Like Highmark Health and Google Cloud, they utilize the AI ​​gene to improve medical claims and improve care: 6 key lessons

Among the numerous educational and surprising insightful panel discussions on the integration of AI Enterprise with the participation of industry leaders in VentureBeat’s Transform Conference 2025 This week was run by Google Cloud Platform vice president and technology director (CTO) Will Grannis AND Richard Clarke, High level healthSenior Vice President and Director for Data and Analysis.

This session “A new pile of artificial intelligence in healthcare: architecting for multimodal and multimodal environments “ He provided a pragmatic view on how both organizations cooperate to implement AI on a scale of over 14,000 employees in a enormous American health care system Highmark Health (based in western Pennsylvania).

In addition, cooperation in the deck of all these employees and turned them into lively users without losing the complexity, regulation or trust of clinicians.

So how did Google Cloud and Highmark go? Read on to find out.

Partnership based on prepared foundations

Highmark Health, an integrated system of payers-PROPRESSIONALLISTS supporting over 6 million members, uses AI Google Cloud models and infrastructure to modernize older systems, boost internal performance and improving patients’ results.

What distinguishes this initiative is the focus on platform engineering – treating artificial intelligence as a fundamental change in the scope of work, not just the next technological layer.

Richard Clarke, Director of Data and Analytical Highmark, emphasized the importance of early building malleable infrastructure. “There is nothing more in Cobol than the employment platform,” noted Clarke, but Highmark even integrated these systems with the AI ​​model based on the cloud. Result: up to 90% of load replication without system interference, enabling smoother transitions and real -time observations in intricate administrative processes.

Google Cloud Cto Brannis will repeat this success starts from scratch. “It may take three, four or five years,” he said, “but if your data is ready, you can run the experiments and grades loops that make artificial intelligence useful on a scale.”

From the proof of the concept for everyday utilize

Over 14,000 over 40,000 Highmark employees regularly utilize the company’s internal AI generative tools, powered by VERTEX AI and Gemini models from Google Cloud.

These tools are used in various cases of utilize – from generating personalized membership communication to downloading documentation to process claims.

Clarke emphasized an example on the supplier’s side covering the certificate and verification of the contract. Earlier, a personnel member manually searched many systems to verify the supplier’s readiness.

Now AI aggregates that the data, requirements for rapporteur and generate adapted output data-with quotes and contextual recommendations.

What drives this high adoption indicator? Combination of structured swift libraries, lively training and user feedback loop. “We just don’t throw tools and we hope that people use them,” explained Clarke. “We show them how their work is easier, and then scaling based on what contributes to adhesion.”

Agency architecture over chatbots

One of the most -running topics from the session in the future was the transition from the interaction based on chat to multi -magent systems capable of performing tasks from end to end. Grannis described this as a departure from a quick chat models in the direction of synthesis and automation of tasks.

“Think less about the chat interface and more about saying:” Go, bring it back and let me decide, “said Grannis. These funds coordinate many models, potentially casculating in various functions – from translation to research to work flow.

Highmark is currently piloting one -time agents for specific work flows, and the long -term goal is to set them in facilities for autonomous activities. This will reduce the need for many interfaces or connector and allow centralized control with a broader range.

First task, not model

Both speakers emphasized the key mental change for enterprises: stop starting with the model. Instead, start with the task and choose or organize models accordingly.

For example, Highmark uses Gemini 2.5 Pro for long, demanding research and twins Bligini to quickly interact in real time. In some cases, even classic deterministic models are used when they fit better into this task – such as translation of patients’ communication into many languages. As Branis put it: “Your business processes are your IP address. Think about completing the task and organize models for this.”

To support this flexibility, Google Cloud invests in routing and open standards. The last initiative of the agent protocol, introduced from Linux Foundation, was designed to promote interoperability and stability in the environments of many agents.

Practical advice for enterprise leaders in various sectors

For those who want to repeat the success of Highmark, panels offered specific tips:

1. Lie the Foundation early: Invest now in data readiness and system integration. Even if the full distribution of AI has been for years, the payment depends on the early basics.
2. Avoid building your own fundamental models: Unless your company Is Building models is steep. Focus on orchestration and refinement for specific cases of utilize.
3. Adopt the way of thinking of the platform: Centralize access to the model and tracking of utilize. Create a structure that supports experiments without devoting management.
4. Start with tasks, not tools: Define the result first. Then adapt it to the agent model or architecture that fits best.
5. Measure and share: Internal adoption increases when employees see practical results. Follow the utilize, capturing success history and constantly update the libraries of approved hints and flows.
6. Design, not just information: The future of AI Enterprise is to perform tasks, not a immobile insight. Build agents who can safely and safely cause real -world activities within your systems.

Looking to the future

While the partnership between Highmark and Google Cloud is still developing, the previous progress is offered by a model for other people in healthcare – and outside – who want to build scalable, responsible and very useful AI systems.

As Clarke summed up: “It’s not about flashy functions; it’s about what really helps people do their job better.”

Leaders of enterprises who left the session can comfort themselves in this: success in generative artificial intelligence is not reserved for people with the largest budgets, but for people with the most clear plans, malleable platforms and patience to build strategic.