The development of agentic AI marks a departure from the deterministic systems of traditional software and classical machine learning. In those worlds, inputs and outputs are predictable, enabling precise testing and validation. But agentic AI brings with it a level of non-determinism with which enterprises must now grapple.
As Jayeeta Putatunda (Director of the AI Center of Excellence at Fitch Group) highlighted in her discussion with Fiddler AI CEO Krishna Gade, organizations can’t simply apply old methods to these new systems.
The Real Shift: From Determinism to Non-Determinism
For decades, enterprise software operated in a deterministic world where every input had an expected output. With agentic AI, unpredictability is a feature, not a bug. Large language models can generate varied outputs for the same prompt, because workflows are probabilistic rather than guaranteed. This introduces new complexities that traditional testing frameworks are ill-equipped to handle. Enterprises must design systems that account for this uncertainty by embedding resilience, auditability, and layered safeguards. In an environment where outputs vary with each execution, entirely new strategies are required to build trust into our most powerful tools.
The opportunity here is equally significant. Non-determinism, when properly managed, can make systems more adaptive and responsive to real-world complexity. Instead of brittle pipelines that fail when encountering edge cases, agentic systems can reason through uncertainty, provide alternatives, and expand what automation can achieve. The challenge is not to eliminate unpredictability but to harness it responsibly, treating it as a feature to be governed rather than a bug to be fixed.
ROI > Novelty
Enterprises are inundated with exciting new frameworks, but novelty is not a strategy. As Jayeeta stressed, the winners in this space will focus relentlessly on business impact. That means prioritizing high-value use cases where automation or augmentation frees analysts to focus on higher-order tasks. Chasing hype-driven experiments often leads to wasted months of prototyping with tools that may be obsolete before they see production. Every decision must be driven via the ROI lens, from model selection to deployment strategy.
A true ROI-first perspective forces organizations to ask harder, second-order questions. It's not enough to ask if a new framework reduces costs or increases efficiency. We must ask if we are even defining efficiency in a meaningful way. What does it really mean to free up an analyst's time? Is that time being reinvested in higher-value strategic work, or is it just disappearing into the ether? By resisting the allure of experimentation for its own sake, enterprises position themselves to deliver tangible value quickly. Sustainable adoption requires anchoring every initiative in measurable, meaningful outcomes.
Data: The Ultimate Enterprise Asset
When it comes to data discipline, Agentic AI raises the stakes. What Jayeeta calls the “data prep tax” is unavoidable. Enterprises must invest in making their data AI-ready, versioning prompts and evaluations with the same rigor as APIs. This approach ensures reproducibility and traceability, which are vital when dealing with systems whose outputs can’t always be anticipated. Without this, enterprises risk opacity at the very moment transparency is most needed.
This tax, while burdensome, is also an investment in long-term scalability. Enterprises that embrace rigorous data practices can iterate faster, debug failures more effectively, and build confidence in their outputs. Treating data as a strategic asset also enables better cross-team collaboration: business leaders gain trust in the systems when they know that data is traceable and auditable. In a landscape where trust is fragile, disciplined data management becomes the bedrock of responsible AI adoption.
Hybrid Architectures: Bridging Predictive and Generative
Pure reliance on LLMs is risky, particularly in industries like finance where a single error in numerical reasoning can have catastrophic consequences. Hybrid architectures — combining predictive models, LLMs, causal inference, and knowledge graphs — are the most pragmatic way forward. These systems validate and ground outputs, ensuring correctness where it matters most. Blended intelligence will always surpass blind reliance.
This “blending” is more than a technical safeguard, it represents a cultural shift in how enterprises should think about AI. Instead of treating new technologies as replacements for old, the goal is to orchestrate their strengths in complementary ways. Predictive models provide reliability, LLMs offer flexibility, and causal models explain why outcomes occur. Together, they can build systems that are both powerful and trustworthy, striking a balance between innovation and responsibility.
Observability as a First Principle
The key to this cultural shift is that observability can no longer be an afterthought. Traditional MLOps often treat monitoring as something to add post-deployment. In agentic AI, observability must be architected from day one. That means logging tool calls, tracking token usage, measuring drift, and embedding human-in-the-loop oversight. Done right, observability transforms uncertainty into insight, supercharging the development loop.
A proactive approach to observability accelerates innovation. By tracking where agents fail, organizations build invaluable feedback loops for improvement. Patterns in failure data can reveal blind spots in business logic, weaknesses in training data, or opportunities to streamline workflows. Observability, then, is not just about risk management — it’s about creating a virtuous cycle of continuous learning and refinement that keeps agentic systems aligned with enterprise needs, and allows each contributor to maximize their impact.
Trust: a Human-Technology Partnership
Collaboration is the foundation for adoption, accountability, and ultimately, trust. It may seem simple, even obvious, but this is where even the most capable organizations can fall down. Agentic AI cannot succeed without stakeholder trust. Developers alone cannot define success metrics. Business stakeholders must be partners in setting criteria that reflect actual needs, not just technical feasibility. What seems trivial to a technologist may be critical to an analyst, and vice versa. Blending perspectives is critical to developing impactful tooling.
Trust also requires education and transparency. Jayeeta underscored the importance of demystifying AI systems for leadership and business users, ensuring they understand not just what a system does, but its limitations. By bringing stakeholders into the journey — through pilots, shared evaluation criteria, and open conversations about risks — enterprises reduce fear and build confidence. Trust emerges when both humans and machines are co-architects of a system’s purpose and performance.
A Pragmatic Path Forward
Do not abandon ambition, channel it responsibly. Start small with ROI-driven use cases. Pay the data prep tax. Design for observability and accountability from the outset. Embrace hybrid architectures. And above all, make business stakeholders co-owners of the journey. Agentic AI is not just about building more powerful systems; it is about building systems people can trust.
Pragmatism is the antidote to hype. As enterprises race to operationalize AI agents, the differentiator won’t be who adopts first, but who adopts responsibly. Those that take a measured, disciplined approach will not only avoid costly missteps but will also be the ones to truly unlock the promise of agentic AI. The future of this technology will be defined not by the speed of adoption, but by the wisdom of how it is adopted.
Watch the full AI Explained fireside chat:
