Something shifted in the last two to three weeks. Not incrementally. Visibly.
Anthropic started throttling access to Claude Code for heavy users — not because the product isn’t working, but because token consumption has gone through the roof. People living in agentic workflows are burning through capacity at a pace that has providers actively pushing them toward API plans just to keep infrastructure stable. When the supply constraint is too many people using the product too enthusiastically, that’s a different kind of signal than the usual AI hype cycle.
Meanwhile, I came across something Meta is apparently doing internally: token leaderboards. Employees ranked by how many tokens they consume. Not as a quirky experiment — as an actual performance signal. The logic: usage correlates with experimentation, experimentation correlates with improvement, and improvement is what they’re optimizing for. They’re literally incentivizing their people to use as many tokens as possible.
That’s one end of the adoption curve.
The Two-Speed World
At the other end, someone recently told me about a colleague in their audit team asking ChatGPT what COSO stands for.
I don’t say that to be dismissive. A year ago, that was a reasonable way to interact with these tools. But the gap between “asking AI generic questions” and “deploying agents that do your work while you sleep” has become enormous. And it’s widening every week.
I see it in my own team. Some people have completely transformed how they work — using AI tools for tasks they wouldn’t have dreamed of automating six months ago. Building workflows. Experimenting with agents. Pushing into territory that feels genuinely new.
Others have the same tools available and barely touch them. Or they use them the way they used Google five years ago: type a question, get an answer, move on.
Both groups are professionals. Both are smart. But they’re diverging at a speed that should concern anyone thinking about organizational capability.
This is the two-speed world. And it’s not a technology problem — it’s a governance problem.
114 Usage Trackers in One Hour
Here’s something that crystallized the absurdity for me.
I was browsing Reddit and came across a post pointing out that 114 people had posted within a single hour about building their own “usage tracker” app. The same app. Essentially the same functionality. One hundred and fourteen times. In sixty minutes.
And those are just the ones who posted about it. The actual number of people who built the same thing is probably an order of magnitude higher.
This raises a question that nobody in the AI enthusiasm bubble wants to ask: what are we actually spending our tokens on?
Are we learning? Sure, probably. Building is one of the best ways to learn. But at what point does “I built the same thing as a thousand other people this afternoon” stop being education and start being waste?
This isn’t just a consumer phenomenon. It’s happening in enterprises too. Departments building internal tools that already exist. Teams automating workflows that three other teams have already automated. The duplication of effort is staggering — and largely invisible because nobody has governance over what’s being built, by whom, or why.
For auditors, this should set off alarm bells. Not because building things is bad, but because ungoverned building at scale is a risk that barely anyone is tracking.
The Sandbox Question
Which brings us to the real governance dilemma: do we keep AI in sandboxes, or do we connect it to our actual systems?
I’ve been thinking about this a lot, partly because I’ve been living it. My own projects become dramatically more useful the moment I connect agents to real data, real APIs, real workflows. An AI trading system that’s connected to a live broker account behaves fundamentally differently from one running in a test environment. An audit agent that can access the actual ERP system is orders of magnitude more useful than one that analyses sample data you paste into a prompt.
The systems become way more effective once they’re linked. I’m convinced of that.
But “more effective” and “more risky” are the same thing here.
An agent connected to your email can draft and send messages on your behalf. That’s powerful until it sends the wrong thing to the wrong person. An agent connected to your financial systems can execute transactions. That’s transformative until it executes the wrong one. An agent that can read your company’s confidential documents can synthesize insights across the entire knowledge base. That’s incredible until you consider what happens if those documents leak through the agent’s context window to a third-party provider.
The delicate balance is this: if we don’t connect AI to our systems, we under-utilize it. We get the toy version. The demo. The “neat but not really useful” version that makes for good conference presentations but doesn’t change how work actually gets done.
But if we do connect it — without proper controls, without understanding the risk profile, without governance — we’re onboarding risks that most organizations aren’t remotely prepared to manage.
Two Camps, Neither Right
I keep running into two positions that are both wrong.
Camp One: “The only risk is not adopting fast enough.”
These are the token-leaderboard people. The “move fast and break things” crowd, now armed with AI agents that can actually break things at scale. Their argument has merit — there is genuine competitive risk in falling behind. But “adopt everything immediately with no guardrails” is not a strategy. It’s a prayer.
Camp Two: “AI is too risky. We’re not touching it.”
These are the organizations with AI policies that consist primarily of the word “don’t.” Don’t use AI for client work. Don’t input company data. Don’t experiment without approval from a committee that meets quarterly. Their caution is understandable — the risks are real. But prohibition doesn’t eliminate risk. It just pushes usage underground, where you have zero visibility and zero control.
The governance gap lives between these camps. Most organizations haven’t figured out what responsible adoption actually looks like in practice. Not in theory. Not in a policy document. In the daily reality of people doing their jobs with increasingly powerful tools.
What Governance Actually Needs to Look Like
Here’s where the internal auditor in me starts itching to build a framework. Because governance isn’t just policies. It’s answering specific questions:
What’s the purpose? Every AI deployment — whether it’s an individual using Claude Code to write scripts or an enterprise deploying an autonomous agent fleet — should have a clear answer to “what problem does this solve and for whom?” The 114 usage trackers didn’t have this answer. Neither do most corporate AI experiments.
What’s the quality standard? When an AI agent produces output, who determines if it’s good enough? What’s the review process? Are we just trusting the model and shipping, or is there a human checkpoint? And if there is — is that human actually qualified to review what the AI produced?
What’s the accountability model? When something goes wrong — and it will — who owns it? “The AI did it” isn’t acceptable. But the traditional audit trail (who approved, who executed, who reviewed) breaks down when autonomous agents are making decisions at machine speed. We need new frameworks for accountability that acknowledge the reality of how these systems work.
What are the boundaries? What data can AI access? What actions can it take? What’s the blast radius if it makes a mistake? These questions need answers before the agent gets connected, not after something goes wrong.
What’s the monitoring? How do we know if AI usage is productive or wasteful? How do we detect drift — the gradual expansion of what agents are allowed to do beyond their original scope? How do we ensure that the controls we put in place today are still relevant next month, when the tools have evolved again?
The Auditor’s Double Role
This is where it gets particularly interesting for people in our profession.
We’re not just spectators in the governance gap. We’re playing both sides of it simultaneously.
On one hand, we need to adopt AI ourselves. The 100x Employee thesis applies to us as much as anyone. If we’re not using these tools to transform how we do audit work, we’re falling behind — and our assurance becomes less relevant with every passing month.
On the other hand, we’re the ones who need to assess AI governance in the organizations we audit. We need to ask the hard questions about policies, procedures, controls, and accountability. We need to evaluate whether our auditees are managing AI risk or just hoping for the best.
You can’t credibly audit something you don’t understand. And you can’t understand AI governance without experiencing it firsthand — including the messy parts. The experiments that fail. The agents that do unexpected things. The moments where you realize the tool is more powerful than you anticipated and your controls weren’t designed for this.
That’s the real value of the adoption gap closing within audit teams. Not just efficiency. Understanding. The auditor who has built their own AI workflows, who has connected agents to real systems and dealt with the consequences, who has felt the tension between “this is incredibly useful” and “this is genuinely risky” — that auditor is worth more than a hundred who can recite the NIST AI Risk Management Framework from memory but have never touched the tools.
Where Does This Leave Us?
I don’t have a tidy conclusion. The governance gap is real, it’s widening, and it won’t be closed by policy documents alone.
What I do know is this: the conversation needs to shift from “should we adopt AI?” to “how do we govern its adoption?” The first question is settled. The second is wide open.
And if we — as auditors, risk professionals, and governance practitioners — don’t step into that gap, someone else will fill it with frameworks that don’t account for the realities of how these tools actually work.
That would be a waste of everything we bring to the table.