by Tony Santiago, Managing Director, Global Head of AI & AWS Practice

In my six years at AWS, I held roles that gave me an unusual vantage point on how enterprises actually adopt AI. As a Worldwide Senior Partner Solutions Architect, my job was helping AWS's largest global strategic partners build out their Generative AI and Agentic AI practices. I worked alongside their leadership teams to shape how they'd serve our joint customers. Sometimes I got pulled into highly strategic deals where I worked directly with the partner to win. But most of my time was spent helping these firms stand up the capabilities, methodologies, and teams they needed to deliver AI at scale.

That vantage point showed me the same pattern over and over, across partners and their customers. A team would build a proof of concept. The demo would land. Leadership would get excited. Budget would be approved. And then somewhere between "successful pilot" and "the business runs on this," momentum would stall. The technology was fine. What stalled things was that the operational questions still hadn't been answered: Who owns this once it's live? How do we monitor drift? What happens when the data changes? Who retrains, and on what schedule?

MIT Sloan's 2025 research put a number on what I'd been watching firsthand: 95% of generative AI pilots never make it to production. That number didn't surprise me. The technology layer has been ready for a while. What these organizations need is the layer between a working model and a business outcome someone is accountable for. Call it operationalization, call it production readiness, call it whatever you want. The point is that most organizations haven't built it yet.

The Bottleneck Is Organizational, Not Technical

At AWS I helped develop the Cloud Adoption Framework and the CAF for AI. Those frameworks exist because AWS recognized that cloud and AI adoption aren't purely technical problems. They're organizational ones. The CAF for AI walks organizations through the business, people, governance, platform, data, and operations perspectives required to generate value from AI. The organizations who made progress weren't the ones with the best models. They were the ones who had aligned their operating model around AI before they deployed anything.

Most enterprises today have access to powerful models and scalable infrastructure. What they need is a deliberate system for getting AI from experiment to operation. That system has to cover four domains that all depend on each other: how you govern AI, how you structure your data, how you build and deploy the platform, and how you run and monitor it in production. Weakness in any one of those domains holds back the others. You can build a strong platform, but if your data is ungoverned, your models are making decisions on information nobody has validated. You can have governance policies on paper, but if operations has no way to enforce them in production, they're theater.

Most organizations invest heavily in the build, the platform, the models, the infrastructure. That's the exciting part. But governance, data foundations, and production operations tend to lag behind, and they're the domains that determine whether AI actually runs reliably at scale. None of this is optional. It's the core of what makes AI work in an enterprise.

This problem gets worse in regulated industries. The questions that stall AI projects everywhere become hard requirements when regulators are involved: Who owns the output if it's wrong? How do you prove to an auditor what the model considered? What happens when the underlying data shifts and the model starts behaving differently than it did in testing? SAS's own research found that only 11% of banks have figured out trustworthy AI. The other 89% are stuck in exactly this gap, their models perform fine, but they haven't yet built the governance processes to answer these questions well enough to put them into production.

Platforms Are Starting to Build for This

I've been paying attention to how platform vendors are responding, and some of them are finally building for operationalization rather than just capability. SAS is one I've been looking at closely. Their Intelligent Decisioning framework in SAS Viya lets you build agentic AI applications that pair LLM reasoning with deterministic business rules so the final output stays explainable and auditable. Their AI Governance Manager puts lifecycle monitoring, drift detection, bias checks, and compliance documentation inside the AI pipeline from the start rather than bolting it on later.

SAS picked up a Leader position in the inaugural 2026 Gartner Magic Quadrant for Decision Intelligence Platforms. That reflects a broader shift I think matters: the market is starting to reward platforms that help you run AI in production under real governance constraints, not just platforms that help you build models faster.

The industry needed this correction. Operationalization has been treated as a phase that happens after the interesting work for years. It is the work.

The Same Problem, Two Layers Deep

Here's what caught my attention. SAS is solving the operationalization problem from the analytics and decision layer. But on the AWS side, Amazon Bedrock AgentCore (GA since October 2025) is solving a version of the same problem one layer down, at the agent infrastructure level. AgentCore handles the runtime, identity, observability, and policy enforcement for AI agents. You write your agent in whatever framework you prefer, and AgentCore wraps governance around it using Cedar policies so the agent operates within defined boundaries.

These two platforms operate independently today. But they're working the same principle from different starting points. SAS Intelligent Decisioning pairs LLM reasoning with deterministic business rules so the decision output is auditable. AgentCore enforces identity, policy, and evaluation controls so the agent itself stays within governance boundaries. SAS governs what gets decided. AgentCore governs the thing doing the deciding. Both are about making sure AI in production stays explainable, controlled, and accountable.

The infrastructure for a deeper connection between these layers already exists. SAS and AWS signed a strategic collaboration agreement in 2023. SAS Viya runs natively on AWS with marketplace listings and published reference architectures. SAS Customer Intelligence 360 already uses Amazon Bedrock for generative AI features like audience copilots and content generation, with Bedrock Guardrails handling responsible AI filtering. I spent years on the AWS partner side watching how these ecosystems grow toward each other. When two platforms independently arrive at the same architectural principle, the integration usually follows faster than either side expects.

A Proof Point

At CapNexus, we've been testing this idea in practice. We recently partnered with a major North American retailer to build a proof of concept using SAS Viya that connects AI driven demand forecasting with packaging carbon footprint metrics. The concept improved forecast accuracy by 63% by enriching traditional sales history with external signals like weather, economic indicators, and regional search behavior. That accuracy gain became the foundation for linking every forecast to its sustainability impact.

What made the proof of concept successful wasn't the model alone. It was what happened before anyone touched the platform. The team started with a meaningful business question, built a small cross functional group that combined retail domain knowledge with analytical capability, and worked in a low risk environment where they could explore without the weight of a full enterprise program. Those conditions let the team move from question to insight fast enough to demonstrate real value. The concept now has a path to operationalization, and the approach is repeatable across industries.

We're presenting this work at SAS Innovate on April 28th. The session gets at the shift I keep coming back to: moving from proving that AI works to proving that your organization can work with AI.

Why CapNexus Exists

I joined CapNexus because this is the problem I want to spend the next chapter on. The company was formed by bringing together three firms with distinct strengths under one roof. One team handling strategy, development, implementation, hosting, and support. That structure exists because the fragmented model, where different firms own different phases, makes it harder to maintain momentum and accountability across the full lifecycle.

The models are capable. The platforms are catching up. The question worth asking now is whether your organization has the operating model to make AI work at scale. That's the question I spent six years at AWS watching companies wrestle with. And it's a better question than "which model should we use?"

We've been building a structured approach to diagnosing exactly where organizations get stuck across governance, data, platform, and operations, and which of those domains is the constraint holding everything else back. More on that soon.

‍