Dear CIO,

Not long ago, leaders were telling employees, "Don't use AI." It was risky, uncertain, and surrounded by concerns about security, compliance, and legal exposure. Then, almost overnight, the message flipped. Use AI everywhere. Put it in your goals, your demos, and your status updates. Make sure everyone knows you're AI-forward. Burn as many tokens as possible.

Now many of those same leaders are suddenly worried about token consumption. Some companies have removed token dashboards and leaderboards. Others are restricting access to AI assistants, steering employees toward cheaper models, or questioning whether all this token burn is producing meaningful results. A few are probably beginning to wonder whether their original "everyone use AI" mandate was a knee-jerk reaction. This is AI Token Whiplash.

Best Regards,
John, Your Enterprise AI Advisor

Corporate AI strategy has officially entered its whiplash era

To be clear, AI isn't the problem. The problem is how many organizations have approached AI adoption. The swings between fear and enthusiasm were predictable. I've been talking about token economics for more than a year, not because I'm anti-AI, but because I believe AI is one of the most significant technological shifts we've seen in decades. That's precisely why economics matter.

You can't tell an enterprise to "go use AI" and then act surprised when the costs start piling up. Tokens aren't magic pixie dust. Large organizations don't manage labor, real estate, manufacturing capacity, energy, or supply chains through intuition and enthusiasm. They manage them as cost structures. AI should be no different.

Instead, many companies have treated AI usage as a vanity metric. More usage means more adoption. More adoption means more innovation. More prompts must mean more productivity. As long as the charts keep moving up and to the right, everyone can feel reassured they aren't being left behind. The problem is that token consumption and business value are not the same thing. A team burning through millions of tokens isn't automatically creating more value. A developer using the most expensive model for every task isn't necessarily more productive. An employee applying AI to simple work that doesn't require it isn't transforming the business. A dashboard showing high usage proves only one thing: people are using the tool.

I often ask leaders a simple question: What productivity gains or business outcomes do you expect your AI investments to deliver over the next 18 months? Very few can answer it, and most admit they don't know. Some offer a rough estimate and say, "At least 2x." For those who say 2x, I ask a follow-up question: What do you expect your total AI costs to be over that same period? I've yet to meet a leader who can answer that with confidence.

Then comes the question that usually changes the tone of the conversation: What happens if your AI costs, including model consumption, governance, rework, hallucinations, validation, cleanup, and the people required to manage all of it, end up exceeding the productivity gains you expected? What if the fully loaded cost is greater than the return? The answer is usually some version of, "That's a scary thought." It should be. Responsible organizations don't manage critical operational resources on vibes. AI isn't just a capability story. It's also a cost structure. 

Fortunately, some of the dust is beginning to settle. The Wall Street Journal recently reported that corporate America is starting to ration AI as costs rise. Companies are taking a harder look at where the money is going, whether usage is producing measurable outcomes, and whether expensive models are being applied to work that less costly tools could handle just as well. Maybe now leaders will start paying attention. The answer isn't to panic and shut everything down. That's no smarter than telling everyone to use AI without any economic guardrails. The answer is to mature the operating model. Leaders need to stop lurching from one extreme to another. Don't ban AI because it feels risky, don't mandate AI because it looks good in a board presentation, don't celebrate token consumption simply because it creates the appearance of innovation, and don't start rationing tokens blindly just because the invoice finally became uncomfortable.

This is no longer theoretical. The sticker shock is already showing up inside large organizations. Uber reportedly exhausted its 2026 AI budget by April. Microsoft reduced some Claude Code access, although the company said standardization was the primary driver. Amazon reportedly scrapped an internal AI usage leaderboard after previously encouraging employees to consume more tokens. Meta, Salesforce, DoorDash, and others are confronting the same reality: AI usage is growing rapidly, and the economics are starting to bite. That's the part leaders need to sit with.

A year ago, companies wanted visible AI activity. They wanted adoption metrics, experimentation, and usage numbers they could point to as evidence of transformation. Those numbers made for a compelling story. Now they come with a price tag. This is the difference between AI as a demo and AI as an operating model. In a demo, more usage feels exciting. In an operating model, increased usage has to be justified. Eventually the bill arrives and asks the question the hype cycle avoided: What did we actually get for all those tokens? That's why the sticker shock may ultimately be healthy. It forces organizations to stop treating token consumption as innovation and start treating it as spend. Not spend that must be eliminated, but spend that must be managed. Like any other investment, it should be connected to outcomes like productivity, quality, cycle time, shipped work, revenue, or better decisions. The companies that figure this out won't ask, "How do we reduce token usage?" They'll ask a much better question: "Which tokens are creating value, and which ones are just making us feel busy?"

Answering that requires understanding where AI actually creates leverage. It means knowing when the expensive model is worth the cost and when a cheaper one will do the job. It means measuring outcomes rather than consumption and teaching people how to use AI effectively instead of simply encouraging them to use more of it. Most importantly, it means connecting token spend to real business results. This is the next phase of enterprise AI. The first phase was fear. The second phase was hype. The third phase is economics.

It's what happens when a technology moves from novelty to operating model. The free-for-all begins to fade. The costs become visible, the tradeoffs become real, and organizations are forced to make decisions grounded in value rather than excitement. The winners won't be the companies with the highest token counts. They'll be the ones who learn how to turn tokens into measurable outcomes. AI isn't going away, and neither are token costs. The question is whether leaders can finally manage AI as a business capability instead of treating it as a slogan, a science project, or a panic button. The dust is settling, and now it's time to stop managing the bill and start managing value.