Ask most companies to define AI governance, and they’ll describe an approval process: a request comes in, someone reviews it, a tool gets a yes or a no. Donna Medeiros, VP of AI and Data Advisory at Data Society, thinks that definition is missing half the job.
“So part of governance is knowing what is actively in use and what is enabling the business. There’s what gets approved, but there’s also the view of what is currently in place for AI, data, and tooling that is enabling and empowering the business, maybe even driving some of the business model changes.”
Approval Isn’t the Same as Visibility
An approval process tells you what was allowed to start. It tells you nothing about what’s still running six months later, who’s using it, or whether it’s actually doing anything for the business. Those are two different questions, and Donna is clear that governance has to answer both.
“It does have a governance function: to understand the systems in place, categorizing, even having a prioritization matrix of AI initiatives, AI systems that then fall under those business use cases, business initiatives.”
Nobody Can Actually Count What They Have

This isn’t a hypothetical gap. According to SAP’s LeanIX Agentic AI Survey 2026, 98 percent of companies have deployed or plan to deploy AI agents, but less than half have visibility into an inventory of them. Gartner projects the average Fortune 500 company will be running more than 150,000 agents by 2028, and only 13 percent of organizations currently believe they have the governance in place to manage that scale. A separate survey found that 96 percent of enterprises run AI agents while only 12 percent have any centralized system to manage them, and 94 percent of IT leaders said the resulting sprawl is adding complexity, technical debt, and security risk, according to ibl.ai’s analysis of enterprise agent inventories.
The Cost of Never Retiring Anything
Donna hears the same pattern across nearly every organization she advises.
“What we hear from most organizations is they have way too much in play, way too much systems that aren’t being used, way too much data that’s old, not being used. So there really needs to be some kind of AI governance lifecycle management and data governance lifecycle management. What is the retirement plan for systems, retention of data, retention of the systems itself?”
The waste is measurable even outside AI specifically: 66 percent of enterprise SaaS licenses are entirely untouched or surplus to actual need, with 15 percent sitting completely unused and another 51 percent running at under half capacity, according to ToolRelief’s 2026 analysis of enterprise software spend. That adds up to an average of $19.8 million in annual waste per organization and 25 to 30 percent of total software budgets. Consolidation isn’t catching up either. The share of application portfolios companies actually managed to retire dropped from 14 percent in 2024 to just 5 percent in 2025. Every one of those unused, unretired systems is a line item and, increasingly, an ungoverned AI or data surface nobody’s watching.
Whose Call Is It
Knowing what’s in use is only useful if someone is actually accountable for the answer. Donna is specific about where that ownership sits.
“Who in the business line has that data and AI ownership point? Is it the data owner, the AI ambassador, AI governance lead for that business line, along with the centralized function? But clearly, at the central function, it’s not usually a call that gets made.”
That split ownership matches how data governance already works in mature organizations. A data steward or data owner typically sits inside the business function itself, not IT, often at the level of a VP or business leader with real P&L responsibility for that domain, because that’s who understands whether a given system is actually creating value. A central AI or governance function can set the framework and hold the inventory, but the decision to keep or retire a specific tool has to involve the business line living with the consequences either way.
Donna’s closing point is the one that keeps this from becoming just another cost-cutting exercise.
“But definitely it goes to who owns the part of the governance function that does prioritize what’s in use, and ensuring they’re not killing off things that actually are providing value.”
An inventory without judgment turns into an axe, cutting whatever nobody happened to be watching instead of what’s actually not working. The real job is figuring out which of those 150,000 agents or old SaaS licenses are quietly doing real work, and which ones are just costing money while nobody’s looking.
