How can you break out of the AI ROI trap?

As our survey results indicate, organizations are taking a wide array of approaches to decisions around AI architecture and delivery, with most choosing to work with external partners. Only 9% are building their own large language models (LLMs), while 18% are co-developing with third-party providers. By contrast, the majority are relying on external models — with 41% using closed models, 27% open and 26% hybrid approaches. Also, customization is largely incremental: 62% use standard external models, while 51% add retrieval-augmented generation (RAG) and 46% fine-tune with proprietary data. The latter two approaches are not mutually exclusive, so some organizations use both.

The high degree of pragmatism in companies’ AI architecture choices flows through to their vendor selection and solution design approaches, as shown in the chart below. On build-versus-buy, the most frequent choice is using a hybrid case-by-case strategy to balance customization with expediency. Out-of-the-box solutions come next, allowing for faster deployment and immediate functionality even if this results in higher upfront costs. Fewer organizations are relying on external providers or building solution-specific AI capabilities in-house. When engaging vendors, 61% of organizations prioritize best-of-breed over best-of-suite approaches. 

Organizations also vary in how they balance rapid deployment of point solutions against comprehensive process redesign, typically aligning their approach to intended outcomes. Quick-to-implement solutions — such as AI-powered customer support systems — still remain prevalent compared to larger-scale objectives like integrating AI across all aspects of the finance function. This preference for rapid-deployment options is largely driven by immediate benefits such as implementation speed, cost efficiency, scalability and, to a lesser extent, ease of integration with existing processes. However, process redesign tends to outperform quick solutions in delivering greater flexibility and control, as well as strategic alignment.

The overall message is clear. In the absence of a clear enterprise vision for AI, deployment choices default toward speed, cost and risk containment — effectively reinforcing pilots and point solutions despite enterprise goals of process transformation and strategic alignment. Those pragmatic choices are rational — but they create measurement and governance gaps that prevent enterprise ROI.

As companies navigate the difficult transition from AI pilots to enterprise-wide transformation, further factors limiting ROI on AI, include the fact that measurement of the benefits realized is largely qualitative rather than quantitative, and that the KPIs being used are often underspecified.

Some 61% of respondents believe that implementation of AI is creating more value than they are able to quantify accurately. And while companies point to gains in productivity and transparency, they lack clear KPIs to measure impacts at an enterprise level. This appears to reflect a lack of upfront definition of the desired outcomes and metrics.

These issues are compounded by governance that is both inconsistently deployed and still in its early stage of development. As the charts below show, companies are taking diverse approaches to the ownership of AI governance — with the most popular options being ownership by the CTO and distribution across various functions — and are also applying various strategies to embed AI leadership, with a fairly even split between a centralized approach and locating it at a functional level. Meanwhile, the vast majority of respondents have either established a cross-business council or are planning to do so.

The central issue is not the structure of governance, but that governance processes are applied unevenly. In a NAVI world — nonlinear, accelerated, volatile, interconnected — this inconsistency dampens ROI and heightens risk.

The same pattern appears in data management. Nearly half (43%) of respondents use hybrid data models that split control between centralized and decentralized teams, and another 20% use federated approaches where business units manage their own data under enterprise-wide standards. To manage these distributed efforts, more than 50% of respondents report formal processes to assess architecture, data, workforce, customer and business readiness for AI, yet application lags: only 38% say they apply these processes regularly for architecture readiness and just 26% for data readiness.

The takeaway is straightforward: there isn’t a single “preferred” governance approach, but there is a persistent execution gap. Even where governance processes exist, they’re not being used consistently. Closing that gap — by making roles, standards and readiness checks repeatable in practice — is what will move AI beyond pilots and make ROI demonstrable. Agentic AI amplifies the cost of that gap because autonomy increases both scale and consequence.

Agentic AI is raising the stakes

As organizations’ AI journey advances to investments in agentic AI, the effect can be to repeat the excessive focus on pilots experienced with GenAI — but at larger scale and risk.

For organizations, the step forward from GenAI to agentic AI compounds both opportunity and complexity. In immature governance environments, agentic deployments risk replicating GenAI’s pilot trap at larger scale — introducing more moving parts and more risk. As with GenAI, some stakeholders express confidence that agents are creating more value than they can capture with existing metrics, a view with which 33% say they agree. Also, respondents’ expectations for how agentic AI adoption will impact their current AI capabilities reveal diverging views – ranging from 40% expecting them to coexist and complement each other, to 12% anticipating minimal impact.

GenAI has the potential to be transformative, but realizing its value requires discipline. Success depends on treating AI as a strategic investment, anchored in readiness, rigor and measurable results. Ultimately, escaping the AI ROI trap requires disciplined capital allocation, strong foundations and portfolio rigor. GenAI is increasingly a competitive necessity, but the risk lies in investing without discipline, patience and a clear understanding of time-to-value.

A portfolio mindset helps leaders balance foundational “table stakes” investments with longer-term bets that may take time to mature but are critical for future growth, setting realistic expectations for value realization. Where appropriate, apply a private equity (PE)-like approach: implement and gear investments toward steady, predictable returns. For transformational impact, adopt a venture capital (VC)-like mindset: diversify across many small bets, accept that some will fail, and pursue those that succeed. The key lies in adopting a dual mindset — optimizing for predictability where needed and embracing risk where transformation is possible.

At the same time, successful AI deployment treats governance, risk and cybersecurity as enablers, not brakes. The key is to evolve these principles from static standards to dynamic, embedded pillars, ensuring they are applied consistently across the organization. Organizations that get this right will see stronger returns, not by slowing down, but by making innovation nimble, sustainable and repeatable. Early GenAI programs often favored tight central control; the next phase will reward disciplined scale with consistent guardrails.

A clear set of actions can increase and clarify AI ROI, supporting enterprise-wide scaling and value growth. Based on EY research and experience, five actions can help organizations boost demonstrable value from AI and set the scene for scaling: