Companies tighten their AI bets amid costs
After several years of widespread adoption, companies are finally looking at the AI bill. Between governance, licenses, measurable gains, and cash flow trade-offs, the central criterion is once again return on investment.
Companies that have expanded their use of AI are beginning to take stock, and the question is no longer just about adoption, but about choosing what to fund. In discussions reported with PwC, Jack Henry & Associates, and Meta Platforms, return on investment, governance, and compute consumption emerge as the new filters. In the background, the market is shifting from a generous experimentation phase to a much stricter arbitration logic. The assessments, however, do not like slogans.
When enthusiasm gives way to calculators
At Meta Platforms, Heidi Rhodes describes an initial phase where the instruction was simple: test, fail, learn. Mark Zuckerberg had made AI adoption an internal priority, to the point of telling teams: "This is going to be the year that everyone is going to adopt AI." Then the token bills, in other words the costs related to model usage, reminded everyone that AI is not free. "In 2023, it was: Any single problem that you have, rub some AI on it. Then we got our token bills," she summarizes.
The signal is now clearer: "Unless you can show me real ROI, we can end the conversation there." This phrase says a lot about the state of the market. Adoption remains broad, but tolerance for blind spending is decreasing. Financial departments now demand proof of concrete impact, not just appealing demonstrations.
Dans les secteurs régulés, la prudence vient avant l’outil
Jack Henry & Associates, active in financial services, started by setting safeguards. Mimi Carsley explains that the company first defined approval rules for AI tools and the data accessible to employees. This sequence is logical in a regulated sector, where the risk of leaks, non-compliance, or poorly controlled use is as important as productivity gains.
The cost remains very concrete. Providing Microsoft Copilot licenses to all 7,300 employees could add 3 million dollars in expenses, according to Carsley. The company therefore regularly reviews access. "If someone’s not using their license, we’ll go talk to them and coach them up. If they’re still not using it, we actually seize it back and redeploy it elsewhere in the organization," she says. Enterprise AI here joins an old finance habit: nothing is lost, everything is reallocated.
At PwC, the problem is not a lack of ideas, but sorting. Karen Young explains that teams test many use cases, but the trigger for funding remains the client and the ability to scale the tool. The internal guideline is clear: "We lead them to use the right tool for the right model for the right reason." In other words, not all tasks deserve the same model or the same bill.
The real test is not time saved, but what you do with it
Measuring gains is becoming more sophisticated. An example cited by PwC shows a client who transformed 14 finance processes, with forecasts and scenarios reduced from several weeks to a few minutes. However, the real issue is not just speed. Karen Young asks what employees do with this freed-up time: more analysis, more strategy, or just a longer coffee break?
Mimi Carsley poses the same question from a budgetary perspective. If AI saves five minutes on writing an email, the value depends on how that time is reallocated. "If you’re not redeploying that savings to another productivity or efficiency, did you really gain anything by the spend?" she says. This logic changes the discussion in executive committees: a productivity gain is only useful if it turns into something else; otherwise, it remains a nice but useless counter.
Heidi Rhodes goes further and warns against a too quick reflex to reduce staff. She cites a client whose sales team became 20% to 30% more productive after AI took over hours of research. The question then becomes as much political as economic: should teams be reduced, or should this efficiency be used to sell more? Her answer clearly leans toward the latter option.
Le calcul compute devient la ligne de fracture de l’IA
As AI grows in power, the issue of compute takes a central place. At Anthropic, Dario Amodei described this item as "the deepest business dilemma" of his company. The problem is simple to state, difficult to solve: capacity must be purchased one to two years before knowing if the revenue will justify it. Too little compute, and clients or new products lack resources. Too much compute, and the company puts itself in financial danger.
According to Bloomberg, Anthropic’s commitments would reach nearly 901 billion dollars, including about 71 billion due by September 2027 [to be verified]. The company wants to strengthen its cash cushion through its IPO, presented as a way to absorb the possible gap between expenses and revenues. In this context, the IPO serves less to finance a classic expansion than to buy time. It’s less glamorous than a product launch, but more reassuring than a shaky line of credit.
Amodei also explained that the company limits its initial capacity purchases and can supplement with the spot market if demand exceeds expectations. This strategy works like insurance, he says. It reduces the risk of being stuck with too few resources while avoiding overinvestment before the market confirms the trajectory. The subtext remains clear: generative AI promises a lot, but its industrialization still faces the same constraint, that of capital to immobilize before the market verdict.
Ce que ce virage dit du marché
Testimonies converge on one point: AI is entering a tougher execution phase. Companies no longer just want to know if the tool works, but how much it costs, who benefits, and how value is reinjected into the business. The debate no longer pits enthusiasts against skeptics; it pits uses that truly transform the organization against those that consume budget for unclear benefit.
In this context, three criteria already dominate: governance, measurement of return, and the ability to absorb compute costs. Companies that can connect these three dimensions will retain room to maneuver. Others risk discovering, a little too late, that AI loves demonstrations. Bills, however, prefer proof.