As enterprises move from isolated AI deployments to scaled, system-wide intelligence, the challenge is no longer building models, but coordinating them. The rise of agentic AI is introducing a new layer of complexity, where autonomous systems operate across workflows, make decisions, and interact with other agents. Without orchestration, this does not create intelligence. It creates fragmentation.

In oncology, where patient journeys span diagnostics, treatment, and long-term care, disconnected AI systems can quickly erode both efficiency and trust. In this final installment of a three-part series, Nasim Eftekhari, Chief AI and Analytics Officer at City of Hope, speaks with Erik Pupo, Director of Commercial Health IT at Guidehouse, on what it takes to move from AI experimentation to coordinated intelligence at scale.

In Part 1, Eftekhari unpacked how oncology AI is evaluated in real-world settings, where it is already delivering impact, and how organizations should define success versus failure. In Part 2, she detailed how AI is being embedded into clinical workflows, where it is delivering value today, and what it takes to scale it responsibly.

This final part brings the conversation forward, focusing on the systems-level challenge: how to orchestrate multimodal intelligence, coordinate AI agents across ecosystems, and build the data foundation required to make it all work in practice.

Multimodal AI: The next breakthrough in oncology intelligence

A central theme of the conversation is the rapid evolution of multimodal AI, systems capable of analyzing and reasoning across multiple types of data simultaneously. “It is the next big thing and maybe also the best thing that is coming down the pipeline, not just for oncology, but for AI in general,” says Eftekhari.

Today’s AI systems have already demonstrated value in processing structured electronic health record data and unstructured clinical text. But the next leap lies in combining these with imaging, genomics, and even socioeconomic or sensor data to form a more complete picture of the patient.

“When we say multimodal, it means discovering patterns or comprehending images, genomics, texts, EHR data, all at the same time,” she explains.

The implications are significant. Instead of isolated insights, clinicians could receive integrated recommendations that account for disease progression in imaging, genetic mutations, treatment history, and available clinical trials, all within a unified model.

Agent orchestration: Avoiding AI chaos in the enterprise

As AI adoption accelerates, another challenge is emerging: the proliferation of disconnected AI agents across enterprise systems. At City of Hope, Eftekhari and her team are proactively addressing this through the development of Hope LLM, a framework designed to orchestrate multiple AI agents across internal and external ecosystems.

“In the next couple of years, if we do not have the connective tissue or nervous system that brings together all these agents, we will not have intelligence. We will have chaos,” she warns.

Healthcare organizations are increasingly deploying AI across platforms such as electronic health records, CRM systems, and cloud data platforms. Each introduces its own agents, often optimized for specific tasks but lacking interoperability.

“Imagine you have a thousand agents running around the hospital. If they do not communicate, that’s missing the point of implementing these agents in the first place.”

The solution lies in orchestration, a layer that enables agents to coordinate, share context, and align with the broader patient journey. This includes introducing supervisory logic and defining when humans must intervene.

“The difference between an agent and a chatbot is that the agent has some level of autonomy and can make decisions. So what happens if there is a tie? Where do we hand off to the humans in the loop?”

For Eftekhari, solving agent interoperability is not optional. It is foundational to scaling AI safely and effectively.

The reality of data: Why the “clean slate” still doesn’t exist

Despite advances in AI, the underlying data challenge remains persistent and, in many cases, unresolved.

“For any AI or machine learning scientist, getting access to data all in one place is key. And it’s not there,” Eftekhari states.

Even in leading institutions, patient data often exists across fragmented systems, making it difficult to construct longitudinal views that AI models can effectively use.

“Do we have all patient data — images, genomics, everything — in one place, easily accessible and analyzable in a longitudinal manner? That’s something we are still working on,” she elaborates.

This challenge is particularly critical for generative AI, which performs best when it can process sequences of events over time.

“Having the sequence of all events for one patient in one place is something that would unlock a lot of different AI applications.”

Beyond data integration, Eftekhari highlights the importance of operational infrastructure, including MLOps and LLMOps capabilities, continuous monitoring, and automated workflows.

“These have already been solved in a lot of industries. We have to take the best practices and apply them to healthcare in a HIPAA-compliant manner.”

A consistent theme in the discussion is the importance of cross-industry learning. Healthcare, despite its complexity, does not need to build every capability from scratch.

“A lot of the stuff we are trying to do has been applied for decades in other industries,” says Eftekhari.

At City of Hope, this approach is reflected in team composition and strategy, blending expertise from outside healthcare with deep domain knowledge in oncology, genomics, and medical imaging.

“We try to merge the best-in-class technology and leading-edge AI, not just in healthcare but anywhere else in the world, with the best skillset and expertise in science.”

This convergence allows the organization to move faster while maintaining clinical rigor.

The strategic data assets that will define the future

Looking ahead, Eftekhari identifies several data assets that will shape the next phase of oncology AI. “For City of Hope specifically, the biggest data asset is the data on rare diseases.”

Rare disease data presents both a challenge and an opportunity. While limited in volume, it holds significant value for developing specialized AI models that can address underserved patient populations.

In addition, the organization’s investment in genomics is a major differentiator: “The genomics data we have through our precision medicine program is a differentiator for us. We can combine it with clinical and imaging data, going back to multimodal being important.”

Equally important is the diversity of the patient population, which supports the development of more representative and generalizable AI models: “We have data from very diverse patient populations that can enable developing AI tools that represent the whole population.”

Many of City of Hope’s AI models are already deployed in production and delivering measurable clinical outcomes. The impact is tangible, from reducing ICU admissions to lowering surgical complications.

“We have measured the impact they’re making in keeping patients out of the ICU or reducing surgery complications.”

This real-world success is driving external interest, with other institutions exploring how these tools can be adapted to their own environments.

“There is a lot of interest in Hope LLM and the other tools we have developed; they might be helpful for other cancer centers as well,” Eftekhari concludes.

CDO Magazine appreciates Nasim Eftekhari for sharing her insights with our global community.