You can put a clock on the way an AI discussion will go in any boardroom. The pitch is always the same: the technology has come of age, you have the productivity numbers to back it up, and your rivals are putting money into it. So leadership gives the green light, a team is put together and pilots get under way. And then, without any fanfare, the project comes to a standstill. There is no dramatic implosion, just a steady build-up of inconclusive results and shifting priorities and dependencies that can no longer be left hanging until there is no case for going on. You write off the investment and start the whole conversation over in the new year.
These days, that is the norm by the numbers. A 2024 study from the RAND Corporation, which involved sitting down with 65 top data scientists and engineers, puts the failure rate of AI at more than 80% when it comes to getting to production – double what you would see with non-AI tech. Gartner was sounding the alarm as early as July of last year, forecasting that poor data and unclear value would lead to the scrapping of at least 30% of generative AI efforts by 2025. They were right, and it has been worse. S&P Global Market Intelligence surveyed a thousand or so enterprises in North America and Europe and found the attrition had become material: 42% of them had walked away from most of their AI work in 2025, compared with only 17% the year before.
Most in the industry will tell you it is the technical side to blame – the models aren’t up to snuff, the compute is too expensive, and integration is a headache. But the data does not bear that out. It is rarely the algorithm itself. What decides if an AI project makes it from the lab to production are the four structural things around it: the governance, the team, how good the data is and whether you have a true product-market fit. We will look at each of those here.
Accountability Without Infrastructure: The Governance Gap
If there is one precondition for rolling out AI that is starved of investment, it is governance. And the fallout from that neglect is out of all proportion. Take an AI system that has put out a discriminatory result, made a factually wrong recommendation or compromised privacy. An organisation’s problem then goes beyond simply patching the model; they have to answer who is on the hook for it, what was done to stop it in the first place and what those affected are owed by way of redress. For those unable to give clear answers, the issue is not merely technical. They are looking at legal exposure and regulatory heat, to say nothing of the loss of stakeholder trust you need before you can deploy any more AI.
The chasm in governance is plain to see. You only have to look at the numbers from IBM’s 2025 AI at the Core research. Some 74% of organisations say their frameworks for managing model, third-party and technology risk offer little more than moderate or limited coverage. A mere 23.8% have something in place that deals with these risks in any meaningful way. That is a telling statistic when you consider the same report puts AI integration in at least one part of the business at 72%, a jump from 55% last year. In short, companies are moving ahead with deployment well ahead of their ability to govern it, and the divide is only getting larger.
Let’s have a look at the numbers to see the disconnect. A 2024 study from McKinsey put it plainly: while 42% of enterprises rolling out generative AI name “content integrity and governance” as a top operational risk, they have yet to put in place the structures to actually govern it. Then there is the matter of cost. According to IBM’s 2025 Cost of a Data Breach Report, 13% of organisations have seen an AI-related breach, and in 97% of those cases it was because access controls were not in order. With the tab for an average breach coming to $4.44 million, an ungoverned AI system is no longer just a compliance headache but a very real financial liability for the enterprise.
The issue of ownership is just as pressing. The IAPP Governance Survey for 2024 shows that a mere 28% of organisations have formal oversight roles for AI. What you get instead is accountability spread so thin across legal, IT, compliance and the business side that you end up with collective responsibility and individual inaction. If a model commits a serious error, there is no one with the mandate to fix it or the authority to pull the plug on deployment. Gartner would have you believe this is why 80% of data and analytics governance efforts will be dead in the water by 2027; most firms treat governance as a reactive gatekeeper to be invoked when things go wrong, not as an infrastructure worth investing in.
But the evidence to the contrary is telling. Take IBM’s internal programme, which has overseen more than 1,000 production models and cut data clearance time by 58%. Or consider what McKinsey reported in 2024: firms with a mature approach to governance are able to scale their AI 2.5 times quicker and for 30% less than their peers. Done right, governance is the infrastructure that makes for defensible, repeatable deployment at speed. It is hardly an impediment to innovation.
The Challenge in Front of the Model: No Product/Market Fit
The second structural failure mode is conceptually simple and organisationally widespread. Organisations develop AI solutions before they are able to confirm that a real business problem exists in the particular form that the solution addresses. This is the AI equivalent of the classic product development failure: delivering something technically impressive that no one really needed.
This pattern is well documented in the failure literature. A Gallup poll in late 2024 found that just 15% of US employees said their workplaces had shared a clear AI strategy. But McKinsey reported that 92% of executives planned to spend more on AI in three years. Investment is outpacing strategic clarity by a factor of six to one. A survey of 250 AI leaders revealed that 67% of business leaders expect AI to transform their organisations by 2025, but only 36% have a clear vision for how to implement it. Most AI projects actually die in the gap between aspiration for transformation and clarity of implementation.
The problem-solution mismatch exists in various forms. In some organisations the use case is valid, but the problem has been scoped at the wrong level of specificity: “improve customer service” is not an AI product spec, and the lack of crisp success criteria means it is impossible to evaluate whether any solution actually works. For others, the use case is a workflow that employees have already informally optimised, meaning the efficiency gains promised by AI are marginal rather than transformative and therefore cannot be justified once the novelty of the pilot has worn off. Most destructively, some initiatives are driven by competitive pressure to appear that they are deploying AI rather than a documented workflow problem that AI specifically solves better than alternative solutions.
The clearest empirical signal on this dimension is McKinsey’s 2025 State of AI survey, which finds that organisations that have seen significant financial benefits from AI are twice as likely to have redesigned end-to-end workflows before choosing modelling techniques. The direction of causality is important. Successful organisations start with a workflow problem, map the human decision points of that workflow, identify where AI can either replace or augment human judgement at scale, and then decide on an approach. Failed organisations start with a model capability, whatever the latest vendor offering is selling, and look for a workflow problem to fit. A late-2024 report by Boston Consulting Group found that 74% of companies had yet to show tangible value from their AI initiatives, and the best predictor of that gap wasn’t model quality but the degree to which the initiative was connected to a redesigned operational process.
The consequences of chasing AI with no validated product-market fit are structural. Pilots that cannot demonstrate improvement in outcomes against a pre-defined baseline cannot make the business case for production infrastructure investment. Without a business case, the initiative dies in prototype, not because it failed technically, but because success was never defined in terms the business could measure.
The Data Foundation Problem: Ecosystems Not Designed for AI
If governance is the most neglected precondition and product-market fit is the most frequently neglected validation step, data quality is the most technically destructive failure mode, the one that destroys initiatives that get everything else right. The causal mechanism is unforgiving: AI models learn patterns in data. If the data is incomplete, inconsistent, or historically biassed, the model will replicate those patterns at scal - with the speed and confidence that make AI valuable in the first place.
The size of the problem is consistently recorded by the major research bodies. In a Q3 2024 survey of 248 data management leaders, Gartner found that 63% of organisations either don’t have or are unsure whether they have the right data management practices for AI, and Gartner now predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. According to Informatica’s CDO Insights 2025 survey, data quality and readiness were the number-one impediment to AI success, with 43% of respondents citing it – tied with technical maturity and ahead of skills shortages at 35%. On top of this, Gartner says that poor data quality costs organisations an average of 15% of revenue per year, regardless of AI programmes. That baseline cost is increased for AI initiatives because model inaccuracy is proportional to deployment volume.
Most organisations don’t know the difference between traditional and AI-ready data management, and that lack of clarity is a major failure driver in itself. Traditional enterprise data architectures were built for structured reporting: data warehouse schemas, ETL pipelines, and batch processing optimised for periodic analytical queries. AI applications, especially large language models and real-time inference systems, need low-latency access to high-volume, diverse, frequently updated data from heterogeneous sources, along with metadata lineage, bias documentation and governance provenance – things that most legacy data systems don’t capture. More than 75% of organisations see AI-ready data as one of their top five investment areas over the next two to three years, according to Gartner’s 2024 Evolution of Data Management survey - which means most have yet to make that investment.
The fragmentation problem exacerbates the quality problem. In most large enterprises, relevant data is scattered across business units, often in incompatible formats, under different policies, and accessible by different teams under different permission schemes. An AI application that needs to synthesise customer transaction history, support ticket data, product usage telemetry and external market data to produce a meaningful recommendation faces not one data problem but four – each with its own cleaning requirements, access controls and update cadence. Research from Gartner, Deloitte and McKinsey consistently shows that over 70% of AI project failures are tied directly to data issues, not algorithmic issues. This was demonstrated at scale in the widely documented Watson Health experience at IBM, where a programme aimed at providing oncology treatment recommendations was trained on hypothetical clinical scenarios rather than real-world patient data, resulting in recommendations that clinicians found inconsistent with clinical practice - a failure of data provenance, not a failure of modelling.
The main point for technology leaders is that data readiness assessment should happen before model selection and infrastructure investments. Organisations that treat data preparation as a downstream task, something to be handled after the model architecture has been chosen, inevitably find that the data problems are foundational enough that the architectural decisions need to be re-evaluated. Successful programmes allocate 50-70% of time and budget to data readiness before commencing model development, according to the McKinsey 2025 AI survey.
The Organisational Dimension: Silos, Ownership and the Talent Deficit
The fourth mode of structural failure is organisational and in many respects the most intractable. Deploying AI is a multi-disciplinary problem: it requires domain expertise to frame the problem correctly, data engineering to build robust pipelines, machine learning expertise to select and tune the right models, software engineering to integrate AI into production systems, product management to help ensure the system actually meets user needs, and legal and compliance expertise to help ensure it does so within regulatory boundaries. These capabilities are seldom found together, and in most large organisations, they reside in separate functions with separate reporting lines, separate budgets and separate success metrics.
The dimension of talent shortage is well documented, and it’s getting worse. In fact, CompTIA’s 2024 Tech Workforce Report found that 87% of organisations find it difficult to hire AI developers, with average time-to-fill hitting 142 days, a timespan that doesn’t match most project schedules. A survey of 250 AI leaders found that 94% of technology leaders cite talent shortages as the biggest hurdle to AI innovation, with 85% saying their initiatives have been delayed as a direct result. IDC estimates the global AI skills gap could cost the economy as much as $5.5 trillion in product delays, quality degradation, missed revenue and reduced competitiveness by 2026. Big tech companies, including the FAANG companies, alone hire 70% of top AI talent directly out of universities, McKinsey Global Institute data suggest, leaving enterprise organisations to compete for a shrinking supply of leftovers.
The silo problem is also organisationally structural beyond the talent supply constraint. According to McKinsey’s 2025 State of AI report, 88% of organisations are using AI in at least one business function, but only about one-third have scaled AI successfully across the enterprise. Many of the distinctions between localised adoption and enterprise-scale deployment can be attributed to organisational structure: business units pursue AI in isolation, data is siloed within functional teams, and there’s no shared infrastructure for the tooling, evaluation, and deployment systems that production AI needs. When an AI prototype built by a data science team needs to move to production, it must cross organisational boundaries – to software engineering for deployment, into legal for review, and into the product team for integration, and each crossing introduces friction that many initiatives do not survive.
The silo structure is directly a consequence of ownership ambiguity. What happens if a deployed AI system produces a harmful output? Who is liable? Who can stop the rollout? It's who owns the continuous monitoring and retraining cycle that is needed by production AI systems. In organisations without clear answers, these responsibilities fall through the cracks between teams. The result is a production AI system that’s technically deployed, but operationally unmanaged, drifting in quality, accumulating technical debt and eroding user trust until it’s quietly deprecated.
The difference with high-performing organisations is telling. According to a.team, which surveyed 250 AI leaders, blended teams – structured, cross-functional teams that combine specialised freelancers with full-time employees – are twice as likely to succeed in implementing AI as those that rely solely on internal or homogeneous team structures. BMW’s rollout in late 2024 of its AIconic multi-agent AI system was a replicable model: digital training for all employee levels; AI innovation spaces with technical and domain experts in shared physical proximity; and obvious delineation of roles between AI development, deployment, and operational ownership. This is not a technological advantage; this is a structural advantage.
What Makes the 20% Different: Organisational Patterns of Successful Deployments
There are certain characteristics of organisations that are successful at taking AI from experiment to production that aren’t necessarily technological. They begin with workflow redesign, not model selection. They create governance infrastructure before deployment pressure arises. They think about data readiness as an architectural requirement, not a cleanup job. And they form cross-functional teams with clear ownership boundaries instead of putting AI in the data science group and hoping engineering will pick it up later.
According to McKinsey’s 2026 AI Trust Maturity survey, organisations with well-defined accountability for responsible AI outperform those with no clear ownership structures in terms of maturity scores. Furthermore, heavy investors in responsible AI governance achieve significantly better business outcomes, including measurable EBIT impact, than governance laggards. The survey was conducted among approximately 500 organisations with direct AI governance responsibility. The average AI trust maturity is expected to increase to 2.3 in 2026 from 2.0 in 2025, but as important, only about one third of organisations have achieved a governance maturity level of three or higher, which implies that two thirds remain in governance configurations that leave them vulnerable to the failure modes outlined above.
The story is in the speed at which AI leaders deploy production systems – organisations in the top quartile of governance and data maturity. 93% of organisations that have deployed AI in production say custom solutions built for specific workflows generate more value than off-the-shelf tools and the more experience they have with deployment, the more strongly they believe this. The learning curve is real: organisations that have already been through the governance, data and team challenges are in a much better place to go through the process again. Those organisations that remain mired in pilots are also, in a meaningful sense, accruing the organisational learning needed to eventually succeed – but the financial cost of prolonged experimentation, measured against competitors already realising deployment value, is significant.
What Technology Leaders Need to Do Next
The findings across these four dimensions of failure converge on concrete organisational investments that distinguish deployable AI programmes from permanent pilots. They are not vendor recommendations or research agenda entries. These are structural choices that are directly in the control of technology leaders.
Set up governance before you feel the pressure to deploy, not as a reaction to it. Before any system goes into production, define ownership roles for model accountability, monitoring and incident response. The deployment spec should include expected model behaviour, acceptable ranges of output, and escalation paths. McKinsey’s research is clear: good governance helps organisations move faster, not slower, and spend less, not more. Investment in governance infrastructure is not a tax on innovation; it is the enabling condition for a sustainable speed of deployment.
Validate the problem, and then select the technology. Each AI project should begin with a documented workflow analysis that maps out the current steps of human decision-making, quantifies the time and error rate of the current process, and details the success criteria against which the AI solution will be evaluated. Only 36% of organisations have a clear vision for implementation – that statistic should be an organisational target to improve before the next pilot launch, not a benchmark to accept.
Run a data readiness assessment prior to building a model. Identify the data sources required for the intended application and assess their quality, accessibility, update cadence and governance provenance. Quantify the delta between current data infrastructure and AI-ready data infrastructure and explicitly budget for that delta. But that Gartner prediction that 60% of AI projects that don’t have AI-ready data will fail is a prediction about organisations that don’t do this, not a law of nature. It is avoidable.
Build the team first, then the model. Cross-functional ownership is critical for production AI, with data engineering, ML engineering, software engineering, product management, domain expertise and legal/compliance review all having defined roles and handoff protocols. Organisations that successfully implement AI in production using blended, cross-functional teams have twice the likelihood to succeed. Designate a responsible product owner (not a technical lead) for each AI project. Accountability without authority is inertia. The product owner must have the authority to stop deployment if success criteria are not met.
Embed the feedback loop in the deployment specification. AI systems require active maintenance to prevent decay. User feedback mechanisms, output monitoring, and retraining pipelines are not post-deployment improvements. They are deployment requirements. IBM’s management of more than 1,000 production AI models demonstrates that operational discipline at scale is possible, and organisations developing these feedback capabilities are building structural advantages that compound over time. If there is no feedback loop, AI tools can’t improve after launch, and user disengagement is itself a known cause of AI project abandonment.
The AI deployment gap is not a tech problem, with a better model to come. This is a problem of an organisational and engineering discipline which the current tools and frameworks are fully capable of solving. The organisations that will close the gap will not be those with the biggest AI budgets or access to the latest foundation models. They are the ones who will bring to AI the same structural rigour – clear ownership, validated requirements, quality-controlled inputs, and operational monitoring – that has always separated reliable software from failed software. All organisations have access to that discipline. Most just aren’t choosing to apply it yet.
Note: This article is the editorial opinion of ExpertStack and is based on publicly available research from RAND Corporation, Gartner, McKinsey & Company, IBM, S&P Global Market Intelligence, MIT NANDA Initiative, IDC, Informatica CDO Insights, CompTIA, BCG, and the IAPP Governance Survey as cited in the text. ExpertStack is editorially independent and impartial. This content is not provided by any vendor.