“We need AI”
“Engineering is really excited about this”
“We have a budget for AI/LLMs so…let’s find projects”
Sound familiar? This is not a story about poor data science. It is not a cautionary tale about technical teams failing to understand the business. It is a story about what happens when the most important question – the one that determines whether everything downstream is worth doing – gets quietly skipped in the rush and urgency to start building.
A 2025 S&P Global Market Intelligence Study1 points to an upstream failure pattern: organizations with lower AI failure rates appear to be better at selecting and prioritizing projects prior to execution. The study reports that the share of companies abandoning the majority of AI initiatives before production rose from 17% to 42% year over year, and that organizations scrapped an average of 46% of projects between proof of concept and broad adoption. It also found that lower-failure organizations use a more holistic approach to project prioritization, including compliance, risk, and data availability criteria when selecting projects. This suggests that AI and analytics failures are not only execution problems; they often begin upstream, when organizations rush initiatives, choose poorly prioritized use cases, or launch pilots without clear feasibility, risk, data availability, and value criteria. The cost shows up as abandoned proofs of concept, projects that fail to reach production, weaker KPI performance, rising concern over cost, and investment with limited measurable business impact.
The INFORMS2 Analytics Framework TM is a widely recognized, industry-leading framework used by analytics professionals and organizations to apply rigorous analytics principles into practice. It also supports professional certification in analytics and data science roles. The INFORMS Analytics Framework TM organizes analytics practices into seven domains: business problem framing, analytics problem framing, data, methodology selection, model development, deployment and lifecycle management. Domain one – business problem framing – is, arguably, the most important because it is determinative. The quality of every downstream decision, model and recommendation is bounded by the clarity of the problem definition that precedes it. Therefore, no amount of sophistication in the methods, modeling accuracy, or deployment discipline can compensate for a poorly framed business problem. The output may be precise and well-executed. But if it answers the wrong question, it will generate little to no business value.
This logic is not unique to analytics. In Good Strategy Bad Strategy3, Richard Rumelt argues that good strategy built on a weak or unexamined diagnosis risks becoming activity without strategy. Without that diagnosis, leaders risk mistaking goals, plans, or activity for strategy. The most common strategic failure is not poor execution of a clear plan. It is the confident execution of a plan whose underlying premise was never examined. The same failure mode runs through analytics at every scale, from a single dashboard project to a multi-year AI transformation. You can be excellent at building and still be building the wrong thing.
The mechanism of how we all fall into this trap lies in Tversky and Kahneman’s landmark research4 in psychology on framing effects, which showed that how a decision problem is presented can systematically alter the choices we make. In analytics then, how a problem is framed determines which solutions appear visible before any analysis begins. The frame is not neutral. It forecloses options before deliberation starts, privileges certain data sources over others, shapes what counts as a useful result and defines what success looks like. When you choose the frame implicitly — which is what happens when you skip the structured problem-definition conversation — you do not get a blank slate. You get an invisible frame that nobody agreed to, which then cannot be examined or challenged because it was never made explicit in the first place. Yet, it will quietly govern every downstream decision on the project until something expensive enough forces a reckoning.
The sequence of failures now runs faster and burns more capital. McKinsey’s 2023 Global AI Survey5 shows that generative AI adoption accelerated quickly: one-third of respondents said their organizations were already using generative AI regularly in at least one business function, and 40% of organizations using AI expected to increase AI investment because of generative AI. However, the report also showed that value capture was still uneven. Overall, AI adoption had not grown meaningfully from 2022, remaining concentrated in a small number of business functions, and only 23% of respondents said at least 5% of their organization’s EBIT was attributable to AI. McKinsey also found that lower-performing organizations were more likely to struggle with strategic issues, such as setting a clearly defined AI vision linked to business value. In 2024, Gartner reported6 that the top barrier to AI adoption was difficulty of estimating and demonstrating business value, cited by 49% of survey participants. Gartner also predicted7 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 because of poor data quality, inadequate risk controls, escalating costs, or unclear business value.
Together, these findings suggest that AI and analytics failures are often not just execution failures. They frequently begin upstream, when organizations pursue technology adoption faster than they define the decision, process, use case, risk profile, and value target the technology is meant to improve. The cost shows up as abandoned proofs of concept, projects that fail to reach production, weaker value realization, rising concern over cost, and investment with limited measurable business impact. These findings suggest that the pattern is widespread, not isolated.
The structural reason this keeps happening is worth naming directly. When budgets are approved for AI or analytics before use cases are defined, the incentive structure inside organizations can reverse. Teams are no longer tasked primarily with identifying problems worth solving; they are tasked with finding applications for already-approved spend. The question — what decision are we trying to make, and does this technology actually help us make it better? — gets skipped, not necessarily through negligence, but through the sequence of approvals. The money moves before the question is asked. By the time anyone examines the question, the budget has already begun to justify itself. This is how organizations can end up with technically impressive analytics infrastructure that generates little measurable business value. The infrastructure is real. The investment is real. The missing piece is the problem statement that would have clarified what the infrastructure was intended to support.
The solution to this expensive problem is conceptually simple: framing discipline. This is closely aligned with the INFORMS Analytics Framework TM, which begins with business problem framing and analytics problem framing before moving into data, methodology, model development, deployment, and lifecycle management. A structured framing exercise asks: Who owns the decision? What does stakeholder alignment need to look like before work proceeds? What constraints — regulatory, capital, timing, operational capacity — are non-negotiable? What would we need to observe, and by when, to call this a success? And critically, what evidence would tell us that we had made a mistake - what would the cost be of discovering we were wrong? How will we measure that this decision improved the problem we were solving?
The value of the method does not lie in technical sophistication. The value is in the sequence. The problem-definition conversation happens first, before anyone opens a data source, briefs a research vendor, or commits to scope. Only then should the organization ask whether the problem is amenable to analytics. If it is, the work can then move into data, methodology selection, model development, and deployment. If it is not amenable to analytics, then the organization avoids spending money solving the wrong problem with impressive tools.
In a 2017 Harvard Business Review article8, “Are You Solving the Right Problems?”, Thomas Wedell-Wedellsborg argued that organizations often struggle not with solving problems, but with diagnosing what the problems actually are. His research found that 85% of surveyed C-suite executives believed their organizations were poor at problem diagnosis, and 87% believed this weakness carried significant costs. The implication for analytics is direct: problem definition should not be treated as a project-management formality or an assumed starting point. It is the work that determines whether the subsequent analysis is worth doing. Organizations that build repeatable practices for surfacing, examining, and stress-testing the problem before resources are committed are better positioned to avoid solving the wrong problem with an unbounded scope. With LLM-driven solutions, that discipline matters even more because costs can compound quickly while true value remains obscured for too long. This is the role Frame Decisions9 is designed to play: not replacing analytics work or human judgment, but creating a disciplined, guided process for improving the quality of the decision before analytics work begins. Frame the problem first. Then build.
1S&P Global Market Intelligence Study. https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning
2INFORMS Analytics Framework TM: https://www.informs.org/Professional-Development/Professional-Development-Classes/INFORMS-Analytics-Framework. INFORMS is a registered mark of the Institute for Operations Research and the Management Sciences.
3Rumelt, Richard P. Good Strategy/Bad Strategy: The Difference and Why It Matters. Crown Business, 2011.
4Tversky, Amos, and Daniel Kahneman. “The Framing of Decisions and the Psychology of Choice.” Science, vol. 211, no. 4481, 1981, pp. 453–458. https://doi.org/10.1126/science.7455683.
5McKinsey & Company. (2023, August 1). The state of AI in 2023: Generative AI’s breakout year. McKinsey & Company. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-in-2023-generative-ais-breakout-year
6Gartner. (2024, May 7). Gartner survey finds generative AI is now the most frequently deployed AI solution in organizations. Gartner. https://www.gartner.com/en/newsroom/press-releases/2024-05-07-gartner-survey-finds-generative-ai-is-now-the-most-frequently-deployed-ai-solution-in-organizations
7Gartner. (2024, July 29). Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025. Gartner. https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025?
8Wedell-Wedellsborg, Thomas. “Are You Solving the Right Problems?” Harvard Business Review, Jan.–Feb. 2017, pp. 76–83. https://hbr.org/2017/01/are-you-solving-the-right-problems?
9Frame Decisions by Virtual Strategy Tech www.virtualstrategytech.com is an independent platform and is not affiliated with, sponsored by, or endorsed by INFORMS. INFORMS materials are referenced or reproduced with permission where applicable.