Everyone is talking about the cost of AI. Usually they’re talking about GPUs, model licensing, or token consumption. I think they’re looking in the wrong place.
The biggest cost of enterprise AI may turn out to be the people needed to supervise it.
A recent study found employees save about 11 hours a week using AI, but spend more than six hours checking outputs, fixing mistakes, adding missing context and making sure the results are actually usable. Someone coined the term “botsitting” for it. It’s a catchy name, but it points to a much bigger issue.
AI was supposed to reduce work. Instead, many organizations are creating an entirely new category of work around managing AI itself. That’s not a technology problem. It’s an operating model problem.
A Faster Sequel to the SaaS Era
I’ve seen this movie before.
Fifteen years ago, organizations embraced SaaS because it made software easier to buy and easier to deploy. Business units could solve problems without waiting for IT. Productivity went up. Innovation accelerated. Then Shadow IT arrived.
Suddenly nobody knew how many applications employees were using, where company data was being stored, or who had access to it. IT was still accountable, but it no longer had visibility or control. AI feels remarkably similar, but the difference is speed.
A SaaS application might expose a handful of systems; an AI agent can connect to dozens on its first day. It can read documents, summarize contracts, generate code, update records, and trigger workflows before most organizations have even documented that it exists.
The technology is moving faster than the operating model built to manage it.
That’s why I wasn’t surprised by IBM’s recent study showing that two-thirds of CIOs and CTOs are accountable for AI systems they don’t fully control. Most responders also said AI is spreading faster than governance can keep up. That isn’t because CIOs have suddenly become less capable. It’s because AI doesn’t fit neatly into traditional ownership models.
Security owns risk.
Legal owns compliance.
Developers build integrations.
Business units own outcomes.
Vendors control pieces of the underlying infrastructure.
Meanwhile, AI cuts across all of them. When something goes wrong, everyone owns a piece of the answer, but nobody necessarily owns the whole system. That creates a gap, and organizations almost always fill that gap with people.
Someone reviews AI-generated reports before they go to customers.
Someone checks AI-written code before it’s deployed.
Someone verifies recommendations before decisions are made.
Someone investigates unexpected behavior.
Some of that oversight is exactly what we want. AI shouldn’t operate without human judgment, especially in regulated industries or high-consequence environments. The problem is when people become the primary control.
From Chatbots to Autonomous Agents
Think about what happens as organizations move beyond chatbots.
Today, an AI assistant helps someone draft an email.
Tomorrow, an AI agent sends it.
Today, AI suggests an update to a customer record.
Tomorrow, it makes the change itself.
Today, AI recommends the next step in a business process.
Tomorrow, it starts the workflow automatically.
Every new action introduces another decision:
Should the AI be allowed to do this?
- Should it access that system?
- Should it retrieve that data?
- Should someone approve it first?
Most organizations answer those questions manually, and that works when you have ten AI use cases. It doesn’t work when you have hundreds…or thousands. The irony is that organizations are investing in AI to improve productivity, then asking highly paid employees to spend their time supervising it.
That’s not scaling AI. That’s scaling overhead.
Microsoft reached a similar conclusion in its 2026 Work Trend Index. After analyzing trillions of workplace signals and surveying 20,000 AI users, the company found that the biggest barrier to AI success isn’t the technology; it’s the organization around it.
Employees are adopting AI faster than their companies are adapting. The real challenge now is redesigning how work gets done: deciding what AI should do, what people should do, and how the two work together. I think that’s exactly the conversation enterprise leaders should be having. AI doesn’t create value simply because it’s deployed. It creates value when it’s deployed with a clear operating model that lets people focus on judgment instead of supervision.
Rethinking AI Governance: From Policy to Operations
This is where the conversation around AI governance needs to change. When people hear the word governance, they often think about policies, committees, and documentation. Those things matter, but none of them make a decision in real time.
None of them stop an AI agent from accessing data it shouldn’t see or pause a workflow that’s behaving unexpectedly. None of them explain, at the moment an action is requested, whether that action aligns with company policy.
That’s the gap.
Organizations don’t need more governance documents. They need governance that operates. Routine actions should happen automatically when they fall inside approved policies; higher-risk actions should require human approval, and sensitive systems should have clear boundaries around what AI can and cannot do.
And if an AI process starts behaving in ways nobody expected, there should be an immediate way to stop it. Not because AI can’t be trusted, but because every business process needs guardrails.
Positioning Humans Where They Matter
The companies getting the most value from AI aren’t eliminating humans from the equation. They’re putting humans where they’re most valuable.
Not reviewing every output.
Not checking every prompt.
Not babysitting every agent.
They’re focusing people on the exceptions, the edge cases, and the decisions that actually require judgment. That’s a much better use of human expertise.
The Real Enterprise Question
AI has reached the point where the technology is no longer the biggest question. Most organizations have already proven it can generate content, write code, summarize information and automate repetitive work.
The harder question is whether they can operate AI confidently at enterprise scale:
Can they see what it’s doing?
- Can they control what it has access to?
- Can they enforce policy consistently?
- Can they explain why an AI system took a particular action?
Those are operational questions, and they’re increasingly economic ones. That’s because every hour spent supervising AI chips away at the productivity gains that justified the investment in the first place. The companies that win with AI won’t be the ones running the most models or deploying the most agents. They’ll be the ones that don’t need an army of people watching them.
AI should amplify your workforce. It shouldn’t quietly become a new one.
