"The Wrong Objective": BNY's CEO Just Told Corporate America It Wasted Two Years on AI

"Maximizing token usage is the wrong objective."

That was BNY CEO Robin Vince in a CNBC interview this month, a line his bank later amplified on LinkedIn. It sounds like a throwaway remark. It isn't. Vince runs America's oldest bank, an institution that built its own AI platform, signed a multi-year deal with OpenAI, and just reported record quarterly revenue while crediting AI for part of the gains. When a CEO with that track record dismisses the metric everyone has been chasing for two years, it's worth asking what he sees that the rest of the market missed.

Call the last two years the tokenmaxxing era.

How token consumption became a vanity metric

Since late 2023, token consumption has served as the enterprise world's proxy for AI maturity. The logic seemed reasonable at the time. Tokens are the unit of account for large language models, so more tokens burned must mean more AI happening. Companies handed out subscriptions in bulk and pushed employees to use them. As VC Cafe noted in its own autopsy of the trend, Shopify went as far as factoring AI usage into performance reviews, and across Silicon Valley more prompts, more agents and more tokens per employee were read as proof a company was becoming AI native. Boards asked CIOs for usage dashboards, usage went up, and the dashboards looked great.

The flaw is that token consumption is an input, not an outcome. It measures activity rather than value, the corporate equivalent of judging a sales team on calls dialed instead of deals closed. In mid-2026, the bill for that confusion came due.

Karp lights the match

On July 1, Palantir CEO Alex Karp went on CNBC's Squawk Box and attacked the token-based business model of the frontier labs, saying that "something has gone completely wrong" with how AI is sold to enterprises. In his telling, companies have burned enormous sums on metered token consumption with little to show for it, while potentially handing their proprietary data and competitive edge to the model providers. A day earlier, Palantir had published a nine-point "AI sovereignty" manifesto on X that named tokenmaxxing directly and criticized it as a business model, a sequence documented in detail by Digital Applied. Palantir's stock climbed roughly 8% on the interview.

Karp is obviously talking his book. Palantir sells the layer that sits between foundation models and enterprise value, and a fresh open-weight partnership with Nvidia gives him every commercial reason to disparage metered access to closed models. Still, his critique resonated widely enough to move markets, which suggests he touched a real nerve. Quartz reported around the interview that Uber and Microsoft have capped or restricted employee access to expensive AI coding tools after budgets blew out, and that a growing number of enterprise customers are demanding clearer returns or switching to cheaper open-weight alternatives.

The pressure is visible on the vendor side too. Palo Alto Networks CEO Nikesh Arora told CNBC the following week that token costs should fall as much as 90% within two years for enterprise deployment to make economic sense. Even OpenAI has changed its pitch. Sam Altman appeared on CNBC in early July promoting not raw capability but token efficiency, claiming the company's newest model is 54% more efficient on agentic coding tasks. When the company that sells tokens starts advertising how few of them you'll need, the tokenmaxxing era is over.

The BNY counterexample

Vince's comment carries weight because BNY is not an AI skeptic backing away from the technology. The bank is one of the most aggressive AI adopters in global finance, and one of the few that can point to measurable results.

BNY moved within weeks of ChatGPT's launch and built Eliza, a proprietary multi-agent platform named for Elizabeth Schuyler Hamilton, wife of the bank's founder. As Funds Society reported from BNY's INSITE26 conference in June, Eliza connects to the major frontier models and serves nearly 50,000 employees across three levels: individual productivity, automation of complex operational processes, and AI solutions offered directly to clients. Roughly 15,000 employees have built their own AI agents and the firm has been developing well over a hundred AI-powered solutions across the organization, according to Banking Dive. In February, TIME reported, BNY deepened the bet with a multi-year OpenAI partnership that brings advanced reasoning models and tools like Deep Research into the platform.

The results, at least by BNY's own accounting, are real. American Banker reported that the bank posted record quarterly revenue of $5.7 billion in Q2 2026, up 13% year over year, while headcount fell 7%, with executives crediting AI-driven productivity for part of the operating leverage. Notably, CFO commentary in the same report put the firm's AI spending at a modest level relative to its roughly $4 billion engineering budget. BNY is not winning at AI by outspending anyone on tokens. It is winning on something harder to put in a dashboard.

Vince has been consistent about what that something is. "Ultimately, adoption and embedding in a company is going to be the differentiator," he said alongside the Q2 results, per American Banker. At Money20/20 last fall he made the same point from the other direction, arguing per Banking Dive that the binding constraint on AI is not the technology's capability but human adoption, culture, and organizational inertia. His internal mantra, "AI is for everyone, everywhere, everything at BNY," sounds on the surface like tokenmaxxing. It is the inverse. It is a culture metric, not a consumption metric. The question is not how many tokens the workforce burns. The question is whether AI has changed how the work actually gets done.

What replaces the token dashboard

If maximizing token usage was the wrong objective, what is the right one? The executives navigating this shift keep pointing to the same few things.

The first is outcomes over usage. Revenue per employee, cycle-time reduction, error rates, client-facing capabilities shipped. These are the numbers BNY put in its earnings, and they are the numbers that survive contact with a CFO.

The second is embedding over experimentation. Vince's framing implies AI woven into workflows and processes rather than a chatbot sitting next to them. The 15,000 employee-built agents at BNY matter less as a count than as evidence the tools have crossed from novelty into infrastructure.

The third is efficiency as a feature. With Altman marketing token efficiency and Arora demanding steep price cuts, the vendor ecosystem itself is repricing around value per token rather than volume. Enterprises that signed contracts during the tokenmaxxing era should expect better unit economics, and should ask for them.

The fourth is ownership and control. Strip away the theatrics and Karp's sovereignty argument raises a fair question every AI customer should be able to answer: who keeps the data, who captures the learning, and where does the competitive advantage end up?

The uncomfortable middle ground

Token consumption was never meaningless. In the early experimentation phase, heavy usage genuinely correlated with learning, because the organizations that burned tokens fastest also discovered use cases fastest. The tokenmaxxing era, wasteful as it was, functioned as a large-scale training program for the workforce, paid for by employers. The mistake was letting a phase-one metric persist into phase two.

The companies best positioned now treated token burn as tuition rather than the diploma. BNY spent modestly, embedded deeply, and can show the results in its income statement. That, and not a usage dashboard, is what winning at enterprise AI looks like in 2026.

Vince's ten words may end up being remembered as the moment the industry said so out loud.