How Highmark Health and Google Cloud are using Gen AI to streamline medical claims and improve care: 6 key lessons

Among the many educational group discussions and surprisingly insightful on IA business integrations featuring industry leaders in Venturebeat conference transform 2025 This week was a Google Cloud The Vice-President of the Platform and Head of Technology (CTO) Will Grannis And Richard Clarke, High quality healthsenior vice-president responsible for data and analysis.

This session, “The new AI battery in health care: architecture for multimodal and multimodal environments “,” Delivered a pragmatic overview of how the two organizations collaborate to deploy large -scale AI in more than 14,000 employees in the large American health system Highmark Health (based in the western Pennsylvania).

In addition, collaboration has integrated all these employees and transformed them into active users without losing sight of the complexity, regulations or confidence of clinicians.

So how did Google Cloud and Highmark proceed? Read the rest to discover it.

A partnership built on prepared foundations

Highmark Health, an integrated payer system serving more than 6 million members, uses Google Cloud’s models and infrastructure to modernize inherited systems, stimulate internal efficiency and improve patient results.

What distinguishes this initiative is the emphasis on engineering of platforms – the treatment of AI as a fundamental change in the way the work is done, not just another technological layer.

Richard Clarke, Highmark data and analysis manager, stressed the importance of building flexible infrastructure early. “There is nothing more inherited than an employment platform coded in Cobol,” said Clarke, but Highmark even integrated these systems with AI models based on the cloud. The result: up to 90% replication of the workload without systemic disruption, allowing higher transitions and real -time information on complex administrative processes.

Google Cloud CTO WILL GRANNIS has echoed that success begins with the basics. “This can take three, four, five years,” he said, “but if your data is ready, you can execute the experimentation loops and the assessments that make AI useful on a large scale.”

From proof of concept to daily use

More than 14,000 of the 40,000 Highmark employees regularly use the internal generative tools of the company, powered by the models of Green and Gemini from Google Cloud.

These tools are applied in a range of use cases – from the generation of personalized communications from members to the recovery of documentation for the processing of complaints.

Clarke highlighted an example on the supplier side involving accredications and a contract verification. Previously, a staff member would manually seek several systems to check the preparation of a supplier.

Now, AI brings together this data, the cross -country requirements and generates the custom output – with quotes and contextual recommendations.

What motivates this high adoption rate? A combination of rapid structured libraries, active training and user feedback loops. “We are not just having tools and hope people will use them,” said Clarke. “We show them how it facilitates their work, then the scale according to what obtains the traction.”

Agent architecture on chatbots

One of the most prospective themes of the session was the transition from cat-based interactions to multi-agent systems capable of finishing end-to-end tasks. Grannis has described it as a distance from fast response cat models to the synthesis of tasks and automation.

“Remember less of having a cat interface and saying:” Go do this, bring it back and let me decide “,” said Grannis. These agents coordinate several models, potentially in cascade on different functions, from translation to research to the execution of the workflow.

Highmark is currently piloting single -use agents for specific workflows, and the long -term objective is to integrate them into backend systems to carry out actions independently. This will reduce the need for several interfaces or connectors and will allow centralized control with a wider range.

Task first, not model first

The two speakers highlighted a key mental change for companies: stop starting with the model. Instead, start with the task and select or orchestrate the models accordingly.

For example, Highmark uses Gemini 2.5 PRO for long intensity of research and a gemini flash for rapid and real -time interactions. In some cases, even conventional deterministic models are used when they are better suited to the task, such as translating patients in several languages. As Grannis said, “your business processes are your IP. Remember to perform a task and orchestrate the models to do this. ”

To support this flexibility, Google Cloud is investing in model routing capacities and open standards. The recent agent protocol initiative, introduced with the Linux Foundation, is designed to promote interoperability and stability in multi-agent environments.

Practical advice for business leaders in all sectors

For those who seek to reproduce the success of Highmark, the panelists have offered concrete advice:

1. File the bases early: Invest in data preparation and system integration now. Even if AI’s full deployment is in the years, the gain depends on the first works.
2. Avoid building your own fundamental models: Unless your business East Construction models, it is prohibitive costs. Focus on orchestration and fine adjustment for specific use cases.
3. Adopt a platform state of mind: Centralize access to the model and monitoring of use. Create a structure that supports experimentation without sacrificing governance.
4. Start with tasks, not tools: Define the result first. Then associate it with the agent model or architecture that is best suitable.
5. Measure and share: Internal adoption develops when employees see practical results. Follow the use, capture success stories and permanently update the approved prompted and flow libraries.
6. Action design, not just information: The future of the company AI is the execution of tasks, not static information. Build agents who can trigger real actions safely and safely in your systems.

Ahead

Although the partnership between Highmark and Google Cloud is still evolving, progress has so far offered a model for others in health care – and beyond – who want to build evolutionary, responsible and very usable AI systems.

As Clarke summed up, “these are not flashy features; This is what really helps people do their job better. ”

Business leaders who have missed the session can comfort themselves in this area: the success of the generating AI is not reserved for those who have the biggest budgets, but for those who have the clearest plans, flexible platforms and the patience to build strategically.