BNP Paribas AI chief ditches tokenmaxxing for dollars

Charles Holive, the chief AI officer at BNP Paribas CIB, is not playing the Silicon Valley token game. At Mistral AI's summit in Paris last week, he said his team does track token consumption, but he treats it as a cost-control input, not the headline measure of success. The real scorecard, he told Business Insider, is dollars and productivity. "We try to go away from vanity metrics - billions of tokens per day," Holive said. "We try to make sure that what we track is an outcome, not a vanity metric," he added.

That is a direct shot at the year-long obsession with "tokenmaxxing" in Silicon Valley, where maximizing AI usage, and therefore token consumption, has been treated by some as proof of productivity. Holive's version of the story is simpler and a lot less fashionable: "What did you do that you didn't do before? How much faster did you do it?" he said. In other words, if AI is not changing work or speeding it up, the token count is just a fancy receipt.

The timing matters because Holive's comments land as more companies are openly questioning whether AI spend is earning its keep. Amazon recently shut down an internal AI-use leaderboard after employees reportedly started doing tasks just to climb the rankings. Uber COO Andrew Macdonald has publicly questioned whether rising AI costs are translating into more useful products, while GitHub recently moved Copilot to usage-based pricing as AI bills continue to rise. The message for executives is hard to miss: AI adoption is no longer enough on its own, and the companies footing the bill want proof that the spend is moving something real.

Holive's job gives that argument extra weight. BNP Paribas CIB is the investment banking arm of one of Europe's largest banks, which means its AI program sits in a world where every initiative has to justify itself against revenue, risk, and productivity. For the projects he oversees, Holive said the process starts with explicit assumptions about how much revenue or productivity they could generate. His team then creates KPIs and checks progress monthly or quarterly against those goals. That is a very different posture from a loose internal experiment where teams are told to use the tool and see what happens. Here, the outcome is the point from the beginning.

That does not mean BNP Paribas is ignoring the plumbing. Holive said the bank still has dedicated teams monitoring token consumption so it can keep costs under control. "Then we look at token consumption because I need to control my costs," he said. That distinction is the whole story: tokens are useful for managing the bill, but they are not the business case. For AI leaders inside regulated financial firms, that difference matters because the pressure is not just to innovate, but to do it without spraying money around in a way that cannot be defended to the board.

Holive's comments also echo a broader pattern among enterprise executives who are trying to shift the AI conversation from usage to impact. Amit Kapur, chief AI and transformation officer at Tata Consultancy Services, told Business Insider that companies should focus on "business outcomes" and "business impacts" rather than "only token as one line item." Antoine Pichot, director of innovation, digital, and data at La Banque Postale, said his bank evaluates AI projects based on efficiency gains, customer satisfaction, employee satisfaction, and financial impact. That is a useful reminder that even when firms are all buying the same underlying technology, they are not necessarily defining success the same way. The metric becomes the strategy.

Holive did leave room for token usage as a useful signal in one place: adoption, especially in software engineering. But even there, he said his bank is not approaching AI like a popularity contest. "We didn't say, 'use the tool, do your best,'" he said. "We did the opposite of that," he said. That is the part leaders in other industries should probably pay attention to. The easy phase of AI was getting people to click. The harder phase is proving that the clicks created value, and in banks, that proof has to survive scrutiny from finance teams, operating leaders, and anyone else who gets handed the model bill. For peers in banking, tech, and any company now trying to justify AI budgets, Holive's framework is the blunt version of a coming standard: track the tokens if you must, but report the outcome if you want to keep the spend.