Token Bills Arrive: Meta, PwC and Jack Henry Recalibrate AI Spend

Microsoft Copilot licenses for all 7,300 associates at Jack Henry & Associates would add $3 million in costs, CFO Mimi Carsley said. That single figure now anchors the enterprise AI conversation. The Information's finance editor, Meredith Mazzilli, assembled three executives to explain what changed: Karen Young, U.S. board chair at PwC; Carsley of Jack Henry; and Heidi Rhodes, global director of enterprise customer success at Meta Platforms.

What ended the free-spending phase?

Meta ran the most permissive experiment in the group. Mark Zuckerberg declared AI adoption a company-wide mandate. Rhodes recalled his framing: "This is going to be the year that everyone is going to adopt AI." Employees were told, in her words, to "figure out what works and go fail."

That freedom worked, up to a point. Meta identified which workflows benefited from AI and standardized them. It also burned tokens without counting them, a luxury Rhodes admitted most companies cannot afford.

"In 2023, it was: Any single problem that you have, rub some AI on it," she said. "Then we got our token bills. Now organizations are saying, 'Unless you can show me real ROI, we can end the conversation there.'"

The message is blunt. Inference costs converted AI from a strategic virtue into a budget line. Nothing in the discussion suggests Meta has retreated from adoption. It simply stopped subsidizing unfocused usage.

How does a regulated company govern AI?

Jack Henry took the opposite path. Operating in financial services, it built governance first. Carsley said the company established guardrails for assessing and approving AI tools and defined what data employees could access. Only then did deployment follow.

The $3 million Copilot estimate forced a hard discipline. Jack Henry reviews license usage continuously.

"If someone's not using their license, we'll go talk to them and coach them up," Carsley said. "If they're still not using it, we actually seize it back and redeploy it elsewhere in the organization."

A license, in this model, is a recoverable asset. Unused seats represent measured waste, not an acceptable overhead.

PwC faces a different constraint. Karen Young said the firm has no shortage of AI ideas from its own staff. The filter is the client. "We think about the solutions our clients are looking for and how we can scale [our AI tools] for their benefit," she said. PwC has also moved to training employees on usage discipline: "We lead them to use the right tool for the right model for the right reason."

Notably absent from all three accounts: any disclosed dollar return on the AI spend itself.

Does efficiency actually convert to value?

Young cited one concrete client result: 14 finance processes transformed, with forecasting and scenario planning cut from weeks to minutes. Her follow-up question is the one that matters. "What are those employees in finance now doing to be more strategic with their time?"

Carsley pressed the same point with a sharper edge. "If someone's gonna use AI to draft an email and you save five minutes, like, did you go get a cup of coffee or did you do something with the five minutes?" she asked. "If you're not redeploying that savings to another productivity or efficiency, did you really gain anything by the spend?"

Five saved minutes is not a return. It is an input to one. The claim of productivity gains survives only if the reclaimed time lands somewhere auditable.

Rhodes cited a client whose sales team became 20% to 30% more productive once AI absorbed hours of research. She framed the fork that follows: cut headcount, or hire against the new capacity. "Do you make a reduction in humans? Or now that you have one of the world's most efficient salespeople, do you go out and hire more?" Her vote went to hiring.

That is a preference, not evidence. No one on the panel quantified what the redeployed capacity produced.

Where should the next dollar go?

Carsley offered the closing rule. "Ask yourself, what do you have an advantage in? Then apply AI to that as an accelerant," she advised. "Do the thing you do best as an organization, then turn it up 10% to 20%."

The pattern across all three companies is consistent. Blanket adoption is over. Licensing, token consumption and reclaimed time are now managed with the same scrutiny as any other input cost, and the next wave of funding will track demonstrated ROI rather than enthusiasm.