GenAI is the fastest adopted technology in history, but speed of uptake has not necessarily translated into durable enterprise value. Despite unprecedented investment, many organizations struggle to convert experimentation into lasting operational impact.
At the start of the AI surge, Gartner projected that 30% of generative AI initiatives would fail to deliver lasting impact, while 40% of agentic AI projects could be cancelled by the end of 2027. IDC and Lenovo highlight the same pattern: of every 33 proofs-of-concept launched, only four make it into production. The rest may stall, become abandoned, or fail to scale—leaving behind rising costs and mounting skepticism.
Part of the explanation appears to lie in how early projects were funded. In 2024, most initiatives were treated as experiments, paid for from innovation budgets. These funds were designed for exploration, with low expectations of long-term delivery. Failure was tolerated as the cost of learning.
But 2025 marked a shift. Enterprises began moving AI into core IT and business budgets, presumably subjecting it to the same scrutiny as any other enterprise-wide investment. CFOs are increasingly less willing to write off AI projects as experiments. They want outcomes that can be measured on the same terms as not only a digital transformation but a business transformation: efficiency improvements, lower risk exposure, stronger customer service, or entirely new revenue streams.
This new funding environment raises the stakes. Simply connecting to a model provider’s API is not enough. Deploying AI at scale requires specialist talent, disciplined engineering, and continuous support. Implementation already accounts for one of the largest categories of AI spend, often exceeding the cost of the models themselves. In some cases, implementation has become the decisive factor separating projects that deliver impact from those that remain stuck in perpetual proof-of-concept mode.
And when projects fail, the impact goes beyond wasted effort. Nearly 40% of enterprises report higher operating costs from stalled AI initiatives. Research suggests that up to 70% never progress beyond proof of concept. These aren’t minor technical glitches. They are structural failures that often weaken competitiveness and damage confidence.
In this post, we’ll examine common reasons why AI projects stall, what it often takes to make them work, and how Moody’s helps enterprises bridge the gap between promise and performance.
Learn more about how Moody's does AI
Why do AI projects often stall?
The data challenge
Even the most ambitious AI systems falter when the underlying data is fragmented, incomplete, or inconsistent. Enterprises with advanced infrastructure may still struggle with outdated records, inconsistent metadata, gaps in sources, and poor governance. High-value AI depends not just on access to data, but on data that is accurate, comprehensive, well-structured, and contextually relevant to the workflows it supports. Without this foundation, even the most advanced models risk producing misleading, incomplete, or non-compliant outputs.
Data foundations are often incomplete or unreliable. Teams may lack centralized data systems, leading to duplication, inconsistency, and risk. Without metadata or lineage tracking, trust in AI outputs erodes—teams can’t trace or trust what they’re building on.
Data interoperability is a growing challenge. Today, 74% of enterprises manage or expect to manage more than 500 data sources, creating significant integration complexity. Pipelines linking these sources are fragile. When upstream feeds shift, ETL systems fail, and teams struggle to trace outputs back to inputs. In industries like finance, that lack of clarity may translate directly into compliance risk.
The impact is tangible. In credit risk modelling, missing or inconsistent company filings can distort outcomes and may expose institutions to losses. In insurance underwriting, fragmented data across claims, customer history, and third-party sources may lead to models that overprice or underprice risk. In sales, incomplete or siloed customer data can cause teams to overlook viable prospects or misjudge demand, potentially leading to missed opportunities. In KYC and sanctions screening, poor data quality may drive false positives that overwhelm compliance teams, possibly slowing onboarding and increasing the risk of overlooking true matches.
LLMs are not designed to be consistent
The models themselves introduce their own complications. Large Language Models (LLMs) are probabilistic by nature. Ask the same question twice and you might get different answers. Public-facing models are intentionally designed this way, as a bit of randomness makes them more engaging, conversational, and human-like. This randomness is controlled by a parameter called temperature: higher values make responses more varied, while a temperature of zero makes them fully predictable.
In enterprise contexts, where consistency is vital, setting the temperature to zero and combining the LLM with RAG (Retrieval-Augmented Generation) significantly increases the rate of predictable outputs anchored in proprietary and auditable data, instead of hallucinations or shifting answers.
AI moves fast, models drift, and keeping up is costly
AI models are inherently tied to the data on which they were trained. As real-world conditions evolve, model performance gradually degrades—a phenomenon known as model drift. Left unmanaged, this may result in declining accuracy and reduced business value.
Compounding this challenge are the high costs of migration between successive model generations. Moving from GPT-3 to GPT-4, and now GPT-5, often requires significant redevelopment of integrations, retraining of custom workflows, and revalidation of compliance and safety measures. These transitions can consume substantial time and resources, while introducing operational risk.
Enterprises that lack systematic monitoring and retraining pipelines face a stark choice: continue relying on models that no longer perform to standard, or undertake costly, disruptive upgrades. In both cases, the absence of robust lifecycle management may expose organizations to potential efficiency losses, reputational risk, and increased total cost of ownership.
Misaligned value
Another common issue is the disconnect between AI teams and business users.
Many enterprises seemingly fall into the trap of building AI in a silo. Developers chase technical breakthroughs without anchoring projects in real-world workflows. The result: tools that may be elegant in design but irrelevant in practice.
The pattern is familiar. An “AI innovation team” produces a prototype, such as a chatbot for compliance reporting. But compliance officers already rely on structured dashboards and audit-ready PDFs. The prototype is admired in a demo but ignored in day-to-day work.
In retail, we have seen AI assistants designed to recommend products, but built without input from sales teams. Adoption often falters when the system doesn’t reflect how staff actually interact with customers. In healthcare, pilots that generate automated summaries of patient records may be ignored by clinicians who need outputs formatted to match existing case management systems. In financial services, failures may arise when firms pilot AI to draft credit memos or perform financial analysis. The systems may generate fluent text but struggle with calculations and structured data, producing incomplete or misleading outputs. In many cases, the lack of purpose-built tools and analyst oversight can result in teams abandoning new systems and reverting to existing processes.
Unless AI is solving a defined, high-value business problem, it risks becoming shelfware.
Security and compliance as afterthoughts
Finally, even technically successful projects can stall when they collide with regulatory and security requirements.
Many large language models originated as general-purpose tools, optimized for broad public use rather than enterprise-grade requirements. Even with large-scale deployments, organizations will often require additional contextual layers and supporting infrastructure to enable secure integration, policy-compliant data access, and firm-wide governance. Without secure access controls, audit trails, and robust oversight, adoption can stall.
In addition, existing governance requirements can be extensive. GDPR and data residency rules limit how and where information can flow. Basel guidance demands interpretability in financial risk models. Securities regulators such as the SEC and FCA require auditability of decision-support systems. And sector-specific mandates in healthcare, energy, and government add further layers of constraint.
Enterprises that delay grappling with these issues until late in development may find themselves forced into expensive re-engineering—or perhaps even cancelling projects outright.
What success looks like: Moody’s-Grade AI
Moody’s brings decades of experience helping customers make confident decisions leveraging world-class financial data and analytics. That same discipline informs our approach to AI: prioritizing aligning with business value, sourcing systems with consistent, comprehensive data, and embedding governance principles from the onset.
Business alignment and strategy
Often, the first determinant of AI success is alignment with business strategy. In many cases, projects begin with a focus on technology and only later search for a problem to solve. That approach rarely survives scrutiny once budgets and executive oversight come into play.
Moody’s aims to take a more strategic path, starting with a well-defined business objective. This helps foster AI efforts that are directed toward outcomes with potential to deliver measurable business value.
Clear use cases: Effective AI integration typically targets well-scoped, high-value opportunities. This can entail mapping workflows, identifying which data is used where, and understanding how analysts and managers make decisions today. Once those steps are clear, AI can be positioned to replicate parts of the process and begin enhancements. This strategic approach optimizes opportunities to save time, reduce bottlenecks, and surface insights while keeping decision-making firmly within human hands.
Measurable ROI: Establishing clear metrics is an important way to assess whether an initiative is delivering. Tying success to meaningful outcomes from the outset such as faster turnaround times, lower error rates, higher coverage, or reduced regulatory exposure, may result in a pilot that isn’t judged on novelty but on tangible impact.
Continuous monitoring: AI is not “set and forget.” Models can drift, data pipelines may shift, and business needs evolve. Continuous monitoring and feedback loops help track and maintain performance, which in turn helps protect against deterioration and costly surprises. For enterprises, this can include building monitoring systems that surface early warnings and integrating those insights into governance routines.
Adoption and integration
When considering business alignment strategy, it is important to consider that AI tends to deliver the greatest impact when thoughtfully embedded into the workflows users already rely on. Achieving production-grade adoption is more obtainable with deliberate integration across the enterprise.
Ability to deploy enterprise-wide: Systems should be designed to grow with demand, to help maintain performance as usage expands.
Workflow fit: To help achieve maximum impact, AI solutions should be deployed where they add the most value. Depending on the use case, this could mean being accessed through existing workflow platforms users are already familiar with or through a standalone platform.
Change management: As technology evolves, it’s important that processes and structures remain adaptable. Thoughtful workflow redesign, communication, and robust governance frameworks can help AI complement business operations rather than disrupt them. Moody’s approach emphasizes practical application to help AI become an integrated part of daily workflows.
Workforce capability
Embedding AI into workflows is only part of the equation. Long-term success depends on people. Technology is unlikely to deliver sustained value unless employees are equipped, supported and confident in its use.
Analysts benefit from understanding when to rely on AI-generated outputs and when to apply human judgment. Engineers should be aware of the governance risks, especially when linking proprietary datasets with external APIs. Similarly, compliance officers should have tools that offer transparency, auditability, and are positioned to hold up under regulatory scrutiny.
Equally important is culture. Some employees may worry that AI could replace aspects of their role rather than support them. Without clear communication and enablement, hesitation can grow.
Moody’s seeks to address these concerns through role-specific training, hands-on experimentation, and structured programs that make AI tangible in daily work. These initiatives help demonstrate that AI is a complement to human expertise, not a substitute, and opens the door for employees to see the technology as a productivity multiplier and a source of new capability.
Governance and compliance
Strong governance is often a prerequisite for AI adoption and a safeguard against operational risk. Moody’s incorporates governance into systems from the outset, so compliance and oversight are not simply retrofitted into the design. They are part of the requirements from the beginning.
Security by design: Systems are supported by secure pipelines, audit trails, and monitoring aligned with the organization’s risk profile.
Data stewardship: Customer data is processed with a focused intent on delivering meaningful outputs, guided by principles that prioritize privacy and analytical integrity and a strong commitment to responsible data management.
Regulatory focus: Moody’s AI platforms are designed for environments such as banking, asset management, insurance, and other compliance-driven sectors, and apply robust governance frameworks to support precision, transparency, and auditability.
This structured approach helps organizations drive confident innovation supported by governance-by-design AI deployments that can help accelerate adoption by removing barriers that might otherwise stall deployment.
Data excellence
AI performance is closely tied to the quality of its data foundation. Moody’s data is rigorously maintained, governed and layered with critical principles to help optimize AI solutions including:
Structured datasets: Curated for usability, these data sets include company financials, sanctions lists, economic indicators, industry-specific data, and news
Interoperability: Seamless integration across hundreds of sources, specifically useful for multinationals managing diverse data flows.
Transparency: Outputs that users can be audit, trace and verify.
The idea is straightforward: poor data management can lead to poor outcomes. Moody’s focuses on curating comprehensive datasets that help downstream systems perform as intended.
Information and interaction design
AI performance depends not only on the model itself but also on how it is deployed and used. Two key disciplines working as the proverbial two sides of the same coin play a critical role in optimizing an AI system’s impact: context engineering and prompt engineering. Context engineering governs what data the model sees, while prompt engineering shapes how users interact with the model.
While data quality is a key component to achieving sound AI outputs, how much data the model can process at once and what information is selected also plays an important role. Models operate within finite “context windows,” and overloading them with poorly matched or excessive data can create contradictions or even trigger hallucinations.
Context engineering aims to provide a way for a model to see the right data, in the right volume, calibrated to its context window to drive relevant outputs. This has the ability to transform a general-purpose model into one that behaves in a domain-specific way, not by changing the model itself, but by managing the information it receives. This practice helps deliver more consistent, actionable outputs aligned with enterprise needs.
Prompt engineering
If context engineering governs what the model sees, prompt engineering governs how it is asked to respond. The way a question is framed can significantly affect the clarity and usefulness of the output. Poorly structured prompts can yield vague or misleading results, while well-designed prompts support responses that are sharper, more targeted, and more actionable.
Prompting is not just about word choice; it is about intent. A strong prompt signals purpose, defines boundaries, and sets expectations for the model, helping to reduce irrelevant output and better align results with enterprise needs. In practice, it can mean the difference between a generic answer and a response that supports real decisions.
Moody’s approach combines both disciplines: the user takes the lead by prompting with well-framed questions, and we curate the context window to help the system see the relevant information. Together, context and prompt engineering help to reduce noise, sharpen clarity, and accelerate adoption across workflows.
Human-in-the-loop
Even with strong data foundations and curated context, performance often depends on how people interact with the model. Prompts are expressions of judgment, framing what the system is asked to do and how its outputs are interpreted and applied.
Moody’s advocates a human-in-the-loop approach, where AI is used to expand analytical capacity while ultimate decision-making and oversight remain with people. The model may surface insights, highlight patterns, and accelerate analysis, but it is the human who shapes the questions, interprets the answers, and provides accountability. Oversight is what keeps outputs aligned with regulatory expectations, institutional standards, and professional judgment.
Placing human-in-the-loop at the center of our approach reinforces a simple principle: AI is most effective when it enhances expertise and decision-making, not when it attempts to substitute for them.
Conclusion
AI is reshaping enterprise technology. The question is no longer whether adoption will happen, but how organizations will use it to deliver meaningful value.
Many challenges stem from recurring, familiar patterns: fragmented data, projects developed in isolation, and compliance treated as an afterthought. In contrast, successful implementations tend to follow a different path: grounded in business alignment, consistent data foundations, governance by design, and sustained support for adoption.
Moody’s brings these elements together to help enterprises move from pilots to production. Collaboration is a critical element of our approach and we are working closely with corporates, banks, insurers, and other institutions to apply generative AI solutions across functions from credit risk management and treasury to origination and supply chain. With multi-model architecture, proprietary datasets, RAG pipelines, and decades of experience supporting customers in regulated industries, Moody’s helps transform AI from an experiment into an enterprise asset, capable of delivering measurable impact and long-term competitive advantage.
The dividing line is clear. Enterprises that treat AI as an innovation experiment are likely to continue to face stalled projects. Those that operationalize AI with discipline, governance, and a sharp focus on business outcomes are more likely to capture the gains. In an environment where competitive edges are narrowing, the ability to embed AI effectively could become a defining factor in who leads, and who lags.
About the author:
Pavle Sabic is a global expert in enterprise AI strategy, helping Fortune 500 companies and large financial institutions across Europe, the Middle East, Asia, and the Americas embed GenAI and agentic solutions into high-value workflows.
At Moody’s, he leads the integration of domain-specific data and analytics into production-grade AI systems that enhance decision-making, uncover risk, and unlock capital. His expertise spans credit risk, automation, strategic data integration, and cross-functional go-to-market execution, making him a trusted partner in driving adoption of AI at global scale.
A published thought leader and frequent speaker on AI in financial services, Pavle has been featured on CNBC, and in the Wall Street Journal, Financial Times, Barron’s, and Fortune. His cross-sector experience, client-facing presence, and deep understanding of regulated markets position him to help scale transformative technologies at the next frontier of enterprise AI.
Moody's-grade AI: What does success look like?
Learn more about Moody's Agentic solutions
Learn More
Beyond prompts: Why enterprise AI demands context engineering
Enterprises have begun to discover what the GenAI hype can obscure: large language models are convincing but inconsistent unless fed the right data. That’s why the true differentiator in enterprise grade GenAI isn’t style, but substance — specifically, context engineering.