You can’t defend what you can’t account for. And most organizations can’t account for what’s actually running.
The Question Nobody Wants to Be Asked
In my last piece, Who Owns the Outcome?, I argued that every enterprise is quietly assembling a digital workforce. I also argued that governance starts with knowing what you actually have.
That sounds obvious. In practice, it’s where most organizations already fail.
Your organization has been investing in AI for a while now. Real money. Real resources. Real expectations from leadership.
At some point… probably sooner than you’d like… someone is going to ask you to show the math. A board member wants to understand the scope of adoption. A potential acquirer wants to know what’s running and what it can touch. A regulator wants an inventory. A client asks what AI is involved in delivering their services.
Can you answer those questions with confidence? Not with a slide deck or a rough estimate. With an actual list.
Most organizations can’t. And that gap… between the AI investment narrative and the operational reality on the ground… is one of the more quietly uncomfortable places for technology leaders to sit right now.
Here’s why that matters beyond the obvious compliance angle. You can’t measure value from something you haven’t inventoried. You can’t optimize it. You can’t course correct when something underperforms. You can’t defend the spend when someone asks you to show the return. And you can’t govern what you can’t see.
The inventory problem isn’t just a security issue with a compliance bow on it. It’s a strategy problem. And it’s sitting underneath almost every serious AI conversation happening at the leadership level right now.
You Probably Know Less Than You Think
The instinct is to push back a little here. We know what we’ve built. We know what we’ve bought. We have a procurement process.
Maybe. But that’s only part of the picture… and increasingly it’s the smaller part.
Think about what’s actually in your environment right now. The SaaS platforms you already pay for that shipped AI features in a product update nobody flagged during renewal. The internal team that built something useful, made it operational, and never ran it through a security review because it didn’t feel like “a system.” The vendor who embedded an agent into a workflow integration and technically disclosed it… somewhere in the release notes. The developer who connected a third-party tool to internal systems because waiting for IT approval would have taken three weeks and the deadline was Tuesday.
None of those people necessarily did anything wrong. They were solving real problems with the tools available to them. That’s exactly how you end up with a digital workforce you didn’t formally hire.
This Is a Strategy Problem, Not Just a Control Problem
And here’s what makes this harder to track than traditional software.
Traditional IT assets are relatively stable. A server doesn’t retrain itself overnight. A database doesn’t quietly expand its own access permissions. You deploy it, you document it, you review it on a schedule.
AI systems don’t behave that way. A static inventory of your AI environment is probably stale before it’s finished.
Here’s what makes it genuinely different:
They’re dynamic. Models get updated, retrained, fine-tuned. Prompts evolve. Integrations shift. An agent you inventoried three months ago may behave meaningfully differently today without anyone having formally changed it.
They’re embedded and often invisible. A SaaS platform ships an AI feature in a product update. A workflow integration gets an agent layer added by the vendor. A copilot gets bundled into a renewal. None of these show up in your procurement system as “AI agent acquired.” They show up as a line item you already approved.
They’re inconsistently labeled. Copilots. Assistants. AI-powered workflow enhancements. Smart triage. There’s no standard naming convention and vendors have every incentive to describe their AI features in ways that don’t trigger your governance process.
They’re cross-functional and decentralized. The teams deploying them aren’t always talking to each other. Marketing has one. Operations has two. The engineering team built a handful. Nobody has the full picture because no single team was responsible for tracking the whole.
Put those together and you don’t just have an inventory problem. You have a category of asset that actively resists being inventoried by traditional means. It’s one reason moving from prototype to platform requires more than just standing up a few useful agents.
There’s a lot of pressure on technology leaders right now to demonstrate that AI investment is paying off. Boards want ROI. Executives want productivity numbers. Investors want evidence of competitive differentiation. And most organizations are trying to answer those questions without a reliable picture of what they’re actually running.
You can’t credibly claim value from systems you can only partially see. You can’t attribute outcomes to agents you didn’t know existed. You can’t optimize a portfolio you haven’t fully mapped. You can’t build a coherent AI strategy on top of an inventory full of holes.
The organizations that will win the AI value conversation… with boards, investors, clients, and acquirers… are the ones that can connect investment to outcomes with something stronger than anecdote. That connection starts with knowing what you have.
When the Gap Becomes Expensive
Two moments where the inventory gap stops being an abstract concern and becomes immediately painful. Both are happening now.
The first is the board conversation. Leadership has approved real AI investment. Someone asks what the organization is getting for it. The honest answer in most organizations is that they don’t have a clean enough picture of what’s running to connect the spend to the outcomes. That’s an uncomfortable place to be in a strategy review… and an even more uncomfortable place to be in a board meeting.
The second is the acquisition process. A PE firm is in due diligence. Standard question in the tech assessment: produce an inventory of AI systems in use. Not just what you built… what’s running, what vendors have embedded, what your teams have stood up, what’s touching customer data and how. The organizations that can answer that question cleanly are demonstrably more governable. That matters to acquirers. It’s starting to show up in enterprise procurement questionnaires. It will matter to regulators.
The thing about diligence moments is they don’t give you six weeks to build the system. The organizations that sail through that conversation built the inventory before anyone asked for it.
What a Registry Actually Gives You
Not a technical deep dive here… just the leadership framing of what visibility actually enables.
Think of it as the CMDB for your digital workforce. When it works, you can answer five questions cleanly for every agent running in your environment:
- What is it, and who owns it?
- What systems and data can it reach?
- Was it sanctioned through a formal process or stood up informally?
- What’s the risk tier for its actions?
- When was it last reviewed, and by whom?
Without a registry, none of those questions have reliable answers. With one, they become operational inputs into everything that follows… guardrails, posture management, audit trails, identity and authorization. It’s not just an inventory. It’s the foundation the rest of the governance stack sits on.
It’s also what turns “we’re investing in AI” from a story into something you can actually stand behind.
Start Before Someone Asks
The organizations that will answer the inventory question cleanly aren’t the ones that build a registry when a regulator or an acquirer asks for it. They’re the ones that build it before anyone asks.
The starting point isn’t a platform or a tool. It’s a question. Ask your team today: if I needed a complete list of every AI agent running in this organization by end of week, what would that process look like… and what would it miss?
The answer tells you where you are. The gaps in the answer tell you what to build.
An incomplete registry that’s actively being maintained is more defensible than a confident answer built on false visibility. Start with what you can see. Build the discipline to find what you can’t. And understand that every AI system you can’t account for is a system whose value you can’t claim and whose risk you can’t manage.
Knowing what you have is the first step. The next one is making sure what you have is properly credentialed, scoped, and authorized to do what it’s doing. That’s where we go next.
Managing the Digital Workforce | Part 2