Chloe Maraina has spent her career at the intersection of big data and human intuition, transforming cold numbers into compelling visual stories that drive executive decisions. As a Business Intelligence expert with a deep background in data science, she has witnessed the evolution of enterprise technology from simple integration to the complex, often chaotic world of generative AI. Her perspective is shaped by a vision where data management isn’t just a back-office function but the very heartbeat of a company’s strategic identity. In this conversation, we explore why so many AI initiatives are currently hitting a wall and how leaders can navigate the organizational and structural shifts required to move from experimental curiosity to sustained business value.
It seems that nearly half of all enterprise AI projects are failing to demonstrate any real business value despite massive investments. From your perspective, why are these initiatives stalling so frequently when the technology itself seems more capable than ever?
The reality is that we are currently navigating a landscape of unrealistic expectations that would be difficult for any technology to meet. While nearly half of enterprises are struggling to show value, the core issue isn’t the model selection or the data pipelines; it is the organizational friction that surrounds these tools. I see projects derail because they are treated as isolated technology experiments rather than integrated business transformations. When CIOs jump in without a granular definition of success, they are essentially building a high-performance engine for a car that doesn’t have a steering wheel. We often see tech teams who are perfectly clear-eyed about architecture, but they’ve chosen the wrong projects because they didn’t have enough collaboration with the people on the front lines. It creates a cycle of disappointment where the structural failures of the company become the Achilles’ heel of the AI.
There is a growing sentiment that CIOs should stop leading these projects alone and instead shift toward a co-leadership model with business owners. How does this shift in governance change the way a project is triaged and selected?
If you ask a CIO to run an AI Center of Excellence in a vacuum, you are setting them up for a platform-centric failure. AI initiatives are fundamentally business projects that happen to use advanced math to achieve a goal, which is why co-leading with P&L owners is the only recipe for success. When business leaders find the opportunities first, they bring a level of pragmatism that prevents the “fishing expeditions” we saw a year ago, which often failed to find any useful cases. I’ve found that having an internal partner who is willing to be conservative and focus on viable opportunities is much better than relying on external consultants who might lack domain context. You need that executive sponsor who is willing to commit to the value and stick their neck out for what success actually looks like. It’s about ensuring the person who owns the financial outcome is also the one defining the metrics that feed into the P&L.
One of the most significant barriers to AI adoption is the fear and cynicism from the workforce, often described as a “fear of being replaced.” How can organizations convert these skeptical users into active participants in the AI journey?
The emotional weight of this transition cannot be overstated; employees often feel like they are being paid to train their own replacements. To overcome this, we have to change the narrative from replacement to assistance, treating the AI as a “brilliant intern” that requires constant human guidance and verification. Involving users from the very first discussion is non-negotiable because if they aren’t in on the process, they won’t be on board with the result. I’ve seen teams of ten people dedicated purely to coaching, helping users navigate these tools without coercing them into a fit that doesn’t exist. By identifying early champions and using phased rollouts with quarterly measures of adoption, you create a sense of incremental change that feels less like a threat and more like a tool. Ultimately, you have to make users accountable for the workflow changes, ensuring they see exactly how the AI acts as a helper rather than a shadow.
Many experts suggest that “first principles” should be applied to workflow redesign when introducing AI, yet others argue for simple task automation. Where is the middle ground for a company trying to avoid the pitfalls of complex process reengineering?
Workflow redesign and change management must go hand in hand, or you end up with a well-conceived AI that sits on a shelf for months. Applying first principles means knowing exactly what problem you are solving, what your inputs are, and how the future state should look, rather than just paving over old, broken processes. However, we have to be realistic; business process reengineering has struggled for 20 years to live up to its potential because it is incredibly expensive and slow. Today, agentic AI is better at handling simpler tasks, where you can tell a Large Language Model the steps and let it determine its own workflow. While we are in the early stages of agents optimizing complex, multi-step processes without human intervention, starting with task automation allows you to build a foundation. The key is to avoid a lack of reimagination while respecting the fact that large companies have thousands of processes that can’t all be redesigned at once.
There is a common belief that an organization must have pristine, perfectly integrated data before even considering an AI project. Is this “data readiness anxiety” actually preventing companies from getting started?
This conventional wisdom can be paralyzing, but the truth is that not every project requires a perfectly manicured data lake. While data is a massive issue for traditional machine learning models, it is often less of a hurdle for large language models, unless you are doing complex forecasting or customer-level personalization. If a project does require specific data, the best approach is to focus exclusively on that narrow slice and use AI-based cleansing and integration tools to solve the problem in real-time. Of course, you cannot ignore legitimate concerns around data quality and metadata gaps that could trigger regulatory issues or prevent compliance reporting. But stumbling on data concerns that don’t apply to your specific use case is an unforced error. You have to balance the need for governance with the need for momentum, focusing only on the data that moves the needle for the desired business outcome.
Given that AI outputs still require human review and can struggle with context, how should leaders manage the inherent unreliability of these models in high-stakes environments?
We have to stop confusing “smartness” with “experience” when it comes to LLMs; they can make dumb mistakes and lack the domain context that a human veteran brings to the table. In practice, shorter and more bounded tasks produce much more reliable results than long-running autonomous processes that tend to lose their rhythm. When a task becomes too complex and exceeds the available token window, the model starts to compress information, which is exactly when hallucinations and errors begin to creep in. In financial services, where 100% correctness is the only acceptable standard, a tool that works 99% of the time is often seen as a liability rather than an asset. This is why every output must have a verification step before it ever reaches a production environment or a client. You need visibility and observability into what these agents are doing to ensure they aren’t drifting into unintended behaviors.
Shadow AI is becoming a major concern for IT departments, with reports suggesting that nearly 70% of organizations have employees using prohibited tools. How can a CIO maintain governance without stifling the innovation employees are clearly seeking?
The rise of shadow AI is a clear signal that the sanctioned tools aren’t meeting the immediate needs of the workforce. When employees build ungoverned processes for critical work, they create a sustainability nightmare because there is no observability when those processes inevitably break. To combat this, companies like Jack Henry have provided over 100 sanctioned AI-based applications to ensure users don’t feel the need to look elsewhere for solutions. It’s about maintaining strict data governance through mandatory training and approval processes while simultaneously providing a path of least resistance for legitimate use cases. If you strictly prohibit unapproved tools without offering a viable, governed alternative, you are essentially inviting risk into your organization. You have to convert the “prohibited” energy into “productive” energy by offering tools that are both powerful and safe.
The “PoC Graveyard” is filled with projects that worked in a lab but failed in the real world. What financial and engineering milestones are necessary to ensure a pilot actually scales?
One of the biggest reasons pilots fail is a total lack of economic analysis before the capital is actually deployed. I’ve seen firms embark on initiatives only to realize later that the ongoing operating costs of the AI would be higher than the projected savings, which is a massive failure of due diligence. You need explicit “proof-of-life” milestones and accuracy thresholds that a project must hit before it moves from one stage to the next. Scaling also requires a specific type of design and engineering talent that is often different from the team that built the initial prototype. I prefer to see the same team take a project from PoC all the way to production to ensure that the tribal knowledge isn’t lost during the handoff. If you don’t have the right people involved and the right financial milestones, those pilots will inevitably get parked or die an untimely death.
Looking toward the future, what are the most significant risks to the long-term sustainability of AI projects once they are finally in production?
Once a project is in production, the real work of monitoring for model drift and managing technical debt begins. Gartner has predicted that within the next four years, 50% of enterprises will face delayed upgrades and rising maintenance costs due to unmanaged generative AI debt. To avoid this, you have to treat AI with the same IT lifecycle management disciplines you would apply to any other core system, including the use of model cards and drift monitoring. There is also a significant risk of vendor lock-in if you allow your data and workflows to be trapped within a single provider’s proprietary API. We should be pushing for open standards and modular architectures that allow us to swap parts of the AI stack as the technology evolves. If you don’t account for these sustainability issues during the design phase, they will be baked into your operations by the time you realize they are a problem.
What is your forecast for enterprise AI?
I believe we are moving toward a period of radical “nitty-gritty” pragmatism where the glamor of AI is replaced by the hard work of process integration. In the next few years, the gap between the leaders and the laggards will be defined not by who has the most advanced model, but by who has successfully redesigned their workflows to support human-AI collaboration. We will see a shift where agents move from simple task automation to autonomously optimizing complex workflows, though this will only happen for those who have mastered their data governance first. Some technologies that failed a year ago will be revisited and find massive adoption because the timing is finally right—sometimes the answer isn’t “never,” it’s just “not yet.” Ultimately, the enterprises that win will be the ones that work down from the desired business outcome and up from the technology foundation, meeting in the middle to create something truly sustainable.
