A 34-year-old woman from Haiphong, Vietnam, presented with a 12-month history of intermittent rectal bleeding that had become more frequent over the prior 2 months and was accompanied by a 5-kg weight loss. Evaluation revealed a lower-third rectal adenocarcinoma with suspicious lymph nodes, and she was advised to transfer to a cancer center in Hanoi. However, because she wanted to remain close to her two young children, she insisted on receiving care locally.

During the 2026 ASCO Breakthrough meeting session, “What Does It Take to Make Clinical Artificial Intelligence a Reality at the Bedside? A Discussion,” experts used her case to explore how AI could support oncology decision-making across the care journey while highlighting the implementation challenges that remain.

AI-Assisted Clinical Decision-Making

David Goldstein, MD, MBBS, FRACP, a senior staff specialist in the Department of Medical Oncology at Prince of Wales Hospital, Randwick, Australia, continued the case presentation through a series of new clinical developments, including a revised pathology diagnosis from proficient mismatch repair (pMMR) to deficient mismatch repair (dMMR), identification of a solitary liver metastasis, and later development of a rectovaginal fistula. These evolving scenarios highlighted additional opportunities for AI-assisted clinical decision-making across her care pathway.

Focusing on diagnostic imaging, Daniel Truhn, MD, MSc, Director of the Lab for Artificial Intelligence in Medicine at University Hospital Aachen, Germany, examined AI technologies that could support patient management in this case. He cited a Scandinavian mammography study, published in The Lancet, in which replacing one of two radiologists with AI reduced the quantity of interval cancers, improved sensitivity, maintained specificity, and decreased screening workload. Relating the example back to the case, he suggested that similar AI approaches could help identify lymph node progression on MRI that might otherwise be overlooked, although most of these algorithms remain limited to research.

Dr. Truhn also reviewed evidence from a Nature Medicine study supporting AI-based prediction of microsatellite instability directly from hematoxylin and eosin (H&E)–stained pathology slides. “Nowadays, the field is just exploding,” he said. “You have many research groups working on developing models to extract different kinds of biomarkers…from the H&E images.”

To address the challenge of navigating multiple AI applications and complex clinical information, Dr. Truhn introduced the concept of tool orchestration, or agentic AI, in which multiple AI tools are coordinated rather than used independently.

Expanding on Dr. Truhn’s introduction to agentic AI, Arsela Prelaj, MD, PhD, Head of the AI-ON-Lab at the Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy, focused on how this emerging technology could inform decision-making in tumor boards. She described agentic AI as an approach in which a large language model–based tool acts as an orchestrator, coordinating multiple AI tools (including narrow, generalist [foundation], and generative AI) while drawing on multimodal patient data and medical knowledge, such as clinical guidelines and web sources, to support clinical decision-making.

Using the presented patient as an example, Dr. Prelaj referenced a Nature Cancer publication describing an autonomous AI agent to illustrate how an AI-assisted tumor board could use GPT-4 to coordinate web-based search tools, a Python interpreter, a vision model (GPT-4 Vision) for generating radiology reports, narrow AI for digital pathology capable of identifying microsatellite instability, and a foundation model for CT scan segmentation within a single workflow. She also highlighted MTBBench as an example of a multimodal agentic framework developed to support molecular tumor board–style decision-making.

The case raised a treatment decision prompted by the revised histopathology report identifying dMMR: Should the patient be transferred to a cancer hospital, undergo local surgical resection, begin chemotherapy locally, or receive immunotherapy? Dr. Prelaj described unpublished work from her laboratory involving a more specialized immunotherapy-focused AI agent built around data from 3,000 patients spanning eight clinical data modalities. Using these routinely collected patient data, the system generated a structured treatment plan that included key findings, guideline-based evidence, predictive tools, recommended regimens, and contraindications.

Beyond informing diagnosis and treatment selection, Huren Sivaraj, MBBS, MRCP, MMed, CEO and Chief Medical Officer of Oncoshot, a company developing a real-time health insights exchange program, discussed how production oncology AI could help bridge the global clinical trials gap. He described an AI data extraction architecture applied in Asia that uses unimodal clinical data, with deidentification and inference performed within the hospital environment, followed by human-in-the-loop quality assurance. Drawing on real-world experience from a hospital in Singapore, he reported improved clinical trial prescreening productivity and suggested that similar approaches could help identify appropriate trial opportunities for patients, such as the presented patient, throughout the course of their disease.

Bringing AI to the Bedside

Implementation and usability, not just model development, remain key challenges for bringing AI into routine clinical practice, Dr. Prelaj explained. Dr. Truhn, meanwhile, cautioned that “we need to be a bit careful on how [we] implement it” because large language models may generate unsupported or inaccurate information or falsely affirm incorrect conclusions.

At a broader systems level, Dr. Sivaraj argued that “the question is…about the access and how do we solve for that access.”

Addressing some of these challenges from a health system perspective, Naveen Nagar, MBBS, MHA, the Senior Vice President and Head of Clinical Strategy at HealthCare Global Enterprises (HCG), India’s largest network of cancer hospitals, argued that successful AI implementation in low- and middle-income countries depends less on sophisticated algorithms than on scalable health systems. Drawing on HCG’s experience transitioning from a fragmented legacy IT environment to a unified AI-powered ecosystem, he described how AI-enabled digital pathology and radiotherapy planning have helped extend specialist expertise through a hub-and-spoke model, while vendor consolidation has reduced the cost of ownership.

“To make this hub-and-spoke [model] a reality, we need standardized, scalable diagnostic imaging, and we need to synthesize multimodal data for our unified tumor boards using agentic AI [to] provide precision cancer care,” he remarked, echoing broader themes of the session. “Once we have this robust database, I think we will be able to democratize global trials.”

DISCLOSURES: Dr. Goldstein has received honoraria from AstraZeneca, Boehringer Ingelheim, Duo Oncology, Merck, and Sun Biopharma; has held a consulting or advisory role with Amplicare, AstraZeneca, Boehringer Ingelheim, Duo Oncology, hC Bioscience, Medison, MSD Oncology, Seagen, Sun Biopharma, and Takeda; and has received institutional research funding from Amgen, AstraZeneca, Bayer, Bristol Myers Squibb, Celgene, Pfizer, Redx, and Zucero Therapeutics. Dr. Truhn has stock and other ownership interests in StratifAI and Synagen AI; has received honoraria from AstraZeneca, Bayer, Gilead Sciences, MSD, Pfizer, Philips Healthcare, and Roche; and has received reimbursement for travel-related expenses from MSD, Bayer, Philips Healthcare, AstraZeneca, Gilead Sciences, and others. Dr. Prelaj has received honoraria from AstraZeneca, Bristol Myers Squibb, and Janssen; has received institutional research funding from AstraZeneca, Bayer, Bristol Myers Squibb, Lilly, and MSD; and has received reimbursement for travel-related expenses from Janssen. Dr. Sivaraj has reported employment with Oncoshot; has held a leadership position with Oncoshot; and has stock and other ownership interests in Oncoshot. Dr. Nagar has reported employment with Apollo Hospitals Enterprise and HealthCare Global Enterprises; and has stock and other ownership interests in Fortis Healthcare.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.