Among the many quite a few instructional and startlingly insightful panel discussions on AI enterprise integrations that includes business leaders at VentureBeat’s Remodel 2025 convention this week was one led by Google Cloud Platform Vice President and Chief Know-how Officer (CTO) Will Grannis and Richard Clarke, Highmark Well being’s Senior Vice President and Chief Information and Analytics Officer.
That session, “The New AI Stack in Healthcare: Architecting for Multi-Mannequin, Multi-Modal Environments,” delivered a realistic take a look at how the 2 organizations are collaborating to deploy AI at scale throughout greater than 14,000 staff on the massive U.S. healthcare system Highmark Well being (primarily based out of Western Pennsylvania).
As well as, the collaboration has onboarded all these staff and turned them into energetic customers with out dropping sight of complexity, regulation, or clinician belief.
So, how did Google Cloud and Highmark go about it? Learn on to seek out out.
A Partnership Constructed on Ready Foundations
Highmark Well being, an built-in payer-provider system serving over 6 million members, is utilizing Google Cloud’s AI fashions and infrastructure to modernize legacy methods, enhance inner effectivity, and enhance affected person outcomes.
What units this initiative aside is its concentrate on platform engineering—treating AI as a foundational shift in how work will get performed, not simply one other tech layer.
Richard Clarke, Highmark’s Chief Information and Analytics Officer, emphasised the significance of constructing versatile infrastructure early. “There’s nothing extra legacy than an employment platform coded in COBOL,” Clarke famous, however Highmark has built-in even these methods with cloud-based AI fashions. The consequence: as much as 90% workload replication with out systemic disruption, enabling smoother transitions and real-time insights into advanced administrative processes.
Google Cloud CTO Will Grannis echoed that success begins with groundwork. “This will likely take three, 4, 5 years,” he mentioned, “but when your information is prepared, you may run the experimentation loops and evaluations that make AI helpful at scale.”
From Proof-of-Idea to Each day Use
Greater than 14,000 of Highmark’s 40,000+ staff frequently use the corporate’s inner generative AI instruments, powered by Google Cloud’s Vertex AI and Gemini fashions.
These instruments are utilized throughout a spread of use circumstances — from producing customized member communications to retrieving documentation for claims processing.
Clarke highlighted a provider-side instance involving credentialing and contract verification. Beforehand, a employees member would manually search a number of methods to confirm a supplier’s readiness.
Now, AI aggregates that information, cross-checks necessities, and generates tailor-made output — full with citations and contextual suggestions.
What drives this excessive adoption price? A mixture of structured immediate libraries, energetic coaching, and person suggestions loops. “We don’t simply drop instruments in and hope folks use them,” Clarke defined. “We present them the way it makes their work simpler, then scale primarily based on what will get traction.”
Agentic Structure Over Chatbots
Some of the forward-looking themes from the session was the shift from chat-based interactions to multi-agent methods able to finishing duties end-to-end. Grannis described this as a transfer away from quick-response chat fashions towards activity synthesis and automation.
“Suppose much less about having a chat interface and extra about saying: ‘Go do that, convey it again, and let me determine,’” Grannis mentioned. These brokers coordinate a number of fashions, doubtlessly cascading throughout completely different capabilities—from translation to analysis to workflow execution.
Highmark is presently piloting single-use brokers for particular workflows, and the long-term objective is to embed these inside backend methods to carry out actions autonomously. This can cut back the necessity for a number of interfaces or connectors and permit centralized management with broader attain.
Activity-First, Not Mannequin-First
Each audio system emphasised a key psychological shift for enterprises: cease beginning with the mannequin. As a substitute, start with the duty and choose or orchestrate fashions accordingly.
For instance, Highmark makes use of Gemini 2.5 Professional for lengthy, research-intensive queries and Gemini Flash for fast, real-time interactions. In some circumstances, even traditional deterministic fashions are used once they higher swimsuit the duty—akin to translating affected person communications into a number of languages. As Grannis put it, “Your small business processes are your IP. Take into consideration fulfilling a activity, and orchestrate fashions to do this.”
To help this flexibility, Google Cloud is investing in model-routing capabilities and open requirements. The current Agent Protocol initiative, launched with the Linux Basis, is designed to advertise interoperability and stability in multi-agent environments.
Sensible Recommendation for Enterprise Leaders Throughout Sectors
For these seeking to replicate Highmark’s success, the panelists supplied concrete steering:
- Lay the muse early: Put money into information readiness and system integration now. Even when full AI deployment is years away, the payoff is determined by early groundwork.
- Keep away from constructing your personal foundational fashions: Except your online business is constructing fashions, it’s cost-prohibitive. Give attention to orchestration and fine-tuning for particular use circumstances.
- Undertake a platform mindset: Centralize mannequin entry and utilization monitoring. Create a construction that helps experimentation with out sacrificing governance.
- Begin with duties, not instruments: Outline the end result first. Then match it with the mannequin or agent structure that matches greatest.
- Measure and share: Inner adoption grows when staff see sensible outcomes. Observe utilization, seize success tales, and repeatedly replace libraries of authorized prompts and flows.
- Design for motion, not simply data: The way forward for enterprise AI is activity execution, not static perception. Construct brokers that may set off real-world actions safely and securely inside your methods.
Trying Forward
Whereas the partnership between Highmark and Google Cloud remains to be evolving, the progress thus far gives a mannequin for others in healthcare—and past—who need to construct scalable, accountable, and extremely usable AI methods.
As Clarke summed up, “It’s not about flashy options; it’s about what truly helps folks do their jobs higher.”
Enterprise leaders who missed the session can take consolation on this: success in generative AI isn’t reserved for these with the largest budgets, however for these with the clearest plans, versatile platforms, and the persistence to construct strategically.
