Despite differences, two enterprises show AI success depends on scaling pilots into production with hybrid infrastructure, low latency, governance as shared constraints.
Two companies that have almost nothing in common at the surface level, CVS Health and Pfizer sat down together and ended up describing the same struggle.
The context was different, the workloads were different, but when you get right down to it, the arc was identical. For both, AI adoption lives or dies by infrastructure, and infrastructure now bends under the weight of adoption.
Alan Rosa, who wears both the CISO and SVP of Global Infrastructure hats at CVS, said that from his view. “eighty-five to ninety percent of all AI investments are currently failing,” closely echoing a Gartner statistic but adding some of his own math.
“We’ve had mixed results, as you can imagine, a lot of failure rates at the beginning,” he explains. “I definitely wouldn’t say we’re at 80% success rates. The truth is probably somewhere in the teens.”
In short, he was telling us what we’re hearing elsewhere. Adoption in the enterprise isn't a smooth curve. It’s pilots stacked on pilots, proof-of-concepts stalling, failures accumulating until something finally breaks through and survives at scale.
Pfizer’s Berta Rodriguez-Hervas, the company’s Chief AI and Analytics Officer, came at it from a different angle but landed in the same place.
“I believe the value never comes from the pilots. It comes from the pilots at scale… you get to learn then where to really apply and how to apply genAI in a way that can bring you value.” In her world, scaling means more than productivity, which can mean accelerating clinical development or broadening patient access. But the rhythm is the same, she says. Pilots don’t yield ROI, scaled workflows do.
Both execs described that transitional moment where AI shifts from proof to production when pressed by Lambda’s Robert Brooks who moderated the session at the AI Infra Summit recently.. At CVS, the experiments are already visible. Claims adjudication platforms are being accelerated by AI so that patients waiting on a decision for a drug authorization get an answer in seconds rather than days. Supply-chain systems are matching drugs to distribution channels faster, shaving precious time off delivery windows that can determine surgical readiness. Even the retail business is touched, with an intelligent IVR system routing calls without human intervention.
Pfizer’s deployments are less consumer-facing but no less dependent on scale. Internal workflows running on genAI, from back-office operations to research modeling, each one a candidate for real-world adoption if the compute and governance pieces align.
But for both, scaling is where infrastructure stops being an afterthought and starts dictating the pace of adoption.
Rodriguez-Hervas described it in terms of latency. “The more the workflows get complex, you cannot have a user waiting to get an answer,” she said. “Latency is also entering the conversation more and more, even for flows or applications that before there was a little bit more tolerance.” This is not the old tolerance of overnight jobs or even minute-scale responses. She thinks multi-step reasoning models and multi-agent orchestration chains change the definition of acceptable.
Rosa added the same point from the customer experience side. “High velocity, low latency applications require very specific hosting,” he said. “Our next generation data center will look very different than the ones that we’ve deployed for the last 15 years.” For CVS, the traditional playbook of generic datacenters and cloud-first strategies no longer fits. On-prem deployments remain mandatory for workloads touching patient data or requiring sub-second responsiveness and cloud absorbs the rest.
The result, according to both Pfizer and CVS Health views, is a hybrid model that is not aspirational but structural, dictated by the character of the applications themselves.
Not surprisingly given his role at CVS Health, Rosa also pointed to security. “Innovation always outstrips security,” he said. “But in this case, it’s not just transformational, it’s revolutionary. And so we’re going to have to catch up.” His checklist of concerns includes not just securing the data sources that train the models, but also controlling which internal or external models can be authorized for use, and enforcing the principles of trust and safety (explainability, transparency, bias mitigation) all at a scale that maps to the company’s regulatory exposure.
Rodriguez-Hervas gave her version of the same story, where the risk is exposure but also, as she describes, inequity. She detailed Pfizer’s new Health Answers chatbot, explaining, “We did put a lot of emphasis there to make sure that we were giving an unbiased answer… to make healthcare more accessible and represent better the different populations.”
For her, responsible AI means building guardrails that reduce bias in data, clean and homogenize inputs, and deliberately seek out representation for under-served groups in studies and trials.
While Rosa is worried about securing a fragmented infrastructure landscape against regulators, Rodriguez-Hervas is trying to ensure that the same infrastructure becomes a vehicle for equity. So in short, different angles, same binding constraint. Without governance and trust, nothing scales.
It was striking to hear two industries as different as pharmacy retail and pharmaceutical research converging so neatly on infrastructure as the true bottleneck.
Pfizer sees it in compute demand. “We do have a hybrid environment as well,” Rodriguez-Hervas said. “Something’s running on prem, something’s running on the cloud. What we are seeing is what is expected, an increased demand for cloud, increased queues… but we are seeing a consistently increased demand on the compute needs, which is a great sign for adoption of AI.”
Rosa echoed the sentiment from CVS: “Long gone are the days where we’re talking about cloud-first, cloud-only. We don’t talk that way anymore. We know it’s going to be a hybrid world.”
What’s interesting is that both acknowledged that hybrid infrastructure creates its own latency cliffs. You can throw GPUs at a problem, but a legacy 10-gig pipe can erase the value instantly.
“You are always gonna be as fast as your slowest,” Rosa said. The math is unforgiving, and it shows up in user experience when inference jobs stall. Rodriguez-Hervas added that the workflows themselves are evolving into multi-stage systems where even non-critical applications are suddenly latency-sensitive. Neither company can ignore those constraints if they expect adoption to grow.
And while it might sound like one of those intangibles in such a rooted conversation, the two leaders also touched on what all of this means for talent. Rodriguez-Hervas pointed to her own background at Nvidia and Tesla, part of a wave of high-tech talent now being drawn into biopharma. The infrastructure pivot has forced Pfizer to import platform engineering and systems skills that were not part of its DNA.
Rosa took a different tack, calling the new generation of practitioners “technical athletes.” The barriers to entry in AI development are rising, he said, requiring more formalized training and sharper aptitude than in the past.
“In a few years, you’re going to be truly working with the best people, and that’s where you’re going to grow the most,” he told the audience.
Both companies are building retraining programs for existing staff, both are open to non-traditional talent pipelines. Again, the business context differed, but the conclusion rhymed: the infrastructure stack is only as good as the people who can run and reinvent it.
By the end of the session, what stood out was not the contrast between pharmacy benefits management and drug discovery but the symmetry of the themes.
AI adoption is experimental and failure-heavy, but scaling pilots into production requires infrastructure ready for low-latency, high-density, hybrid deployments. And infrastructure expansion is necessary for adoption, but it multiplies risk, forcing new approaches to governance and responsible AI.
Both companies understand that the two tracks cannot be separated. Pilots test the art of the possible, infrastructure defines what actually works, and governance dictates what can be trusted at scale.
