For the past few years, companies around the world have been wrestling with the challenge of getting their employees to adopt artificial intelligence. While some businesses initially clamped down on AI use over security concerns, many others pushed hard to drive uptake. That race is now over.
Today, companies are confronting a fresh problem: how to stop AI spending spiralling out of control. The bill is coming due, with Amazon, Walmart and Uber all tightening controls as the cost of AI soars.
The reason? Frontier models are no longer priced like conventional software. Increasingly, they are becoming metered services. Every prompt now adds to the bill, forcing firms to scrutinise far more carefully how much they use them, and whether the returns justify the cost. Often, they do not.
Part of the reason is that for many AI services, flat monthly subscriptions are now giving way to token-based pricing, whereby companies pay for what they use. Anthropic and OpenAI are both doing this, while others charge by outcome: Zendesk, the customer service platform, charges businesses up to $2 each time its AI solves a customer support issue.
And many businesses are now devouring tokens with such a voracious appetite that AI bills keep climbing.
On top of that, the pricing itself is becoming more fragmented. Premium reasoning models already cost dozens of times more than standard versions. OpenAI’s GPT-5.5 Pro, for example, costs 150 times as much to process tokens as GPT-5.4 Nano, an older version. That gap is likely to widen further as ever more powerful models are released.
AI bills v gains
Many companies initially equated higher token usage with higher productivity, and paid the price for it. Amazon, for one, scrapped an internal AI leaderboard after employees gamed the system by assigning AI agents pointless tasks that inflated token consumption – a practice known internally as “tokenmaxxing”.
Token consumption is a terrible proxy for productivity. And that’s why the focus is shifting from volume to value. Uber, for instance, burned through its entire annual AI budget by April, prompting the ride-hailing business to impose caps on employees’ use of tokens.
To keep a lid on costs, firms need a fundamentally different mindset. First, governance must escape the IT department. AI can no longer be the sole preserve of chief information officers and chief digital officers. Procurement teams and even chief financial officers must understand what drives token consumption, what pushes bills higher and how quickly costs can spiral as usage accelerates.
Second, AI breaks the economics of traditional software. Every time it is used, the bill climbs, so forecasting demand is going to be key. But controlling costs is only half the battle, since companies also need to start using AI to generate real value.
Unfortunately, many are still using it too timidly. The common approach is to make existing tasks faster, like taking meeting notes, summarising documents or producing analysis that is a bit sharper. Those gains are decent, but they rarely scale enough to justify the outlay.
The bigger prize comes from revamping entire workflows around AI. Some are already doing this quite well: IKEA, for instance, deployed AI to automate nearly half of its customer calls. But, rather than laying off the 8,500 employees whose jobs had been displaced, the flat-pack retailer retrained them as interior designers, thereby turning a cost-saving exercise into a fresh revenue stream.
But rethinking the workflow is only the first step, as firms also need to rethink the models they use to run it.
You get what you pay for
One trap to watch out for is that the cheapest model is not always the most cost-effective option. Many companies obsess over the price of tokens rather than the cost of completing a workflow. That’s the wrong calculation.
A frontier model may cost many times more per token than a smaller or open-source alternative such as China’s DeepSeek. But if it finishes the job in fewer steps, makes fewer mistakes and demands less human intervention, it can still deliver the lower overall cost.
Increasingly, the best strategy will not be to deploy a single model but a portfolio of them. Frontier models should tackle the work that demands sophisticated reasoning, while cheaper models shoulder the routine tasks.
The implications of that bifurcation will stretch beyond the companies buying AI; it will create a new commercial dilemma for model builders.
In traditional software, heavy users are often the most profitable because the cost of serving them barely rises. Once the software is built, every additional user adds almost nothing to the bill. AI turns that logic on its head. Every prompt consumes computing power, meaning the biggest customers can also be the most expensive to serve.
Whether companies’ bets pay off remains far less clear. The evidence that AI is generating returns beyond merely cutting costs is still thin.
IBM said it was on track to generate $4.5 billion in savings last year owing to AI and automation. JPMorgan Chase says AI is generating about $2 billion in annual savings from an investment of roughly the same size. But there are still remarkably few companies that can demonstrate, at scale, how AI has translated into sustained improvements in productivity or profitability.
The first step is to identify the biggest AI costs. Which workflows rely on the most expensive models, and do you actually need them? Too many companies are putting Ferrari engines in lawn mowers, paying for frontier-level reasoning when a smaller, cheaper model would do the job just as well.
So five years from now, the companies pulling ahead of the pack will not be those that spend the most on AI. They will be the ones that understand their own workflows best and match the right model to the right task.
Get that wrong, and the costs quickly compound. Deploy frontier models where cheaper ones would suffice, and you waste money. Deploy cheaper models where sophisticated reasoning is needed, and you invite inefficiencies and errors that raise the risk of reputational damage.
The competitive edge will, ultimately, not come from spending more on AI. It will come from spending less on the wrong AI. The adoption race is over, but the cost war has just begun.
Amit Joshi is professor of AI, analytics and marketing strategy at IMD
