most-cited statistic in enterprise technology within a month: about 95 percent of enterprise generative AI pilots delivered no measurable return on profit and loss. Not a low return. Zero. Only the remaining 5 percent, the integrated systems, created significant value.
The report's deeper findings are more useful than its headline, because they describe the mechanism of failure rather than its scale. Three of them recur in every sector this report examines. First, spending is misallocated: more than half of generative AI budgets go to sales and marketing, while the highest returns sit in back-office automation, finance, and operations, the functions nobody puts in a press release. Second, the build-versus-buy result is brutal and counterintuitive: solutions purchased from specialised vendors succeed roughly 67 percent of the time, while internally built tools succeed at half that rate. Third, the failure is almost never the model. Gartner attributes the majority of AI project failure to the absence of AI-ready data, and predicts that 60 percent of AI projects unsupported by adequate data will be abandoned through 2026, a rate already running at 42 percent of United States companies.
The hype says everything has changed. In our operations, nothing fundamental has shifted.
To understand the scrutiny, look at what the $2.59 trillion is actually buying. More than 45 percent of it, well over a trillion dollars, is infrastructure, with data-centre spending alone passing $788 billion in 2026, a 55.8 percent jump in a single year, and the four largest hyperscalers driving toward roughly $725 billion in combined capital expenditure. This is the largest infrastructure build-out in the history of corporate technology, and it is concentrated in a handful of companies buying compute from an even smaller handful of suppliers. The circularity of the arrangement, where the same firms are simultaneously the largest investors, the largest customers, and the largest beneficiaries, has drawn comparison to previous capital cycles that ran ahead of demonstrated demand. None of this means the technology is a mirage. It means the capital has been deployed into the supply of intelligence at a scale that has not yet been matched by the enterprise demand to convert that intelligence into value. The supply side raced. The demand side, the institutions that must actually integrate and govern and capture, did not race with it, and the distance between the two is the activation gap priced at trillion-dollar scale.
Gartner's own framing places 2026 squarely in the Trough of Disillusionment, the stage where inflated expectations meet operational reality and the survivors separate from the casualties. The trough is not the end of the cycle. It is the filter. It is where the institutions that built readiness keep compounding while the ones that bought hype quietly write off their pilots, and it is precisely the moment a clear-eyed reading of where value actually accrues becomes worth more than another round of experimentation.
The reckoning is no longer theoretical, because the bill is coming due on a predictable schedule. The pilots launched in enthusiasm during 2023 and 2024 are now reaching their first serious budget-renewal cycles, and the threshold for continuation has risen. Forrester found that enterprises are postponing 25 percent of planned AI spend to 2027 as financial scrutiny tightens. Fewer than one in three corporate decision-makers, in Gartner's survey, could identify a specific financial outcome attributable to their AI investment. Uber's chief operating officer told analysts in May 2026 that AI costs had been harder to justify than the company first anticipated. An Axios investigation the same month described a single large enterprise client that spent $500 million in one month on AI services, the kind of ungoverned consumption that turns a board's curiosity into a board's alarm.
This is what the Trough of Disillusionment looks like from inside an operating company. It is not that AI failed. It is that the gap between deploying a model and generating value from a model was treated, for two years, as if it did not exist. It exists. It is the most important thing in enterprise AI, and it has a shape.
Adoption is not the problem. Activation is. The distinction is the single most important idea in this report, and the data draws it cleanly. Across McKinsey's global survey, 88 percent of organisations now use AI in at least one business function. But nearly two-thirds have not begun to scale it across the enterprise. Only about 28 percent have deployed AI in production at scale across multiple functions with measurable impact. And only 39 percent report enterprise-level EBIT impact. The funnel from access to activation to outcome narrows at every step, and it narrows most sharply at the end, the only place that shows up in earnings.
The Deloitte read on the same population is consistent: driving measurable enterprise value was the top priority for three-quarters of technology leaders, precisely because so few could yet demonstrate it. The behaviour underneath the numbers is a quiet, widespread, unmeasured productivity story, what MIT called the shadow AI economy, employees using consumer tools without sanction, real gains that never reach a ledger because no system was built to capture them.
Two mechanisms quietly drain the value the headline numbers promise. The first is the verification tax. When a generative system produces output that cannot be trusted without checking, the checking consumes the time the system was meant to save. A model that drafts a document in seconds but requires twenty minutes of human review to catch its errors has not saved twenty minutes, it has moved the work and added a step. Across enterprises running models on imperfect data, this tax silently erases the promised efficiency, which is why so many deployments feel productive in a demo and disappear in the aggregate. The second mechanism is the shadow AI economy. MIT found that while official enterprise systems stalled, employees were quietly using consumer tools without sanction, generating real individual productivity that no enterprise system captured and no ledger recorded. The result is a strange double failure: the sanctioned deployments produce no measurable value because they were never integrated, and the unsanctioned usage produces value that is never measured because it was never governed. Both leave the income statement untouched.
Employees are quietly running a shadow AI economy. In MIT’s study, 90% of workers used personal AI tools for work, while only 40% of their employers had bought an official subscription.
Underneath both sits the single largest cause of the 95 percent, and it is not the model. It is the data. Gartner attributes the majority of AI project failure to the absence of AI-ready data, defined as data aligned to a specific use case, governed at the asset level, supported by automated pipelines with quality gates, and continuously assured. Most enterprises do not have it. They have data that was wrong, missing, mislabelled, or scattered across systems that never spoke to each other, and they assumed a powerful enough model would compensate. It cannot. A model trained or grounded on dirty data learns the dirt, and the pilot that looked promising on a clean sample collapses when it meets production volume and variety. The organisations that will report AI-driven gains in 2026 and beyond are, almost without exception, the ones that stopped launching new pilots and started fixing the data foundation first. The data is not a precondition for the AI. The data is the AI's actual substrate, and the readiness to supply it cleanly is what the N dimension of BEACON exists to measure.
The enterprises stuck in the 95 percent did not buy the wrong model. They attempted the destination without the readiness. They pointed at CORE without scoring on BEACON.
The paradox is not evidence against AI. It is evidence that the market bought capability and skipped readiness. The $2.59 trillion is real and the 95 percent is real, and both will hold until institutions stop funding pilots and start funding foundations.
Read every adoption headline through one filter: did it touch a governed, regulated process and move a number a chief financial officer will sign. If not, it is activity, not value.
The market has no shortage of frameworks. It has a shortage of clarity about which question each one answers. The Global AI Forum uses two, and never confuses them.
Every credible AI transformation carries four layers, and a board that cannot name all four is funding pilots, not a transformation. The destination has a shape, and the shape is CORE. It is a map, not a score. It answers a single question: where does AI actually create value in an enterprise. A chief executive is pointed at CORE. They are not graded against it.
The chief executive owns AI personally. Capital is allocated, governance exists, and there is a deliberate strategy for the data moat underneath everything. Without Command, nothing else holds. It is the gateway condition, not one option among four.
Process reinvention and workflow AI. Claims handled in seconds, documents reviewed in minutes, downtime predicted before it happens. This is where MIT found the highest and most ignored returns, and where the 5 percent quietly win.
Personalisation, new markets, product innovation, data monetisation. The hardest layer and the one boards most want. Almost nobody reaches it, which is why almost nobody reports growth from AI, only cost savings.
Customer delight and employee amplification. The advisor who is fully present because the machine took the notes. The clinician who looks at the patient, not the screen. The layer customers and staff actually feel.
CORE comes with an accessible journey, a four-stage maturity ladder that lets a board self-locate without a consultant in the room: Pilot, Scale, Transform, Redefine. Pilot proves a use case works. Scale puts it in production across a function. Transform reshapes the operating model around it. Redefine changes the business model itself. The entry point is a five-question board agenda. Does the chief executive own AI personally. What is our data moat. Is AI producing top-line revenue or only cost savings. Are we training everyone. Will AI change the business model. A board that cannot answer those five has, by its own admission, just located itself at the bottom of the ladder.
One feature of the CORE map is easy to miss and decisive in practice: Command is not one of four equal options. It is the gateway. An institution can run Operations pilots and Experience demos and even chase Revenue plays, but without the chief executive's personal ownership, a deliberate data strategy, and a governance structure that lets AI touch regulated processes, none of it compounds. This is why the failures so often had impressive activity and no accountability, an innovation lab generating proofs of concept that no business owner was obligated to scale. The 5 percent inverted that. They established Command first, made the data moat a board-level strategy rather than a technology backlog item, and only then funded the other three layers. The map has four destinations, but it has one entrance, and the institutions that tried to skip it are well represented in the 95 percent.
Knowing the destination is not the same as knowing whether you can get there. CORE describes what good looks like. It says nothing about where your institution stands, what is blocking it, or what the return would actually be once discounted by your real readiness. That is the gap BEACON closes. BEACON is the one number this report treats as a score, and the only one a chief executive should ever quote. It measures six dimensions, each carrying a signature metric that no generic maturity framework computes.
| Dimension | Signature metric | What it actually measures |
|---|---|---|
| B · Business value | Strategic Half-Life | Years before a use case becomes table stakes. A play with a two-year half-life is a cost, not a moat. |
| E · Engineering | Core Reachability | Share of value-creating processes that AI can actually reach on the core platform, not in a sandbox. |
| A · AI capability | Escape Velocity | Share of pilots that reach twelve months of sustained production. The exit rate from pilot purgatory. |
| C · Compliance | Time-to-Trust | Median days from model-ready to compliance-approved. In regulated sectors this number decides everything. |
| O · Operating model | Augmentation Quotient | The augment-versus-replace ratio, weighted by adoption. Whether the workforce is amplified or merely threatened. |
| N · Numbers | Data Sufficiency Index | Share of required data at usable quality. The single largest cause of the 95 percent, carried at the highest weight. |
The weighting is not arbitrary, and it encodes the lessons of the 95 percent. Numbers carries the highest weight, twenty-five of a hundred, because insufficient data is the single largest cause of failure and no amount of capability compensates for it. Engineering and AI capability carry twenty each, because reaching the core and exiting pilot purgatory are the two mechanical gates between a demo and a deployment. Business value, Compliance, and Operating model carry fifteen each, the dimensions that decide whether a working deployment becomes a durable, governable, adopted one. Read together, the six produce not a grade but a shape, and the shape is the diagnosis. A profile strong on data and engineering but weak on operating model describes an institution that can build but cannot get its workforce to adopt. A profile strong on capability but weak on compliance describes a regulated institution whose best models will never clear the lane to production. The composite number tells a board which band it is in. The shape tells it exactly what to fix first.
That is the difference between a maturity model and an instrument. A maturity model assigns a stage and leaves the institution to infer what to do. An instrument measures the specific constraint, prices it, and points at the move. Nascent, Emerging, Advancing, and Leading are the bands a composite score falls into, and they map cleanly onto the CORE maturity ladder of Pilot, Scale, Transform, and Redefine, but the band is the headline, not the insight. The insight is always in the lowest-scoring dimension carrying the highest weight, because that is the binding constraint, and the binding constraint is the only thing worth funding next.
Here is where the two lenses fuse into a decision. A naive business case for an AI initiative computes a return as if the institution were ready to capture it. It almost never is. Readiness-Adjusted ROI discounts the projected return by the institution's actual BEACON score, dimension by dimension, and the difference between the two numbers is not an accounting nuance. It is the precise reason 95 percent of pilots return nothing. The naive case said the value was there. The readiness was not, so the value never arrived.
Four frameworks, one job each, and only one is ever scored. CORE is the destination, pointed at. BEACON is the readiness, measured. The maturity ladder is the plain-language version of where a BEACON band places you. The roadmap is how execution is structured. The moment any of them competes to be the framework, it is collapsed back. Destination times readiness equals the plan.
A maturity model assigns a stage and leaves you to guess the move. An instrument measures the constraint and points at it. The discipline is to keep CORE as the map and BEACON as the score, and never let either compete to be the whole framework.
Destination times readiness equals the plan. Most institutions have a destination and no honest measure of readiness. That is precisely why most institutions are in the 95 percent.
AI adoption is now near-universal and wildly uneven at the same time. The map that matters in 2026 is not who is using AI. It is who is ready to capture its value, and who is governing it before they scale.
Ask three credible institutions for the global AI adoption rate and you will get three defensible answers, because they are measuring three different things. McKinsey's enterprise survey, covering nearly 2,000 companies across 105 countries, finds 88 percent using AI in at least one business function. The OECD's official firm-level measurement puts realised adoption at 20.2 percent, more than double its 2023 figure. Microsoft's population tracking of working-age adults who actually opened a generative tool lands near 16 percent. None is wrong. They asked different questions. The honest reading is that access is approaching saturation among large enterprises and remains thin among small ones, that realised use lags access badly, and that the gap between the two is the whole story.
The United States remains the gravitational centre of AI. It accounts for roughly 38 to 47 percent of global enterprise AI spending depending on the definition, hosts the frontier labs, and produces more AI startups and venture capital than any other nation combined. Private AI investment reached $109 billion in 2024, close to twelve times China's and twenty-four times the United Kingdom's. Enterprise adoption exceeds 85 percent, concentrated in finance, healthcare, and technology. Yet on population-level use the United States sits around 28 percent, behind Singapore and the Gulf, and its enterprise growth rate has been comparatively flat. The American story is capability and capital, not velocity. It is also where the reckoning is loudest, because it is where the spend is largest and the CFO scrutiny most advanced.
The American market is also where the structural response to the divide is most visible. When the frontier labs concluded that capability had outrun deployment, they built institutions specifically to close the gap: the OpenAI Deployment Company and the Anthropic enterprise joint venture are American constructs, capitalised by American private equity, aimed at the American enterprise base first. The logic is that the United States has the most capable models, the most AI spend, and the largest population of enterprises sitting on unrealised value, which makes it simultaneously the biggest opportunity and the most scrutinised market. The reckoning is loudest here because the stakes are largest here. The same conditions that produced the $2.59 trillion produced the CFO who now wants it justified, and the resolution of that tension in the United States will set the template the rest of the world adopts.
China treats AI as a national strategic priority and leads on enterprise deployment, near 58 percent, second only to the most aggressive adopters. State direction compresses the distance from policy to deployment, and the domestic model ecosystem has matured fast. The constraint is less adoption than the measurement and governance maturity that turns deployment into durable advantage, a constraint China shares with every fast-moving market.
Europe's adoption is lower and its posture is distinct. EU enterprise AI use reached 19.95 percent in 2025 by Eurostat's official measure, with a steep size gap, 55 percent for large enterprises against 17 percent for small. The defining feature is regulatory crystallisation. The EU AI Act has moved from draft to enforcement, classifying credit, health, and life-insurance decisioning as high-risk, which shifts adoption toward governance, documentation, and audit trails before scale. This is slower. It is also, on the BEACON Compliance dimension, a head start. The institutions that build Time-to-Trust into their architecture from the beginning will not have to retrofit it under penalty later.
The United Arab Emirates leads the world on population-level AI usage, above 60 percent, the product of deliberate state investment and a sovereign-AI agenda that treats compute, models trained on local language and values, and national data infrastructure as strategic assets. Saudi Arabia and the wider Gulf follow the same logic. Sovereign AI is not a slogan here. It is procurement policy, and it is reshaping where data is allowed to live, which is precisely the question on-premise and in-perimeter deployment was built to answer.
No market embodies the divide between adoption and readiness more sharply than India. By the inaugural State of Enterprise AI India 2026 analysis of 500-plus deployments, Indian enterprise AI adoption reached 80 percent, surpassing the United States and global averages and making India the most aggressive adopter on earth. The market is racing toward $17 billion by 2027, the talent pool toward 1.25 million by the same year, with big technology committing $67 billion-plus and more than ten thousand crore of domestic capital flowing into AI initiatives.
And yet only 23 percent of those enterprises have formal AI governance. This is India's paradox stated in two numbers: it leads the world in adoption and talent, and trails in the governance maturity that turns velocity into durable value. The question for Indian boards has already shifted from whether they are using AI to whether they can prove it works, and the proof requires exactly the readiness foundation that 77 percent have not yet built.
The texture underneath is specific. EY's work finds that 74 percent of Indian financial firms have launched proofs of concept and only 11 percent have reached production, with generative AI projected to drive productivity gains up to 46 percent in Indian banking operations by 2030. Customer service leads the use cases at 68 percent, and early results are real: 63 percent report improved customer satisfaction, 58 percent report cost reductions. The Global Capability Centres that anchor so much of India's enterprise stack are themselves becoming the centre of gravity for agentic AI, with 83 percent engaging generative AI and 58 percent developing agentic capabilities, shifting GCCs from cost centres to engines of enterprise value. India also leads globally on workforce sentiment: 86 percent of employees report a positive productivity impact from generative AI, the highest AI Advantage of any major economy.
The European posture deserves a closer look, because it is becoming the template others will be measured against. The EU AI Act's risk classification turns adoption into a documentation exercise before it becomes a scaling exercise: a credit or insurance model is not merely built and deployed, it is logged, explained, monitored, and audited, with penalties for failure that reach into the percentage of global turnover. To an enthusiast this looks like friction. To a clear-eyed board it looks like a forced investment in exactly the Compliance dimension that decides whether a model in a regulated sector ever reaches production. The European institutions building Time-to-Trust into their architecture now are buying an advantage that the lightly governed markets will have to retrofit later, under penalty, after the fact. Slower adoption is not the same as lower readiness, and on the Compliance axis Europe is ahead precisely because it was made to be.
Beyond the major blocs, the emerging markets are adopting through a different door. Where Western enterprises layer AI onto decades of legacy systems, many institutions across Southeast Asia, Latin America, and Africa are adopting through the cloud and the smartphone, without the brownfield integration burden that slows the incumbents. This leapfrog dynamic produces fast population-level usage and real consumer-facing innovation, particularly in payments and financial inclusion, while the enterprise-grade governance and data infrastructure lag further behind. The readiness gap there is widest of all in absolute terms, which is also why the opportunity to build it correctly from the start is largest.
Japan occupies a distinct position. With a shrinking workforce and an ageing population, AI is less an efficiency option than a demographic necessity, and the government has paired that pressure with one of the world's most permissive postures on AI and copyright to accelerate development. Adoption is rising fast from a cautious base, concentrated in manufacturing, robotics, and the physical AI where Japan's industrial base is strongest. South Korea follows a similar logic with state-backed compute and a chip-manufacturing spine. Across the broader Asia-Pacific, the pattern is bifurcated: Singapore leads on per-capita usage and governance sophistication, India leads on raw enterprise adoption and talent volume, China leads on state-directed scale, and a long tail of emerging markets adopts through the smartphone and the cloud rather than through on-premise enterprise systems. The region as a whole is adopting faster than the West, and governing more lightly, which means its readiness gap is wider in absolute terms even as its ambition runs ahead.
Cutting across all of these geographies is the rise of sovereign AI, the most consequential policy shift of the period. Nations have concluded that models trained on their own language, culture, and values, running on domestic compute, over data that never leaves the national perimeter, are strategic infrastructure on the order of energy or telecommunications. The Gulf states have made it procurement policy. India is building indigenous foundation models and national compute. The European Union has wrapped it in regulation. The practical consequence for any enterprise operating across borders is that the question of where data is allowed to live, and where a model is allowed to run, is now a board-level constraint rather than a technical detail. This is precisely the condition that on-premise and in-perimeter deployment was built to satisfy, and it is why the institutions treating data residency and model sovereignty as first-class design requirements, rather than afterthoughts, are the ones whose AI strategy will survive contact with the next regulation.
No region leads on every dimension. The United States leads on capital and capability, the Gulf on sovereign ambition, Europe on compliance, India on adoption and talent. The lesson is not to copy a geography. It is to copy a posture.
Adopt with governance, build with data residency in mind, and measure readiness rather than enthusiasm. The institutions that treat sovereignty and governance as first-class design requirements are the ones whose strategy survives the next regulation.
Nine industries, read the same way each time. Where AI creates value on the CORE map. The named institutions that disclosed real numbers. What separated the 5 percent from the 95. And the BEACON dimension that decides the sector.
The cross-sector pattern is worth stating before the detail, because it holds in every industry that follows. The winners are not generic. They are domain-specific and workflow-integrated, deployed where the data is clean and the process is owned. The losers are flashy pilots that never touched the core. MIT found that only two of nine major sectors, technology and media, showed material business transformation from generative AI. This report is, in part, an explanation of why the other seven have not, and what the exceptions inside them did differently.
One finding frames everything that follows. MIT's analysis concluded that only two of nine major sectors, technology and media, have shown material business transformation from generative AI, while the remaining seven, the regulated, data-rich, physically grounded industries where most of the economy actually lives, have produced abundant pilots and scarce transformation. That is not a verdict on those industries. It is a verdict on how AI was deployed in them. The sections below are, in effect, nine investigations into the same question: in each industry, what did the small minority who captured real value do that the large majority did not. The answer rhymes across all nine. The winners went where the data was clean and the process was owned, integrated to the core rather than bolted to the edge, and built the foundation before they chased the headline. The losers ran the demo and called it a strategy.
Banking is the most advanced AI sector in the world by spend, and the clearest case study in the limits of efficiency without transformation. The exemplar is JPMorgan Chase. Its annual technology budget approaches $18 to 20 billion, with roughly $3 billion dedicated to AI, and a data and analytics function reporting directly to the chief executive. The firm runs more than 400 AI use cases in production, on a path to a thousand, and has deployed its proprietary LLM Suite to over 200,000 employees. It estimates $1.5 to $2 billion in tangible annual business value from AI, a figure Jamie Dimon places alongside the printing press and electricity. COiN reviews 12,000 commercial credit agreements in seconds. The fraud platform protects against losses the industry projects could reach $58 billion by 2030. Engineering teams report 10 to 20 percent productivity gains.
JPMorgan’s contract-intelligence engine COiN reviews 12,000 commercial credit agreements in seconds, work once estimated to consume roughly 360,000 lawyer-hours a year.
possibly as transformational as the printing press, the steam engine, electricity.
This is real, disclosed, audited value, and it is also the most important cautionary tale in this report. Examine what the value consists of. Document review, pitch-deck generation, fraud detection, coding assistance. These are Operations-layer efficiency plays. They make existing processes faster and cheaper. They sit at the bottom of the CORE map. And they are replicable: Goldman Sachs has deployed its GS AI Assistant to over 10,000 bankers and traders, Morgan Stanley its OpenAI-powered Assistant to 98 percent of advisor teams, Citi equivalent capabilities, all on the same underlying models. JPMorgan itself swaps its foundation models every eight weeks, which is a confession that the models are commoditising. When the technology is interchangeable, the efficiency advantage converges across competitors, and the moat closes.
Banking has proven AI can take cost out. It has not yet proven AI can build a durable moat. The institution that wins the next decade is the one that moves from Operations to Revenue and Command, from cost savings to a data advantage no competitor can swap in a new model to replicate. JPMorgan has built the most impressive AI infrastructure in financial services and is still missing that layer. So is everyone else.
The banking AI value chain runs from the back office forward, and the disclosed value tracks that order precisely. The deepest returns sit in operations and risk: fraud detection applied across billions of transactions, anti-money-laundering triage, credit-decision support, regulatory-reporting automation, and the document-heavy machinery of commercial lending. The middle of the chain, the developer and analyst copilots, produces the productivity gains every bank now quotes. The front of the chain, client-facing generative AI, is where banks move slowest, because accuracy and compliance thresholds are unforgiving when the counterparty is a customer rather than an employee. JPMorgan's own leadership has been candid that a full return will take years, naming a value gap between what the technology can do and what the enterprise can capture, the exact gap this report is about. The laggards in banking are not the institutions without models. Every bank has models. The laggards are the ones that deployed copilots without touching the core systems where the data and the risk actually live, and so produced shadow productivity that never reached a regulated process or an income statement.
The pattern globalises with a twist. In India, where BFSI is the leading AI sector and 90 percent of financial institutions name AI a core investment, the deployment is real but the maturity is uneven: 74 percent of firms have launched proofs of concept and only 11 percent have reached production, with NBFCs racing ahead on business-intelligence automation while large banks build enterprise-scale underwriting and cybersecurity copilots. The destination is the same everywhere. The readiness to reach it is what varies by institution and by market.
Insurance is the sector where AI is most visibly rewriting the underlying economics, not just the workflow. The exemplar is Lemonade. Its AI chatbot Maya generates a quote in under 90 seconds. Its claims assistant Jim settles a covered claim in as little as three seconds with no human in the loop. The result is not a productivity anecdote, it is a structural cost advantage: loss-adjustment-expense ratios near 4 percent, against the much higher handling costs of traditional carriers. By the first quarter of 2026 Lemonade passed 3.1 million customers, up 23 percent, with $1.3 billion of in-force premium. Its cost to handle a pet claim fell from $44 in 2021 to $14 in 2025, a level it calls industry-leading and a direct readout of AI compounding across operations. With over 90 percent of customers running continuous telemetry, its European gross loss ratio improved 16 points to 74 percent even as premium grew 150 percent. That is Operations and Revenue moving together: cheaper claims and sharper pricing at once.
The incumbents are not standing still, and their numbers are also real. Aviva runs over 80 AI models on motor claims, cutting 23 days from liability determination on complex cases. Across the industry, AI-powered claims now resolve roughly 75 percent faster, compressing a 30-day cycle to 7.5 days, with simple claims clearing in 24 to 48 hours. AXA worked with Sprout.ai to pull processing on thousands of life and health claims down to minutes. Allianz is among the most carrier-wide AI-integrated groups in the world, and Munich Re and Hannover Re lead in reinsurance. The competitive truth, stated by analysts at SG Analytics, is uncomfortable for incumbents: insurtechs have been running AI-native claims for years, their operating costs keep falling while carriers' stay high, and the gap is widening, not closing.
What makes insurance the clearest case in this report is that AI touches every link of its value chain at once, and the links compound. Distribution gets faster quotes. Underwriting gets sharper risk selection from telemetry and alternative data. Pricing gets granularity no actuary could compute by hand. Claims get speed and consistency. Fraud, which costs the United States industry an estimated $80 billion a year, gets pattern detection at scale. Each improvement feeds the next: more telemetry sharpens pricing, sharper pricing improves the loss ratio, a better loss ratio funds more competitive premiums, and the data accumulated along the way trains better models still. This is a flywheel, and it is why insurtech operating costs fall while incumbent costs stay flat. The incumbent dilemma is structural, not a question of effort. A carrier with a thirty-year-old policy-administration system cannot simply bolt three-second claims onto it, because the core was never built to be reached. The winners rebuilt the core or, like the digital-native carriers, never carried the legacy in the first place. The AI is the visible part. The architecture underneath it is the moat.
Lemonade’s claims bot Jim has settled a covered claim in as little as three seconds, with no human in the loop, faster than most people can read the confirmation message.
Beyond retail banking and insurance sits the broader machinery of capital markets, payments, asset management, and research, where AI is reshaping how money is moved, priced, and understood. Financial services leads every adoption survey, at roughly 87 percent, because the sector has three structural advantages: enormous proprietary datasets, quantifiable outcomes, and a single basis-point improvement in fraud or lending applied across vast transaction volumes produces real money. This is why banks invested in data science long before generative AI arrived, and why the sector's data foundations, the N dimension, are deeper than most.
The frontier here is agentic. Gartner expects 40 percent of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5 percent at the start, and financial services is where multi-step autonomous workflows first touch the core: reconciliation, regulatory reporting, KYC, trade processing. Morgan Stanley is opening its trillion-dollar wealth platform to external AI agents, letting clients' autonomous systems pull data directly from its stock-administration platforms, an early sign of where the sector is heading. But the same Gartner data warns that over 40 percent of agentic projects will be cancelled by 2027 on cost, unclear value, and inadequate controls. The sector that adopts fastest also cancels fastest. Readiness, not enthusiasm, separates the two.
Inside the broader sector, three sub-domains show distinct patterns worth separating. In capital markets, AI accelerates research synthesis, surveillance, and the unglamorous reconciliation work that consumes back-office headcount, with the frontier moving toward agents that execute multi-step processes rather than answer questions. In payments, AI is a fraud-and-risk engine first, where a basis point of accuracy improvement is measured directly in losses avoided. In asset management, the value splits between investment research augmentation and the operational efficiency of client servicing and reporting. Across all three, the institutions that capture value share one trait: they treated their proprietary data as the asset and the model as the commodity, which is why the sector's data foundations run deeper than any other and why its Numbers dimension is, on average, its strongest. The sector's risk is the inverse of its strength. Saturated with pilots and flush with budget, financial services generates more abandoned proofs of concept than any other, and the agentic wave will sharpen that divide, rewarding the disciplined and punishing the enthusiastic at the speed and cost that autonomous systems impose.
Healthcare is the sector where AI's first durable wins are about giving people back their attention. Ambient clinical documentation, AI that listens to a clinical conversation and writes the note, has moved from pilot to enterprise reality with disclosed, audited results. Houston Methodist rolled it out enterprise-wide and reported a 40 percent reduction in documentation time, a 27 percent increase in patient face time, and 1.3 additional voluntary visits per clinician per day. The Cleveland Clinic deployed ambient documentation to over 4,000 clinicians, saving 14 minutes per provider per day, alongside Bayesian Health's sepsis detection running at 13 hospitals with 10 times fewer false alerts and 46 percent more cases caught. A University of Wisconsin randomised trial in NEJM AI found ambient AI cut documentation by 30 minutes per clinician per day and meaningfully reduced burnout. A multicentre JAMA study found a 31 percent drop in reported burnout.
The scale is now historic. Kaiser Permanente deployed Abridge's ambient documentation across 40 hospitals and over 600 medical offices, the largest generative AI rollout in healthcare history and its fastest technology implementation in two decades. Mayo Clinic is investing over $1 billion across more than 200 AI projects, extending into diagnostics and nursing. Advocate Health evaluated 225 AI solutions to select 40 for production. The market has organised around two value pools that deliver immediate ROI: ambient documentation, roughly a $600 million market, and coding and billing automation, roughly $450 million. The discipline is striking. Leading systems prioritise production-ready tools, low patient-care risk, and quick wins that build organisational confidence, then stack early wins into operational muscle. The result: more than half of health systems that can quantify AI ROI report at least 2x returns, and 75 percent of US systems now use or plan to use an AI platform.
Kaiser Permanente’s ambient-AI rollout reached 40 hospitals and more than 600 medical offices, the largest generative AI deployment in healthcare history and the fastest technology rollout in the system’s two decades.
Healthcare's success is, paradoxically, a story of restraint. The systems that scaled did not chase the most ambitious clinical AI. They sequenced deliberately, starting where patient-care risk was lowest and ROI was fastest, then stacking confidence. Documentation and revenue cycle came first because they are administrative, measurable, and safe. Diagnostic and patient-facing AI came later, under deeper scrutiny and longer timelines. This sequencing is the discipline the 95 percent lacked elsewhere, and healthcare arrived at it precisely because the stakes of getting it wrong are mortal. The financial mechanics reinforce the choice: a prior-authorisation platform reporting eight times ROI, an appointment-scheduling deployment at 468 percent, a claims-appeals process cut from sixteen days to two. These are revenue-cycle and operations wins, the same back-office gravity MIT identified across every sector. The laggards in healthcare are the systems that bought clinical AI before fixing the data integration underneath it, because healthcare data is fragmented across electronic records, lab systems, and imaging databases that rarely speak to each other, and a generic model on fragmented data hallucinates in a setting where hallucination is dangerous. The 5 percent fixed the data plumbing first. The 95 percent assumed the model would compensate for the mess. It did not.
Manufacturing delivers the highest average AI ROI of any industry, roughly 200 percent, for a structural reason: factory operations provide quantifiable baselines and direct cost-to-savings mappings. A machine that fails costs a known amount per hour. A defect that escapes costs a known amount per unit. When AI moves those numbers, the return is not a productivity narrative, it is arithmetic. Predictive maintenance, the most deployed use case, achieves 400 to 500 percent three-year ROI and cuts unplanned downtime 30 to 50 percent. Continental cut downtime 37 percent across four tire plants for over 8 million euros in annual savings. BMW runs computer-vision inspection across 31 plants and reduced defect escape rates 85 percent against human-only inspection. Siemens drove defects from 500 to 12 per million.
On a single line, Siemens cut defects from 500 per million to 12, a 97 percent reduction, by letting AI catch what the human eye kept missing.
The frontier is the AI-native factory itself. Siemens and NVIDIA are building an Industrial AI Operating System, with the first fully AI-driven adaptive site launching in 2026 at Siemens' Erlangen electronics factory. The model is a factory that continuously analyses its own digital twin, tests improvements virtually, and pushes validated changes to the shop floor. PepsiCo, using Siemens' Digital Twin Composer, simulates entire plants and identifies up to 90 percent of potential issues before physical implementation. Foxconn, HD Hyundai, and KION are evaluating the same stack. Yet even here the divide holds: 42 percent of manufacturers have deployed AI in some form, and only 12 percent have moved beyond single-use-case deployments to enterprise scale. The typical manufacturer sits at structured experimentation, with strong process discipline and real gaps in the OT and IT integration that decides whether the twin can reach the line.
The reason manufacturing both leads on ROI and lags on enterprise scale lives in two letters: OT and IT. Operational technology, the sensors, controllers, and machines on the floor, was built in a different era and a different language from the information technology that runs the model. Bridging them is the entire game. Where the bridge exists, the returns are the highest and most measurable in enterprise AI, because a failed machine and an escaped defect carry known costs. Where it does not, the most sophisticated model in the world cannot reach the line, and the pilot dies in a sandbox that never touched production. The encouraging shift in 2026 is that the bridge is getting cheaper: standard protocols, falling edge-compute costs, and cloud-native industrial platforms have compressed OT and IT integration from multi-year programmes to projects of six to twelve months, and regulatory certainty from the EU Machinery Regulation and AI Act has, paradoxically, accelerated adoption by telling manufacturers exactly what governance to build against. The frontier players are now wiring AI into everything: Siemens deploying hundreds of industrial AI experts alongside NVIDIA infrastructure, copilots reaching the shop floor through smart glasses, and digital twins that test a change in software before a single machine moves. The gap between the 12 percent at enterprise scale and the rest is not a gap in ambition or even in capital. It is a gap in whether the model can reach the metal.
Retail is where AI is most directly attached to the top line, because the levers, personalisation, demand forecasting, and conversational commerce, touch revenue immediately. The standout disclosed number belongs to Amazon: its Rufus shopping assistant is generating an estimated $12 billion in incremental sales. Walmart runs a self-healing inventory system. Across the sector, NVIDIA's 2026 survey found 91 percent of retail and CPG companies actively using or assessing AI, and 90 percent planning to raise AI budgets. The shift in 2026 is from analytics to agency: agentic systems that do not just recommend but act, adjusting messages, recommending products, and guiding purchase decisions in real time based on individual context. Personalisation done well yields three to four times better ROI on engagement campaigns and revenue lifts up to 20 percent.
Amazon’s Rufus shopping assistant is generating an estimated $12 billion in incremental sales, more than the entire annual revenue of many national retailers.
The caution is the same as everywhere. Retail's value is real where the data is clean and the workflow is owned, in demand forecasting, allocation, replenishment, and service automation. It evaporates in the flashy customer-facing pilots that dominate budgets and underperform on return, the precise misallocation MIT identified. The disciplined retailer targets AI at rework loops in the contact centre and trade-spend effectiveness, the unglamorous functions where the money is.
Retail is entering its agentic phase faster than most sectors, and the shift changes what value looks like. The first wave of retail AI was analytical: predict demand, recommend products, segment customers. The second wave is agentic: systems that do not just predict but act, adjusting a price, triggering a promotion, rerouting inventory, guiding a purchase in real time on the basis of an individual shopper's context. NVIDIA's survey found nine in ten retailers raising AI budgets specifically toward agentic and physical AI, the latter automating warehouse and supply-chain operations with robotics. This is genuine Revenue-layer territory, the rare place AI touches the top line directly. But retail also illustrates the misallocation trap better than any sector, because its most visible AI, the consumer-facing chatbot and the flashy personalisation demo, is precisely where ROI is weakest, while the durable returns sit in the contact-centre rework loops, the trade-spend optimisation, and the replenishment engines that never make a press release. The disciplined retailer spends its budget where MIT found the money and accepts that the unglamorous functions are the profitable ones.
CPG is the supply-chain twin of retail, and its AI value is concentrated where physical goods meet prediction. Unilever's AI demand-sensing saved an estimated $300 million annually, and its weather-integrated forecasting lifted ice-cream sales up to 30 percent in some markets by pre-positioning inventory ahead of demand spikes, improving forecast accuracy 10 percent in Sweden alone. Unilever is building a digital twin of its global supply chain using satellite imagery to simulate disruptions before they become crises. On the creative side, Coca-Cola's Project Fizzion with Adobe produces branded content up to 10 times faster and cut time-to-market for creative assets by up to 90 percent, while its Freestyle and smart vending machines feed real-time flavour preferences back into product development. Nestlé works with IBM on generative AI for packaging materials.
Unilever’s weather-fed forecasting lifts ice-cream sales by up to 30 percent in some markets, pre-positioning stock days before a heat wave arrives.
The sector-level prize is large and specific. Industry estimates put AI's value unlock for global CPG at roughly $500 billion, with 2 to 5 percent profitability gains and supply-chain errors halved. The most advanced players, Nestlé, Coca-Cola, L'Oreal, Procter & Gamble, Unilever, embed AI across decisions and product innovation, blending traditional, generative, and agentic approaches. The winners treat AI as supply-chain infrastructure. The laggards treat it as a content-generation novelty.
The defining CPG move of 2026 is the supply-chain digital twin, and it reframes AI from a marketing tool into operating infrastructure. A consumer-goods company that can simulate a global supply chain, model a heat wave or a supplier failure or a port closure before it happens, and pre-position inventory and promotion accordingly, holds an operational advantage that compounds every season. Unilever's satellite-fed twin and PepsiCo's plant simulations are the leading examples, and within twelve to thirty-six months the analysts expect twins to become standard practice across major CPG companies and large retailers. The competitive pressure underneath is the direct-to-consumer shift, a market heading toward $320 billion, which is why incumbents like Unilever and Procter & Gamble have been acquiring DTC brands to defend their position and feed their data engines. CPG's AI value, in the end, is a data-and-supply-chain story dressed occasionally in generative-content clothing. The companies that understand which is the substance and which is the decoration are the ones capturing the $500 billion the sector is projected to unlock.
Wealth management is a $112 trillion industry generating an estimated $312 billion in annual advisory fees, and it is being reshaped by AI in a way that proves a subtle point: in a relationship business, the best AI makes the human more present, not less necessary. Morgan Stanley's AI Assistant, built on OpenAI and drawing on roughly 100,000 research documents, has been adopted by 98 percent of its financial-advisor teams. Its companion tool, AI @ Morgan Stanley Debrief, generates meeting notes and action items with client consent, so the advisor can have a deeper conversation instead of taking dictation. Client assets in its wealth division reached $7.4 trillion, on a path beyond $10 trillion. The firm is now opening that funnel to external AI agents, letting clients' autonomous systems pull data directly from its platforms.
98% of Morgan Stanley’s financial-advisor teams adopted its AI assistant, one of the fastest enterprise AI adoptions ever recorded, in a business built entirely on human trust.
The market is moving fast underneath. Agentic AI in wealth management was valued at $3.2 billion in 2025 and is projected to reach $42.8 billion by 2034, a 37 percent compound rate, one of the fastest-growing segments in financial technology. UBS, FNZ with Microsoft, and Allfunds are building competing capability. But the analysts are unanimous on the limit: the industry is relationship-driven, clients value trust more than cost efficiency, and the advisor's role has always been to earn the fee through judgment, behavioural coaching, and personalisation. AI disintermediates the back office. It does not disintermediate the relationship. The winners use it to scale the human, not to remove them.
The most instructive moment in wealth management came not from a product launch but from a market reaction. When concern spread that AI might disintermediate advisors, Morgan Stanley's wealth leadership made a point that applies far beyond finance: an individual tool is a tiny part of the capability ecosystem required to help a client, and that ecosystem has to connect to products, third-party managers, and a relationship a client can understand and act on. The firm's outsized growth came from orchestrating that ecosystem deliberately over years, not from any single model. Analysts agreed, noting that the industry is relationship-driven, that clients value trust over cost efficiency, and that while components of advice can be automated, the advisor earns the fee through judgment, behavioural coaching, and personal relationships that no model replaces. This is the augmentation thesis in its purest form, and it carries a warning for every sector tempted to read AI as a replacement story. The institutions winning in wealth are scaling the human. The ones that mistook the tool for the relationship are the ones whose clients will leave for a firm that did not.
Private equity has discovered that AI is a third value lever alongside financial engineering and operational excellence, and it is acting on that discovery at the top of the adoption curve. Deloitte's M&A study found 86 percent of organisations have integrated generative AI into M&A workflows, with PE leading at 88 percent. AI scans markets to surface targets at superhuman scale, identifying 195 relevant companies in the time a junior analyst evaluates one. EQT's Motherbrain platform consolidates over 140,000 data points for real-time deal insight. Blackstone has run AI deal-sourcing since 2021 and embedded it across 70-plus portfolio companies in pricing and labour. Blackstone and KKR report cutting due-diligence review times by up to 50 percent. Bain found nearly 20 percent of portfolio companies have operationalised generative AI with concrete results.
EQT’s Motherbrain platform scans more than 140,000 data points to surface acquisition targets, evaluating 195 companies in the time it takes a junior analyst to read one.
Yet the same MIT divide runs straight through the sector that adopts fastest. S&P Global found 41 percent of PE firms still in nascent AI adoption and only 7 percent fully integrated. And the most consequential development of 2026 is structural: the frontier labs are using private equity as a distribution channel into the enterprise. In May 2026, the OpenAI Deployment Company launched with multibillion-dollar backing from TPG, SoftBank, Brookfield, and Bain, while Anthropic formed a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to deploy its models into enterprise operations. Google is negotiating omnibus licensing across Blackstone, KKR, and EQT portfolios. The bet is explicit and it validates this entire report: the next phase of the AI economy will be fought not in benchmark leaderboards but inside portfolio companies, in deployment and readiness. The capital has decided that the missing layer is the one between a capable model and a company that can actually use it.
The deeper story in private equity is that AI is changing the diligence question itself. Sponsors no longer only ask what a target earns. They ask how exposed it is to AI disruption, whether it has the data and technology foundations to adopt AI effectively, and how its management plans to use AI to improve margins and growth. That is a readiness question, asked at the moment of capital allocation, by the most return-focused investors in the world. The implication is profound: readiness is becoming a priced attribute of an enterprise, a factor in its valuation, not a soft capability. And the largest sponsors hold a structural edge, because they can pool data across thousands of portfolio companies and apply learnings at a scale no single operator can match. This is exactly why the frontier labs are paying billions to enter the channel. Blackstone and KKR together manage assets exceeding two trillion dollars across thousands of companies spanning healthcare, logistics, technology, and financial services, and whoever helps that portfolio deploy AI captures the largest new enterprise distribution channel to emerge since cloud computing. The labs have concluded that the bottleneck in enterprise AI is implementation, not capability, and have committed capital accordingly. The whole of this report is an elaboration of why that conclusion is correct.
Nine industries, one shape. The winners were domain-specific and workflow-integrated. The losers ran flashy pilots that never touched the core. Only two of nine sectors show material transformation, and the other seven are not failing the technology, they are failing the integration.
The named institutions that disclosed real value did the unglamorous work first. JPMorgan, Lemonade, Houston Methodist, Continental, Blackstone. The headline followed the foundation. It never leads it.
Nine industries, one recurring shape. The institutions that captured value did the same five things. The ones that did not bought the same models and got nothing.
Read the disclosed numbers across all nine sectors and a single fact dominates. The value that reached an income statement was, with rare exception, Operations-layer value. Cheaper claims at Lemonade. Faster documents at JPMorgan. Less downtime at Continental. Restored attention at Houston Methodist. Compressed diligence at Blackstone. These are efficiency and automation wins. They make existing processes faster and cheaper. They are real, and they are the bottom of the CORE map. The Revenue layer, genuine top-line growth from AI, and the Command layer, AI changing the business model itself, remain almost empty. This is why McKinsey finds firms reporting cost benefits far more often than growth, and why so few report enterprise EBIT impact. The sector has learned to take cost out. It has not yet learned to grow.
It is worth asking why the Revenue layer stays so empty, because the answer is instructive. Cost reduction is internal, measurable, and controllable. An institution can deploy an efficiency tool, measure the hours saved, and book the result without anyone outside the building agreeing. Revenue growth is none of those things. It depends on customers behaving differently, on new products finding markets, on data being monetised in ways that regulators and partners permit. It requires reaching the Revenue and Command layers of CORE, which in turn require the data moat, the integration, and the mandate that most institutions have not built. So the value pools where the foundation is thinnest, and the result is a market full of companies that have made themselves more efficient and almost none that have made themselves grow. This is not a permanent condition. It is a sequencing problem. The institutions that build the foundation will reach the growth layer that the efficiency-only players cannot, and when they do, the gap between the 5 percent and the 95 percent will stop being a gap in cost and start being a gap in revenue, which is the gap that actually decides who survives.
The single most clarifying fact in enterprise AI is that JPMorgan swaps its foundation models every eight weeks. The most sophisticated AI operation in financial services treats the model as a replaceable component, because it is one. When the underlying technology is interchangeable, any advantage built purely on it converges across competitors within a quarter or two. This is the Strategic Half-Life problem, and it disqualifies most of what enterprises currently call their AI strategy. A capability every rival can buy and stand up in weeks is a cost of doing business, not a moat. The durable advantage lives one layer down, in proprietary data and the readiness to use it, which no competitor can swap in a new model to replicate.
MIT's most counterintuitive finding is its most actionable. Solutions purchased from specialised vendors succeed about 67 percent of the time. Internal builds succeed at roughly half that. Enterprises consistently underestimate the integration and workflow cost of building, and stall in pilots that never reach production. The instinct to build everything in-house, strongest precisely in the regulated, data-rich sectors this report covers, is a leading cause of the 95 percent. The winners focused their engineering on integration and data, and bought the capability that was buyable.
Almost everywhere we went, enterprises were trying to build their own tool.
BCG's research on AI transformations gives the allocation that the 95 percent ignored and the 5 percent obeyed: the 10-20-70 rule. Ten percent of the effort is algorithms. Twenty percent is technology and data. Seventy percent is people and process. The failures treated AI as a technology project and spent their effort on the 10 percent. The successes treated it as a transformation of how work gets done and spent their effort on the 70. This is why MIT found the binding constraints were organisational, not technical, and why the institutions with deep change-management muscle, the ones that designed for augmentation and trained the whole workforce, are the ones that scaled.
The next phase is agentic. Gartner expects task-specific AI agents in 40 percent of enterprise applications by the end of 2026, up from under 5 percent, and McKinsey finds 62 percent of organisations already experimenting with agents and 23 percent scaling them. Agents change the ROI model, because the comparison is no longer marginal productivity but fully loaded employee cost. But the same Gartner data forecasts over 40 percent of agentic projects cancelled by 2027 on cost, unclear value, and inadequate controls. Agents do not forgive an unready institution. They punish it faster, and more expensively. The divide that defined 2025 and 2026 will not close in the agentic era. It will widen, and the readiness gap will become the difference between a portfolio of autonomous systems and a portfolio of cancelled projects.
The throughline beneath the other five is measurement. The 95 percent did not only skip the foundation. They skipped the honest accounting that would have told them the foundation was missing. They ran pilots without a defined financial outcome, scaled on enthusiasm rather than evidence, and discovered the readiness gap only when the return failed to arrive. The 5 percent did the opposite. They measured their readiness before they spent, priced the gap, sequenced the fixes, and tracked the value to the income statement. This is not a coincidence of temperament. Measurement is the mechanism. An institution that cannot state its data sufficiency, its time-to-trust, its escape velocity, and its augmentation quotient as numbers is an institution operating on hope, and hope is what the trough of disillusionment is built to punish. The discipline that separates the winners is not a secret technology or a privileged model. It is the willingness to look at the readiness honestly, to express it as a number, and to act on the number rather than the narrative. Everything else in this report is an elaboration of that single habit.
This is the layer the frontier labs just spent billions to reach through private equity. It is the layer the 5 percent built before they started, and the 95 percent assumed they could skip. It does not show up in a model benchmark or a capital-expenditure line. It shows up only in whether the deployment worked. And it is, precisely, what an instrument like BEACON exists to measure and what a destination model like CORE exists to aim at. The market has been buying models for three years. It is about to start buying readiness.
The six lessons reduce to one sentence: the winners measured, the losers hoped. Build versus buy, the seventy percent that is people, the commoditising model, all of it follows from whether an institution was willing to look at its readiness honestly and act on the number.
The playbook that follows is that habit, written as a sequence. It is the order of operations that separated the deployments that worked from the ones that returned nothing.
The 95 percent is not a reason to slow down. It is a map of where the foundation was skipped. Here is the sequence the 5 percent followed, written as a board agenda rather than a technology plan.
Everything in this report converges on a single operational truth: the institutions that captured value built readiness before they chased the destination, and they built it in a specific order. The sequence below is not a maturity model to admire from a distance. It is the order of operations that separated the deployments that worked from the ones that returned nothing, expressed so a chief executive can run it without a translator.
The first move is not a use case. It is ownership. The chief executive takes personal accountability for AI, names where the data moat will come from, and sets the condition that no initiative proceeds without a path to a governed, regulated process. Every institution in the 5 percent had this. Most of the 95 percent had an innovation team and no mandate. Command is the gateway, and a board that cannot say who owns AI has already located its weakest dimension.
Before funding the destination, price the readiness. Score the six BEACON dimensions as they actually are, not as the roadmap wishes them to be, and compute the Readiness-Adjusted ROI of the leading initiatives. The number that matters is not the naive business case. It is the business case discounted by the institution's real capacity to capture it, and the gap between the two is the budget that would otherwise be wasted. This single act, pricing the gap, is what converts an AI strategy from a wish into a plan.
The Data Sufficiency Index is the highest-weighted dimension for a reason: insufficient data is the single largest cause of failure. The disciplined sequence is to stop launching new pilots, identify the specific data required for the highest-value use cases, and bring it to usable quality through governance and automated pipelines before the model is asked to depend on it. Foundation payback, in the engagements that did this, was measured in months. The institutions that inverted the order, deploying first and discovering the data gap in production, are the ones whose pilots populate the 95 percent.
Vendor solutions succeed at roughly twice the rate of internal builds. The implication is not to outsource everything. It is to spend scarce engineering on the integration and the proprietary data loop that constitute the actual moat, and to buy the commoditising capability that no longer differentiates anyone. The model is replaceable. The data advantage and the reach into the core are not. Direct effort accordingly.
Seventy percent of the work is people and process. The deployments that scaled gave the workforce time back rather than threatening it, and they trained broadly rather than concentrating capability in a specialist team. A high Augmentation Quotient is not a cultural nicety. It is the mechanism by which adoption actually happens, and adoption is the only path from a working pilot to a captured return. The institutions that framed AI as replacement stalled. The ones that framed it as amplification scaled.
Run in sequence, these five moves are the difference between the institution that deploys two and a half trillion dollars worth of industry capability and captures nothing, and the institution that deploys a fraction of it and reports a return its competitors cannot explain. The destination has been mapped. The readiness can be measured. What remains is the discipline to do them in the right order, which is, in the end, the only thing the 5 percent ever had that the 95 percent did not.
By May 2026 the world had deployed close to $2.59 trillion into artificial intelligence. The number will only rise from here. And so far, almost nothing fundamental has shifted.
That is not a contradiction. It is the defining condition of the year, and it deserves to be stated without the usual softening. The capital is real. Gartner's $2.59 trillion is not a forecast of intent, it is a measure of money already committing, up 47 percent in a single year, with the four largest hyperscalers alone driving toward $725 billion in capital expenditure and data-centre spending passing $788 billion. The infrastructure is rising on three continents. The chips are sold out. The build-out is the largest in the history of corporate technology, and 2026 is the year Gartner says enterprises finally start spending their own money rather than watching the hyperscalers spend theirs.
And against all of it stands one number that refuses to move. MIT's NANDA initiative, studying 300 deployments, found that 95 percent of enterprise generative AI pilots returned no measurable impact on profit and loss. Not a small return. None. McKinsey, from a different direction, found only 39 percent of organisations reporting any enterprise-level EBIT impact at all. A manufacturing chief operating officer gave the researchers the sentence that should sit at the centre of every board's AI review: the hype says everything has changed, and in our operations nothing fundamental has shifted.
The instinct is to read this as a verdict on the technology. It is not. The technology works, and the 5 percent prove it works spectacularly, with claims settled in three seconds, downtime cut by half, billions in disclosed value, attention given back to clinicians and advisors. The 95 percent is not a verdict on AI. It is a verdict on readiness. The institutions that captured value were not the ones that spent the most. They were the ones that built the foundation first: the data moat, the integration to the core, the compliance lane, the augmented operating model, the mandate owned from the top. They pointed at the destination and they measured their readiness to reach it. Everyone else bought a model and waited for a return that the readiness gap quietly consumed.
The reckoning now underway, the CFO scrutiny, the 25 percent of spend postponed to 2027, the agentic projects already being cancelled, is not the end of the AI investment cycle. It is the moment a market stops paying for potential and starts paying for proof. The capital has already begun to vote: the frontier labs are spending billions to buy their way into the deployment layer through private equity, because they have concluded, correctly, that the bottleneck was never the model. It was everything around the model that an institution either built before it started, or did not.
The $2.59 trillion will keep climbing. The question that decides which institutions are on the right side of the divide is no longer how much they will spend. It is whether they are ready. That question has an answer, and the answer is a number. The work of this decade is to measure it honestly, and then to close the gap.
There is a version of the next three years in which this report ages badly, and it is the version every chief executive should be working toward. In it, the 95 percent shrinks, not because the technology improved, but because institutions stopped treating readiness as something they could skip. The data foundations get built. The integration to the core gets done. The compliance lanes get shortened. The workforce gets trained and the operating model gets redesigned around augmentation. And the value that was always latent in the technology finally reaches the income statement, because the institution was finally ready to receive it. None of that requires a better model. The models are already good enough, which is the quiet scandal underneath the $2.59 trillion. The bottleneck was never on the supply side of intelligence. It was on the demand side, in the readiness of the enterprise to convert intelligence into outcomes, and that is the one variable a board actually controls.
The institutions that understand this are already moving. They have stopped asking how their AI compares to a benchmark and started asking how their readiness compares to their ambition. They have stopped measuring activity and started measuring the gap. They treat the destination as a map to be pointed at and the readiness as a number to be raised, deliberately, in the order the constraints demand. They are, in the language of this report, the 5 percent. And the only thing that distinguishes them from everyone else is that they decided to measure the thing that matters, and then to close it.
CORE is the destination · BEACON is the readiness · Global AI Forum
Every figure traces to a named, dated, primary or near-primary source. Where institutions disclosed their own numbers, those are used in preference to estimates.
A note on the numbers. Adoption figures across this report come from surveys that measure different things, enterprise function usage, firm-level realised adoption, and population-level usage, and where they appear together they are labelled as blended and directional rather than precise. Institutional outcomes, the disclosed values from JPMorgan, Lemonade, Houston Methodist, Unilever and others, are taken from company filings, shareholder letters, and named executive statements in preference to third-party estimates, and are presented as the institutions themselves stated them. The two anchoring statistics, Gartner's $2.59 trillion in 2026 worldwide AI spend and MIT NANDA's finding that 95 percent of enterprise generative AI pilots returned no measurable profit-and-loss impact, are drawn directly from those organisations' 2025 and 2026 publications. The CORE and BEACON frameworks, the maturity ladder, the six signature metrics, and the Readiness-Adjusted ROI model are proprietary instruments of the Global AI Forum, and the figures shown for them are illustrative, chosen to demonstrate the method rather than to report a specific engagement. Throughout, the discipline is the same one the report argues for: numbers in preference to adjectives, sources named, and uncertainty stated plainly rather than hidden.
Updates, the full PDF, and the next forensic readiness brief — sent only to people who ask.