How CIO consider AI fashions

Anthropic captured the lead in U.S. enterprise AI adoption, securing 34.4% of the market in comparison with OpenAI’s 32.3%, in accordance with Ramp 2026 AI Index in Might. Amongst companies adopting AI for the primary time, Anthropic is profitable roughly 70% of head-to-head matchups.

“Anthropic gained early with the individuals who have been already attempting AI inside corporations,” stated Ara Kharazian, lead economist at Ramp. He stated these early adopters have been “a superb main indicator of the place the remainder of the market goes.”

The shift marks a milestone within the AI platform wars. However for CIOs making shopping for choices, the leaderboard has by no means been the principle occasion.

“I’d watch out on the head-to-head,” stated Phil Leslie, chief expertise and innovation officer at Cornerstone Analysis, an financial and monetary consulting agency that helps high-stakes litigation. “The variations among the many main frontier fashions are actual however slim, they usually maintain shifting. The extra helpful query just isn’t ‘which mannequin is greatest this quarter’ however ‘which setup lets us swap because the frontier shifts.'”

Associated:Safety is slowing autonomous AI; how CIOs are responding

That pragmatism is widespread. Because the AI mannequin race intensifies, CIOs say they’re focusing much less on selecting a winner and extra on constructing architectures that do not pressure them to.

Safety first, every thing else second

Earlier than efficiency, pricing, or options enter the dialog, CIOs say AI platforms should first go safety and governance exams.

“Our work depends upon strict shopper confidentiality, so a platform has to clear a number of non-negotiable bars earlier than it’s even a candidate,” Leslie stated. “Shopper knowledge is rarely used to coach fashions, interactions aren’t uncovered to human evaluation, and knowledge stays on U.S. infrastructure. These constraints outline the possible set; every thing else is a alternative inside it.”

Eric Tempo, head of AI at Cox Enterprise, the business providers division of Cox Communications, put it equally.

“Safety is non-negotiable for us,” Tempo stated. “Given the quantity of important infrastructure we handle, we’ve to begin with danger and whether or not an answer matches inside our safety, authorized, knowledge privateness and governance necessities. If it introduces dangers that we’d not be comfy standing behind, it’s not value pursuing.”

At Lowenstein Sandler, a nationwide legislation agency, the bar is simply as excessive. “Safety and confidentiality aren’t one issue amongst a number of,” stated the agency’s chief data and innovation officer, Maureen Naughton. “They are the brink take a look at.”

The mannequin is not the moat

With safety and governance necessities met, CIOs say efficiency issues — however not in the way in which a lot vendor advertising suggests. The hole between main fashions is skinny and continuously shifting. Betting on in the present day’s benchmark winner is a short-term play at greatest.

Associated:Outsourcing contracts weren’t constructed for AI. CIOs are renegotiating now

“The sturdy benefit was by no means the mannequin — the fashions are the simple half,” stated Jeremy Bruck, associate at administration and expertise consulting agency West Monroe. “The benefit is in an organization’s knowledge belongings, context, workflows, controls, and how briskly they can flip a sign into motion.”

Cornerstone Analysis constructed a model-agnostic stack on goal. “The frontier is shifting too quick to wire our structure to any single vendor; the lock-in danger is actual, and the hole that appears decisive in the present day could also be gone in two quarters,” Leslie stated.

Somewhat than selecting a single winner, a number of CIOs describe operating a portfolio of AI platforms matched to completely different use instances.

A number of fashions serving distinct roles

“We do not strategy this as selecting a single winner,” stated Naughton of Lowenstein Sandler. “We consider them as occupying distinct lanes relatively than competing for one seat.”

This strategy is turning into the norm. “The continual mannequin leapfrogging has helped corporations settle for that the speed of change is barely going to speed up,” stated West Monroe’s Bruck. “Corporations are now not targeted on ‘smartest mannequin’ however as an alternative on modular platforms that scale back switching prices as new options emerge.”

Associated:The true warmth behind OpenAI’s new Jalapeño chip

Ramp’s Kharazian sees the identical sample in spending knowledge. “The analysis is shifting from ‘Which AI vendor ought to we use?’ to ‘Which mannequin ought to do that job, at what value?'” he stated. “That pushes corporations towards multi-vendor setups, routing, open supply fashions, and inference platforms.”

Freedom inside a framework

Developer enthusiasm has been the driving pressure behind AI adoption. However CIOs say that bottom-up demand works greatest when there are AI guardrails in place.

Tempo of Cox Enterprise describes the strategy as “freedom inside a framework.” The corporate offers a ruled set of mannequin choices throughout on-prem and cloud environments, and inside that framework, groups have flexibility to decide on what works greatest for his or her use case.

“Developer choice does play a task, significantly in mannequin choice,” Tempo stated. “The secret is that every one of this occurs inside a broader governance framework that we view as an enabler, not a constraint”

The dynamic is analogous at Cornerstone Analysis. “Our knowledge scientists are robust and opinionated; they run their very own evaluations and maintain actual views on mannequin efficiency,” Leslie stated. “That enter is indispensable.”

However a full analysis additionally requires IT to evaluate safety and authorized to vet confidentiality phrases — and people evaluations do not occur on their very own. “The job of management is to not overrule the technical judgment effervescent up from the crew,” Leslie argued. “It is to ensure the entire analysis really occurs.”

At Lowenstein Sandler, Naughton frames it as a design precept: “The healthiest model of that is governance setting the boundaries,” she stated, “and the entrance line driving the priorities.”

Watching the meter: How CIOs handle consumption

Governance solves one downside. Prices create one other.

“AI is the fastest-growing spend class we have ever noticed,” stated Ramp’s Kharazian. “The common enterprise is spending 13x extra on tokens than it was in January 2025.”

Utilization-based pricing means CIOs can see precisely what they’re spending, however not all the time what they’re getting for it. As adoption scales, CIOs are growing new disciplines for managing consumption with out stifling worth.

The primary transfer, stated West Monroe’s Bruck, is tying spend to outcomes. “You’ll be able to’t handle what you may’t attribute,” he saids, “so disciplined enterprises monitor which crew, workflow, and more and more which individual is driving spend, tied to a unit of enterprise worth, earlier than they attempt to spend much less.”

Tempo takes a value-first strategy to value administration at Cox Enterprise. “When AI helps somebody work by way of years of backlog in weeks, the dialog shifts from controlling value to asking what extra we are able to allow,” he stated.

Kharazian sees three traits rising:

“Most corporations aren’t attempting to sluggish AI adoption,” Tempo stated. “They’re attempting to get the upside with out letting the invoice run away.”

What CIOs are asking now: AI mannequin analysis shifts from outputs to actions

A 12 months in the past, CIOs have been questioning whether or not AI instruments really labored. That query is essentially settled.

“Our standards have matured towards integration depth, safety posture, governance match, and more and more readiness for extra agentic capabilities — which means instruments that take actions relatively than simply generate textual content,” stated Naughton. “The bar has moved from novelty to sturdiness.”

When AI methods can take actions, the analysis adjustments completely.

“As soon as a system can search the net, name exterior providers, write and run its personal code, you need to ask a distinct class of query,” stated Cornerstone’s Leslie. “What guardrails exist? What can we cease it from doing inside our surroundings?”

The previous take a look at was about what the mannequin produces. “The brand new take a look at can also be about what the mannequin does,” Leslie stated. “That could be a genuinely more durable downside.”

Mannequin choice nonetheless issues. However for CIOs, so does constructing structure that may adapt when the leaderboard flips once more — as a result of it would.