Why 97% of Companies Have AI Projects and Only 5% Have Data Ready to Use Them
There is a statistic that should bring any boardroom discussion about artificial intelligence to a halt: according to a Dun & Bradstreet survey of 10,000 companies conducted in 2026, 97% declare that they have active AI initiatives, while only 5% consider that their data is truly ready to sustain them. That gap is not a minor technical detail. It is the distance between investing in infrastructure and having something that works reliably in production.
What that number describes is a pattern well known to those who have observed how technology decisions are made in large organizations: first the pilot gets approved, then the problem that justifies having approved it is sought after the fact. The demonstration impresses. The room applauds. The project receives a budget. And somewhere between that moment and day-to-day operations, something breaks down without anyone having formally declared failure.
BCG documented this with precision: only 5% of companies obtain substantial value from AI, while 60% report no material impact. McKinsey, for its part, found that more than 80% of respondents were not seeing any tangible effect on EBIT from generative artificial intelligence, even as declared adoption continued to grow. Those figures are not a condemnation of the technology. They are a snapshot of how investment is being managed.
The Illusion of the Perpetual Pilot
There is a silent form of organizational failure that does not appear in balance sheets or generate press releases: the pilot that never dies. It installs itself under the name of "innovation," chronically consumes technical and human resources, produces reasonably attractive internal presentations, and never manages to become something that changes a single line on the income statement. Organizations with greater maturity in digital transformation have spent years learning that this dynamic is not accidental, but structural.
The problem begins at the origin of the project. When an AI initiative is approved because "the use case is interesting" or because a vendor gave a convincing demonstration, it lacks from the very beginning something that every investment program should have: a measurable baseline, an owner of the outcome, and an exit criterion if value does not materialize. Without those three elements, the pilot has no way to die with dignity or to scale with traceability.
BCG identified that companies with the best AI results prioritize between three and four use cases on average, compared to six or seven at organizations with the worst returns. That difference does not arise from the available budget or the size of the technical team. It arises from the willingness to reject initiatives that cannot demonstrate strategic alignment and economic viability from the moment they request funding. Focus is not an abstract managerial virtue; in this context it is the only condition that makes scaling viable.
What the BCG and McKinsey data reveal in combination is that the majority of organizations are in a phase that might be called AI theater: high visible activity, low operational transformation. Press releases talk about adoption. Internal metrics tell a different story.
The Problem Is Not in the Model, It Is Below the Model
There is an understandable tendency to analyze AI performance from the angle of the model: which architecture was used, which vendor, which version of the system. That analysis has utility in research contexts, but in most business environments the bottleneck is not in the model. It is in what the model needs to function reliably: clean data, shared definitions, redesigned workflows, and clear ownership over the decisions the system is meant to support.
The Dun & Bradstreet survey cited earlier puts it in terms that do not admit much alternative interpretation: if almost no company considers its data to be ready, then the massive problem is not one of technological experimentation but of fundamentals. An AI that receives fragmented data, without a single source of truth, with business rules buried in spreadsheets and exception processes that nobody documented, does not generate more reliable recommendations than the system it aims to improve. In many cases, it simply accelerates existing errors.
PwC identified this pattern from another angle: the most solid results come when companies redesign workflows rather than layering AI on top of legacy processes. That distinction matters economically. Adding an artificial intelligence component to an inefficient process can make that process faster. But it does not change the cost logic or the structure of the operation. The economics of work remain the same, only executed at greater speed.
The case of highly demanding regulatory environments is especially clear. Finance, regulatory compliance, legal review, supply chain: these are contexts where two different responses to the same query are not a sign of system flexibility, but a control problem. Reliability in production has a different standard from that of experimentation. And that difference is what separates systems that get adopted from those that are quietly abandoned after the pilot.
When 70% of the Value Comes from Factors That Do Not Appear on the Technology Roadmap
BCG documented something that tends to make technology teams uncomfortable: in AI-driven transformations that generated real results, 70% of the value came from actions related to people, not technology. That includes redefinition of roles, changes in incentives, active management of the adoption process, and the building of capabilities within the teams that were supposed to use the systems in production.
That finding should not be read as an argument against technical investment. It should be read as a map of where the real blockage typically lies. A language model can process thousands of contracts per hour; but if the legal team does not trust its outputs, if the area's incentives have not changed, if nobody redefined what it means to "review a contract" when there is a system that performs the first pass, adoption does not occur in a sustained manner. The system exists. The value does not.
The Global 1000 companies that are reporting measurable impacts share some operational characteristics: they have redesigned specific processes before implementing the systems, they have established metrics against documented baselines, and they have assigned ownership of results to people with real accountability over those numbers. In some documented cases, the results are material: increases on the order of 30% in manufacturing efficiency, 80% reductions in document analysis times, improvements of 1.7 times in sales conversion rates. Those numbers do not come from superior models. They come from superior integrations.
The difference between a company that uses AI and a company that operates with AI does not lie in the model provider or the size of the innovation budget. It lies in whether the organization has been able to connect the system's output to a concrete decision, within a redesigned process, with someone responsible for measuring whether that moves the number it is supposed to move.
The Real Displacement That These Numbers Reveal
The current phase of the business cycle of artificial intelligence is not being defined by advances in foundation models. It is being defined by the capacity of organizations to move from the legitimacy of the experiment to the demand for results. And that transition is not yet the majority condition.
What the data from BCG, McKinsey, PwC, and Dun & Bradstreet describe collectively is a market with an asymmetric distribution: a small minority of companies is generating measurable and compounding value with AI, while a broader majority continues to accumulate projects that do not touch the income statement. That gap does not close with more technology. It closes with portfolio discipline, with data fundamentals that are today clearly missing in most of the market, and with an organizational willingness to accept that real adoption requires redesign, not superimposition.
The displacement that is occurring, though still incomplete, points in a precise direction: AI is ceasing to be a signal of modernity and is becoming a demand for evidence. Organizations that cannot respond with numbers to the question of what changed operationally since they implemented their systems will face growing pressure — first internally, then from their boards of directors and their investors. The capital that previously flowed toward the experiment will begin migrating to where the experiment proved to be something more.
