AI Is a Copilot, Not a Decision Maker, Oncology Panel Says

Artificial intelligence (AI) is already easing administrative and documentation burdens in community oncology, but panelists at the 2026 Patient-Centered Oncology Care^® conference agreed that it should assist clinicians, not replace them. Moderator Rhonda Henschel, MBA, senior vice president of payer and care transformation at The US Oncology Network, asked the group to separate real value from vendor hype.

Where AI Is Delivering Value

The clearest gains thus far are seen in AI’s ability to assemble information, said Puneeth Indurlal, MD, MS, senior vice president of strategic operations at American Oncology Network. AI can pull data from PDFs, images, and scattered sources into a single chart, work that once fell to abstractors. A clinician confronted with hundreds of pages of hospital notes can find the key detail faster, he said; at scale, this same approach can build databases that help match patients to clinical trials.

For Johnetta Blakely, MD, MS, MMHC, executive vice president of quality and clinical operations at Tennessee Oncology, the biggest change is the adoption of ambient AI documentation. Before ambient AI, she took charts home daily; now she does so about once a quarter. Her notes are more thorough as well, often capturing details she would have omitted. In addition, Blakely practices in a rural area where she more frequently sees cancers she once rarely treated, so OpenEvidence has also become an important decision support tool, which she uses to compare information received from colleagues.

Samyukta Mullangi, MD, MBA, vice president of clinical strategy at OpenEvidence, said the platform’s physician users now represent nearly 65% of US physicians and generate about 2 million queries per day. Because it uses retrieval-augmented generation, she said, the references are drawn from a content library rather than invented. The main failure mode is “wayfinding,” in which a poorly framed question retrieves the wrong sources. Incorporating National Comprehensive Cancer Network guidelines required building a new vision model and 6 to 8 months of quality assurance. A newer feature, Evidence Grade, assigns a letter grade reflecting the quality of the studies behind an answer, a step Mullangi said increases costs but addresses concerns about overreliance.

Who Bears the Risk?

Blakely said accountability is a question too few are asking. Her practice’s tools keep a human in the loop, and she would never let an AI agent place orders on her behalf. “If my name is on an order, in my mind, I should be responsible for that order,” she said. Whether liability would extend to software makers if something goes wrong is unclear. “It’s kind of the wild, wild west out here,” she said.

Indurlal pointed to an attorney sanctioned over hallucinated citations as a signal that clinicians’ licenses are at stake. Larry Bilbrey, senior director of digital innovation at OneOncology, added that vendors are moving to limit their own exposure. A recent contract his organization signed with Anthropic, he said, places no liability on the vendor, and as such terms spread, risk will shift to practices and providers.

Blakely also raised a cognitive concern: clinicians who lean entirely on AI may lose skills they would otherwise keep sharp, much as her children cannot navigate without GPS.

Clean Data Comes First

On investment, Bilbrey said administrative burden is the first target, followed by clinical decision support, which he expects to ramp up substantially in the next 2 to 3 years. He evaluates tools through 3 lenses as follows:

• Do they generate revenue?
• Do they create savings?
• Do they improve quality of life enough to justify the cost?

Scribes fall into the last category, he said, noting there is no real evidence that they increase relative value units, despite marketing claims.

Data quality is the foundation. Faxes remain a primary way hospitals and practices exchange information, so Tennessee Oncology is using AI to capture and digitize them before they reach any system, identify the document and patient, and route the data into the electronic medical record. The data are then exposed through an application programming interface end point so other vendors do not repeat the work and generate additional token costs. Unstructured feeds like these, he said, account for about 80% of the data quality problem.

Indurlal said AI agents with human oversight can convert unstructured chart data into structured fields and validate them. An agent that cannot find a biomarker should not treat that as a negative result. Instead, it can flag what it does not know and ask for the missing information. He said traceability of how a recommendation was reached is as important as the recommendation itself.

“If you cannot trace that back, then you don’t want to implement it,” he said.

How Much Autonomy?

The panel’s liveliest exchange concerned where autonomy might be appropriate. Mullangi proposed weighing value against risk: high-value, low-risk tasks, such as daily patient check-ins that clinics lack staff to perform, or protocol-driven medication titration, could justify some autonomy. She noted that existing rule-based alerts in electronic health records already involve subjective sensitivity settings, so AI is not the first source of variability.

But Indurlal warned that when it comes to risk, the practice is on the hook. If an agent triaging electronic patient-reported outcomes fails to escalate a serious symptom, liability still falls on the clinician. Blakely agreed that checks are needed to detect missed escalations. Bilbrey said health care, as the most risk-averse industry, is not ready for fully autonomous tools. Low-stakes operational tasks, such as creating a duplicate chart from a fax, may be acceptable, Indurlal said, because errors can be corrected.

Talking With Patients

Both clinicians said they ask permission before using ambient documentation. Blakely has had only 1 patient decline. Mullangi said patients increasingly arrive with AI-generated questions, and checking a source in front of them has built trust rather than eroded it. Blakely recalled a patient who refused an iron preparation after an AI chatbot warned of hypophosphatemia, a complication she said she had never seen in practice.

Looking 5 years ahead, Bilbrey predicted stronger decision support and gradual acceptance of autonomy, starting with nonclinical tasks such as rescheduling. He does not expect autonomous clinical decision-making.

“I think the physicians are safe,” he said.