Close the ROI gap when scaling AI

By now, executives are clear in their directive: “Figure out how to use AI.” The urgency is significant, the timelines are challenging, and expectations are exceptionally high. And yet, while experimentation is increasing, meaningful business impact remains elusive for most—and haphazard prototyping is setting organizations up for expensive disappointment.

Harvard Business Review reports that only 26% of companies have developed working AI products, and only 4% have achieved significant returns on their investments.¹ And Gartner predicts that at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025 due to poor data quality, inadequate risk controls, escalating costs, or unclear business value.²

If you’re feeling the weight of these statistics and the gap between AI expectations and reality, you’re not alone. A pattern is emerging with these common elements:

• Leaders who know they need AI but have trouble articulating what success looks like
• Organizations with data scattered across siloed systems
• Teams who’ve successfully run proof-of-concepts (POCs) but can't scale them into operational reality

This results in AI initiatives that burn through budgets, exhaust teams, and leave executives questioning whether the projected benefits of their technology investment were overstated.

As many are realizing, significantly more work is needed to get AI right than initially expected—and they face the real risk of getting it wrong. Understanding these challenges is the first step toward building AI initiatives that deliver lasting business value.

The hidden complexity behind AI success

Forbes states that 75% of executives consider AI/GenAI a top strategic priority, but only 25% notice significant value from it.³ That gap represents misallocated resources and hours of organizational frustration. The most detrimental failures occur when organizations confuse activity for impact. They often pursue trendy AI tools and accumulate countless POCs but minimal viable products. Or they push pilot projects that never advance to production and commend technical achievements that fail to yield results.

These statistics don’t mean that prototyping is inherently problematic. When structured correctly, these validation methods and prototypes build invaluable organizational experience and technical understanding. The challenge lies in how prototypes are conceived and executed. Successful prototypes are designed to start small, build for change, and learn fast while maintaining clear connections to business outcomes. Problematic prototypes become expensive science experiments with no clear path to operational value.

The disparity between successful experimentation and operational impact frequently hinges on four aspects that determine whether an AI initiative—be it an exploratory POC or full-scale deployment—will generate operational value, not just technical validation.

The four aspects that determine AI readiness

Rather than asking “How can we use AI?”, start with your strategy. These aren’t technical assessments—they’re business readiness evaluations that determine whether your AI investment will generate sustainable returns or become a costly learning experience. These questions will help frame your AI strategy:

1. Problem-solution alignment: What specific problem are you solving?

A common pattern for many organizations is to start with solutions that look for problems. “We want to use AI for customer service” isn’t a strategy—it’s a technology product in search of a business case. But asking, “How can AI be used to analyze our customer data to offer tailored financial products that yield increased profitability and better customer retention?” represents a problem you can solve, measure, and scale.

Gartner's research with early AI adopters shows promise when objectives are clear, with respondents reporting on average a 15.8% revenue increase, 15.2% cost savings, and 22.6% productivity improvement. But these gains came from organizations that started with specific, measurable business problems rather than general AI aspirations. The best AI implementations begin with clearly articulated challenges that have quantifiable success metrics. They identify specific decisions that need to be made faster, more accurately, or on a greater scale. Without this clarity, you'll build technically impressive solutions that don’t move business needles.

Problem-solution alignment helps verify every AI initiative—from initial pilot to full deployment—and maintain clear connections between technical capabilities and business outcomes. Whether you’re exploring AI possibilities through pilots or implementing proven solutions, this alignment drives measurable results.

Actions on alignment:

• Translate your AI ambition into measurable objectives with clear success criteria.
• Align your initiatives using workshops and hands-on POC development to generate insights that are aligned to outcomes, not activity

2. Data architecture readiness: Do you have the right data strategy for your AI approach?

Many organizations conflate having data with having AI-ready data. A Salesforce survey showed that while 76% of business leaders say the rise in AI increases their need to be data-driven, only 36% say they are confident in the accuracy of their company’s data.⁴ Your data requirements vary significantly based on your AI approach. While traditional machine learning demands extensive, clean training datasets, generative AI (GenAI) applications require different foundations. Fine-tuning models necessitates curated domain data. Retrieval Augmented Generation (RAG) systems require structured knowledge repositories. And prompt engineering needs representative examples with clear lineage.

The build-versus-buy decision reshapes your data strategy entirely. For GenAI, most organizations will leverage existing foundation models rather than training from scratch, but this doesn’t eliminate data challenges—it changes them. Fine-tuning requires carefully labeled, domain-specific data. RAG implementations need comprehensive, searchable knowledge repositories. Even simple prompt engineering benefits from representative data samples to optimize performance.

A data readiness assessment can evaluate data quality and accessibility, governance policies, and whether you have sufficient volume to train reliable models. If your data team needs more than a week to compile appropriate datasets for your chosen AI approach—whether for fine-tuning, RAG implementation, or model evaluation—your data foundation requires strengthening.

Actions on data readiness:

• Assess your data readiness and infrastructure.
• Establish your data foundation using data assessments, pragmatic data infrastructure roadmaps, or build-versus-buy decision frameworks

3. Infrastructure scalability: Can your infrastructure support production AI?

AI isn't just software—it’s a technology that demands robust infrastructure, monitoring, and governance frameworks. Stanford’s State of AI report revealed that, according to one index, AI-related harm incidents rose to 233 in 2024—a record high and a 56.4% increase over 2023, highlighting the operational complexities of production AI systems.⁵ Many organizations successfully build models in development environments only to discover that they lack the production infrastructure or monitoring to deploy them reliably and at scale.

Infrastructure considerations go beyond computing power. They include model deployment pipelines, monitoring and alerting systems, data security frameworks, complex regulatory requirements and compliance protocols, and integration capabilities with existing business systems. They also include the often-overlooked human infrastructure: teams with the skills to maintain, monitor, and iterate on AI systems. The infrastructure question becomes even more complex when you consider the difference between pilot and production environments. A model that works perfectly in a controlled pilot environment might fail catastrophically when exposed to real-world data variability and volume.

Gartner research indicates that production AI deployment costs range from $750,000 for basic RAG applications to $20 million for customized, domain-specific LLMs.² Understanding these requirements during prototype design ensures that experiments can build toward scalable solutions instead of standalone demonstrations.

Actions on infrastructure:

• Design and implement an AI infrastructure that scales from successful pilots to enterprise wide deployments.
• Make your first AI implementation the foundation for future initiatives and not an isolated experiment.

4. Organizational readiness: Can you act on AI insights?

This is perhaps the most overlooked question, but it’s often the most critical. AI systems generate value through organizational adoption, not just technical performance. Models may provide perfect predictions, but if your organization cannot operationally respond to those insights, you've built an expensive dashboard that creates frustration rather than value.

Acting on AI insights requires more than technical integration; it requires organizational change management. That means retraining staff, updating processes, and often restructuring workflows. Getting stakeholder buy-in from teams whose daily work will be impacted by AI recommendations is essential.

For financial institutions, operational readiness is the critical link between AI-driven insights and measurable outcomes. For banks, insurers, and asset managers in particular, this means ensuring that once AI models flag a compliance risk, identify a fraud pattern, or suggest a personalized product, the organization is equipped to seamlessly act on the resulting insights. This requires aligning operations with AI-enabled workflows—whether that’s automating alerts in risk management, integrating AI outputs into CRM systems, or enabling advisors to act on real-time recommendations.

Without this alignment, institutions risk underutilizing powerful AI tools, leading to inefficiencies, regulatory exposure, and missed revenue opportunities. True operational readiness ensures that AI becomes a trusted, embedded part of decision-making across the enterprise.

Actions on operational readiness:

• Support cross-functional alignment and change management to drive AI-enabled operational improvements and avoid underutilized systems.
• Integrate your AI solutions into operational workflows to enhance usability and impact.

For industry specific checklists and concrete actions to help you accelerate AI adoption with clarity and confidence, download our AI acceleration frameworks.

From readiness assessment to strategic advantage

Organizations that systematically assess implementation readiness achieve something powerful. With every AI initiative, whether pilot or production deployment, they build toward operational advantage. Their exploratory pilots generate insights into both technical performance and implementation requirements. These focused implementations leverage pilot learnings to accelerate value realization.

Instead of using pilots as technology experiments, market leaders use them as implementation preparation. They structure exploratory initiatives to build organizational readiness alongside technical capability, creating systematic pathways from experimentation to operational advantage.

Implementation readiness isn’t about avoiding experimentation. It’s about having the patience to structure AI initiatives in a way that maximizes both learning and operational potential. Organizations that master this balance don’t just deploy AI—they use it as a strategic accelerator.

Knowing when to act

Not every “pause” on any of these questions means abandoning AI entirely. Often it means having the wisdom to invest in the prerequisites that position your organization for sustainable AI success. Seasoned experts offer the following tips to consider as you evaluate your true readiness.

Proceed with AI initiatives when you have:

• Clearly defined, measurable problems
• Accessible, quality data
• Infrastructure that can support production deployment
• Organizational alignment on implementation requirements

Pause and reassess when:

• Your primary motivation is competitive pressure rather than business needs
• Your data exists but isn’t accessible or reliable
• You lack infrastructure for production deployment
• Key stakeholders aren’t aligned on success criteria and/or implementation requirements

Build your AI foundation when:

• You have strategic AI vision but lack tactical readiness
• You can start to identify some specific use cases
• You’re ready to explore robust evaluation frameworks for iterative improvement
• You’re capable of developing adaptable systems to facilitate rapid learning

Organizations that thoughtfully assess their AI readiness before diving into implementation are more likely to achieve sustainable competitive advantages and better long-term ROI. Rather than accumulating failed pilots, they build capabilities that compound over time. Their first successful AI implementation becomes the template for the second, and their infrastructure investments pay dividends across multiple use cases.

Moving forward with confidence

The AI revolution is real—and so is the opportunity for organizations that approach it strategically rather than reactively. The pressure to “do something with AI” is understandable, but the smartest first step is an honest assessment rather than immediate action. Understanding where you stand on problem definition, data readiness, infrastructure capability, and organizational alignment will determine whether your AI initiatives create lasting value or expensive lessons.