The Six Percent Problem Nobody Wants to Talk About
Only six percent of organizations reported meaningful return on investment from their AI initiatives in 2024. Six percent. Tony Santiago, Managing Director at CapNexus Applied Intelligence and co-author of the AWS Cloud Adoption Framework for AI, opened Episode 290 of the Localization Fireside Chat with that number, and it hangs over everything else that followed. The reason for that failure rate is not that companies lack ambition or budget. It is that most organizations are building before they can govern what they have built. They are deploying agentic systems before they can observe what those agents are actually doing, before they have the data foundations to feed them cleanly, and before anyone in leadership has asked the uncomfortable question: if this goes wrong, how would we even know?
Tony left a six-year run at AWS, where he helped global system integrators and Fortune 500 companies stand up AI practices and contributed to foundational cloud governance frameworks, to join CapNexus and do something different. Not write the playbook for others. Apply it directly. The companies CapNexus serves, typically in the two to five billion dollar revenue range, are too large for a boutique and too small for a GSI to genuinely care about. They need enterprise-grade rigor delivered with the flexibility of a team that is not running them through a factory. What Tony brought with him was a methodology he calls the AI Capability Flywheel, built around four phases: govern, build, foundations, and enforce. The insight behind it is disarmingly simple. Most organizations land somewhere in the middle of that wheel, in the build phase, with use cases deployed and systems running, but with the govern and enforce phases sitting almost untouched. That imbalance is where the horror stories come from.
Shadow AI Is Not a Policy Problem, It Is an Enablement Failure
One of the sharpest exchanges in the episode came when Robin raised shadow AI, the quiet epidemic of employees uploading company emails into free-tier AI tools because it saves them forty minutes on a task their organization has not officially sanctioned a better alternative for. Tony’s response reframed the whole conversation. Blocking access does not solve the problem. Giving people monitored, governed, internal access does. When you can observe how your employees are actually using AI tools, you stop guessing about which workflows to invest in and start seeing the evidence directly. The developers and frontline workers using AI creatively are not breaking policy out of laziness. They are telling you exactly where the highest-value automation opportunities live. The organizations treating shadow AI as a compliance violation are missing the signal entirely.
The same logic extends to governance more broadly. Tony made the point that an AI policy written by a Chief AI Officer and distributed as a PDF is not governance. Governance is what happens at the platform level, at the edge, where what comes in and goes out is observable, auditable, and aligned to organizational policy. The guardrails have to be built into the system, not handed down as a memo. This is the enforce phase of the flywheel, and for most organizations, it is the weakest link in the chain.
The CIO and CTO Are Both Right, and That Is the Problem
Tony offered a nuanced take on the tension Robin named directly: some CTOs are already building while their CIOs are still designing guardrails. His read is that this is a healthy push and pull, not a dysfunction to eliminate. Gartner’s recent symposium suggested that CEO pressure is pushing toward a convergence of those roles, a single accountable leader who owns AI, data, governance, and IT infrastructure together, because AI is increasingly where revenue gets made or lost. Whether or not that consolidation happens, the underlying point holds. You cannot separate the decision to build from the decision to govern. When Tony’s team assesses a client who wants to deploy a fully autonomous agentic system for finance operations, the first question is not what tools to use. It is whether the governance layer is mature enough to oversee an agent acting without a human in the loop. If the answer is no, you build governance first. Not because governance is comfortable, but because deploying without it means you have no way to know when it fails.
If this conversation made you rethink where your organization sits on the AI maturity curve, you will want the full episode. Watch on YouTube if you want to see the back and forth in full, or Listen on Simplecast if you are catching up on the go. Either way, Tony’s flywheel model and his six percent number are worth sitting with for a while. Get every episode of the Localization Fireside Chat on your phone by downloading the free LFC app on the App Store or Google Play.
Leave a Reply