Multidisciplinary GU tumour boards in the age of AI: Obsolete ritual or more essential than ever?

In genitourinary (GU) oncology, that alignment often determines whether a patient receives curative therapy, intensified multimodal treatment, or a trial that changes the trajectory of disease. But as AI tools - general models like ChatGPT and clinician-facing platforms like OpenEvidence - become increasingly capable of summarising literature and producing guideline-shaped recommendations, it is fair to ask: do we still need tumour boards? Or are MTBs destined to shrink into a niche role, replaced by “instant consensus” generated at the point of care?

A 42-year-old man presents after painless gross haematuria. Imaging reveals a 14 cm left renal mass highly suspicious for renal cell carcinoma (RCC). He undergoes robotic radical nephrectomy; pathology shows chromophobe RCC, pT3aNxM0. Rather than presenting him at tumour board, the clinician asks an AI platform two questions: 

1. Is there evidence to support adjuvant immunotherapy for non-clear cell RCC, pT3a?
2. Are there open clinical trials for this histology? 

This is exactly the kind of query that makes MTBs feel redundant. AI can rapidly summarize landmark trials, interpret eligibility criteria, and enumerate ongoing studies. In minutes, the clinician receives a polished recommendation that resembles what a tumour board might say.

And yet, that resemblance is also the trap.

Even in 2026, the best argument for MTBs is that their value is not limited to picking a regimen. GU tumour boards provide 

• Diagnostic correction and staging accuracy: pathology rereview, radiology re-interpretation, discordance resolution, identification of variant histology, margin nuance, nodal interpretation, and imaging pitfalls.
• Sequencing and intent clarity: curative vs. disease control decisions that include neoadjuvant/adjuvant/early salvage decisions.
• Constraint-based realism: overall determinants of life expectancy and treatment tolerance/adherence, including comorbidities, organ function, social barriers, trial logistics, insurance hurdles, and institutional capabilities.
• Trial matching with feasibility: not just “a trial exists,” but “we can actually enrol this patient this week,” with a research team that can move.
• Shared accountability and communication: a unified plan across specialties, documented rationale, and reduced downstream fragmentation. 

Prospective GU data support that MTBs change management often enough that clinicians cannot reliably predict which cases will benefit, suggesting MTBs add information, not just ceremony (1).

AI is unquestionably improving. Systematic reviews show rapidly growing performance in oncology question-answering but persistent heterogeneity in evaluation and meaningful error rates (2). In urologic cancers specifically, ChatGPT-4 recommendations demonstrate improved concordance with guidelines compared with earlier models, yet remain imperfect and context dependent (3).

More importantly, AI outputs are usually decontextualised. They can be guideline-shaped but patient-misaligned. 

In the chromophobe RCC vignette, an AI tool might: 

• Mistakenly extrapolate adjuvant evidence from clear-cell–dominant trial populations to non–clear cell histologies
• Overstate certainty when data are sparse
• Omit local trial feasibility
• Miss that the most valuable MTB contribution may alter pathologic diagnoses through pathology clarification 

AI can suggest; MTBs can adjudicate.

Finally, retrieval-augmented generation (RAG) may outperform general-purpose models in producing evidence-based answers. However, RAG represents only one component of a broader architecture in which AI systems dynamically retrieve information, invoke specialized tools, and integrate multimodal data sources to support clinical reasoning (4).

Recent advances suggest that the relevant comparison is no longer between tumour boards and static large language models, but between tumour boards and agentic AI systems.

These systems combine large language models with retrieval-augmented generation, domain-specific tools, and multimodal inputs - including imaging, histopathology, and genomic data. Rather than simply answering questions, AI agents can plan multi-step workflows, select and execute appropriate tools, and iteratively refine their outputs (4,5). Early studies demonstrate that such systems substantially improve completeness and accuracy in experimental case-based evaluations compared with language models alone, particularly in complex oncology decision-making tasks (4,5).

In parallel, multi-agent AI systems are emerging in which different agents assume specialised roles - such as hypothesis generation, critique, planning, and synthesis—engaging in a form of structured computational dialogue (5,6). Indeed, emerging multi-agent systems are explicitly designed around planning, tool selection, self-reflection, and debate among specialised agents, creating a computational analogue of multidisciplinary discussion (6).

This paradigm begins to resemble the structure of a multidisciplinary tumour board itself. The implication is not that AI will replace tumour boards, but that it may increasingly reconstruct aspects of them in silico.

Yet the greatest challenge for multi-agent systems may no longer be their technical design, but their clinical integration. The engineering required to orchestrate specialised agents is increasingly tractable; the harder task is determining how to embed these systems safely and meaningfully into clinical workflows, with appropriate validation, governance, human oversight, and accountability.

• MTBs reduce “single-specialty drift”. Urologists, medical oncologists, and radiation oncologists naturally anchor on the tools they know best. MTBs counterbalance that anchoring.
• MTBs are a quality and safety intervention. Their greatest value may lie in preventing incorrect decisions.
• AI is best used as the MTB’s pre-work engine. AI can: auto-draft case summaries; extract staging-relevant details; generate structured evidence summaries; identify trial opportunities. In this model, AI becomes the pre-processing engine, while the MTB remains the decision authority.
• GU MTBs are training infrastructure. They teach reasoning, not just guidelines. 

There is a legitimate case that many MTBs have become inefficient. 

• Opportunity cost is real
• Variability and “presentation theatre” exist
• Many cases are straightforward and guideline-driven 

With mature pathways and AI-assisted decision support, routine cases may not require full MTB review.

Selective MTB review should focus on variant pathology, discordant imaging, multimodal treatment decisions, trial-eligible patients, recurrence and oligometastatic disease, and complex comorbidity scenarios. AI can standardise routine decisions while MTBs may integrate AI to focus on complexity.

In the age of AI, the question is not whether MTBs will disappear, but how they will adapt. Even as AI evolves from suggestion engines to agentic systems capable of multimodal reasoning and workflow execution, MTBs remain uniquely positioned to adjudicate uncertainty, contextualise recommendations, and assume responsibility for decisions.

More importantly, while AI may generate increasingly sophisticated, disease-specific recommendations, it does not know our patients. It cannot fully appreciate their individual personalities, values, cultural backgrounds, fears, or goals of care. Only human interaction can create the trust and therapeutic relationship necessary for patients to understand, contextualise, and ultimately accept complex treatment recommendations. Even in an AI-enhanced MTB, physicians remain essential - not only for interpreting evidence, but also for recognising hesitation, addressing uncertainty, and responding with empathy and compassion when patients struggle with difficult decisions.

_______

Disclosure of AI use: Generative artificial intelligence was used to assist with language editing, refinement, and revision of this editorial. The author reviewed and takes responsibility for all content.

1. Scarberry K, Ponsky L, Cherullo E, Larchian W, Bodner D, Cooney M, et al. Evaluating the impact of the genitourinary multidisciplinary tumour board: Should every cancer patient be discussed as standard of care? Can Urol Assoc J. 2018;12(9):E403–8.
2. Carl N, Schramm F, Haggenmüller S, Kather JN, Hetz MJ, Wies C, et al. Large language model use in clinical oncology. NPJ Precis Oncol. 2024;8(1):240.
3. Tsai CY, Cheng PY, Deng JH, Jaw FS, Yii SC. ChatGPT v4 outperforming v3.5 on cancer treatment recommendations in quality, clinical guideline, and expert opinion concordance. Digit Health. 2024;10:20552076241269538.
4. Low YS, Jackson ML, Hyde RJ, Brown RE, Sanghavi NM, Baldwin JD, et al. Answering real-world clinical questions using large language models, retrieval-augmented generation, and agentic systems. Digit Health. 2025;11:20552076251348850.
5. Ferber D, El Nahhas OSM, Wölflein G, Wiest IC, Clusmann J, Leßmann ME, et al. Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nat Cancer. 2025;6:1337–1349. doi:10.1038/s43018-025-00991-6.
6. Lee Y, Ferber D, Rood JE, Regev A, Kather JN. How AI agents will change cancer research and oncology. Nat Cancer. 2024;5:1765–1767. doi:10.1038/s43018-024-00861-7.