The Disappearing Specialty, Part III

No Signal for the Work

Color Health's Cancer Copilot, built with OpenAI and tested with UCSF, reportedly achieves 95% concordance on real-world patient cases. That's the headline metric and the marketing wedge. The harder question is which cases were in the test set, and which weren't. The answer to that question is where palliative care lives.

Part I and Part II of this series argued that virtual-first cancer programs are being marketed as equivalent to comprehensive cancer center care while routing around the accreditation regime, and that the language of palliative care has been absorbed by marketing copy that isn't doing the work. This is the structural argument underneath both: the AI infrastructure that will drive the next decade of cancer care doesn't have a signal for the work specialty palliative care does. As the AI hardens, the field gets quieter — not because anyone decided to silence us, but because the algorithm has no pattern for the patient who needs us.

What Cancer Copilot Optimizes For

Cancer Copilot is one of a small number of generative AI systems specifically trained for oncology decision support. It's built on OpenAI's foundation models, fine-tuned on clinical data, and validated against real-world cases at UCSF. Its job, as Color describes it, is to help clinicians deliver "guideline-based cancer care in minutes" — accelerated screening eligibility decisions, treatment plan generation, adherence checks, survivorship follow-up.

Each of those tasks is well-suited to the technology. They are pattern-matching problems with structured inputs and bounded outputs. Is the patient eligible for lung cancer screening based on age, smoking history, and risk factors? Match against USPSTF criteria. Is the proposed chemotherapy regimen NCCN-concordant for this disease and stage? Match against NCCN guidelines. Has the patient missed a recommended follow-up imaging study? Compare against survivorship care plans. The structured-data substrate of these decisions is mature. The training data is plentiful. The right answer, when there is one, is verifiable.

This is the work AI is good at. It is also the work that most directly contributes to the metrics employer benefits programs are paid for: earlier detection, treatment plan accuracy, surveillance adherence, return-to-work timing. Cancer Copilot and its successors will get better at these tasks, and the cancer benefits ecosystem will continue to be built around them. None of this is a problem.

The problem is what isn't in the picture.

What It Doesn't See

Specialty palliative care is the work that happens when the structured patterns break down. The patient with a six-month prognosis who insists on a sixth line of chemotherapy because the alternative feels like giving up on her grandchildren. The family meeting that has to navigate a son who is convinced his father would want everything tried, when the father said the opposite to the night nurse at three in the morning. The seventy-eight-year-old who has not yet realized that the symptom she is calling fatigue is the disease progressing, and the conversation that has to honor what she's prepared to know. The thirty-four-year-old who is dying and whose grief is competing with his desire to make his last weeks meaningful for his children.

These cases don't sit cleanly in structured data fields. They live in progress notes, in family meeting documentation that is often unstructured or formatted differently across institutions, in the inflection of "I think we are approaching a moment" inside a clinician's note. The work itself involves silences as much as words — the work of being present, the work of pacing disclosure, the work of holding two contradictory truths at once. None of those things have a code. None of them have a structured tool that captures them.

When AI is trained on cancer care data, it is trained on the structured parts. It doesn't see the chaplain visit that reframed how a family was thinking about hospice. It doesn't see the nurse-led family meeting that prevented an unwanted ICU admission. It doesn't see the symptom-management adjustment that came from a fellow's intuition during a phone call at eleven at night. The substrate is wrong for what we do.

This isn't a technical problem that better engineering will solve. It's a category problem. The data we generate as a specialty does not look like the data the rest of oncology generates. Our notes are longer. Our endpoints are nuanced. Our work happens in conversations that don't always change a number on a flowsheet. AI trained on cancer care data — the cancer care data that exists in EHRs, that is structured well enough to ingest — is going to be trained on the parts of cancer that aren't us.

The Temel Paradox

In 2010, Jennifer Temel and colleagues published the trial that changed the field. In 2024, Temel and colleagues published the STEP trial in JAMA — the first major trial to compare in-person versus telehealth-delivered specialty palliative care for advanced lung cancer. Telehealth was non-inferior. The implications for access, for rural and underserved communities, for the populations who currently can't get to an academic palliative care clinic, are substantial.

The Temel paradox is this. The most well-evidenced specialty in cancer care for delivery in a virtual format is specialty palliative care. The most well-capitalized virtual cancer care company in the country has built a model that doesn't include it. Either Color's leadership doesn't know about the STEP trial, which strains credulity, or they know and have decided that specialty palliative care isn't a clinical capability worth building. The second interpretation is more likely. It is also what makes the structural argument in this series urgent. The decision is being made by people without HPM expertise on their leadership team, in a market that doesn't reward including us, with AI tools that don't see us, and the result is being marketed as comprehensive cancer care.

Augmented vs. Automated

I have written before about the distinction between augmented intelligence and automated intelligence in clinical care. The distinction is not technical — it describes the relationship between the technology and the human judgment it operates alongside. Augmented intelligence is designed to make a human clinician's judgment better. Automated intelligence is designed to replace it.

Specialty palliative care is the test case for whether oncology AI augments or automates. Our work is irreducibly judgment-laden. The right pace of disclosure for a particular family. The right balance between symptom relief and alertness for a particular patient. The right moment to introduce hospice. These are not problems with verifiable right answers. They are problems where the right answer depends on the patient in front of you and what they are prepared to know, want, and hold.

If the AI in cancer care is built to augment specialty palliative care clinicians — flagging patients with worsening trajectories, surfacing relevant prior conversations from the chart, drafting summaries the clinician then reviews and revises — it can be a real force multiplier for a specialty that has more demand than supply. If it is built to automate the work of specialty palliative care — generating goals-of-care discussion summaries from structured data, pushing standardized advance care planning prompts at patients, replacing family meeting decision support with algorithmic recommendations — it will produce the appearance of palliative care without the substance, and the buyers will not know the difference.

The current trajectory is toward automation, not augmentation. The cancer benefits market wants scalable, replicable, lower-labor solutions. Specialty palliative care is high-labor by design and by necessity ("FRICTION IS GOOD!" I yell into the darkness). The AI tools being deployed are not being built to augment our specialty because our specialty isn't in the buyer's specification. They are being built to make the buyer's specification cheaper to deliver. The work we do isn't being made cheaper. It's being made invisible.

The 2030 Picture

Here is the future I think is most likely if nothing changes.

By 2030, employer cancer benefits programs will cover a larger share of the working-age insured population than they do today (unless we suddenly become more rational). Color, Iris, and one or two of their competitors will have expanded substantially. Their AI tools will have improved. The metrics they report — earlier detection, treatment plan adherence, ROI per member, return-to-work timing — will look better. Their ASCO Certified or successor designations will be widely recognized as proof of clinical legitimacy.

The patients with advanced disease who do not appear in those metrics will continue to need specialty palliative care, and they will continue to live disproportionately in safety-net systems, on Medicare, on Medicaid, in rural areas, and in marginalized communities. The specialty palliative care workforce will continue to grow more slowly than demand. AI tools to augment our work will exist but will be built by smaller, less-capitalized players, because the deep-pocketed market doesn't see us. Our average referrals will continue to come too late, our family meetings will continue to be too short, and our integration with oncology will continue to depend on the relationships of individual clinicians rather than on the structures of the system.

The dashboards will look great. The patients we serve will continue to be where the dashboards aren't.

Where I Might Be Wrong

This is the most predictive piece in the series, and the easiest to be wrong about.

The technical limitations of AI on unstructured data may close faster than I expect. Large language models are getting better at narrative inference. The data substrate problem may be solved by 2027 or 2028, and the tools may be more useful for our specialty than I am projecting. I do not think this changes the structural argument — the buyer still has to want the tools to serve our work — but it changes the timeline.

Our specialty may build the AI tools ourselves. The Vyncas and SanLucas of the world are working on this. Members of our field have been in the room on EHR vendor product roadmaps for years. There is a possibility that the augmented intelligence path I'm sketching here is closer than I realize, just being built somewhere I'm not paying enough attention.

The cancer benefits market may correct on its own. As benefit utilization data matures, employers may notice that the patients with advanced disease are missing from their cancer programs, and demand integration. I don't think this will happen quickly because the patients who would expose the gap have largely already left employment, but I could be wrong about the speed.

What Comes Next

This series has argued that specialty palliative care is being structurally disappeared from the next generation of cancer care across three layers: accreditation policy, field marketing, and AI infrastructure. The disappearance is not happening because anyone decided to do it. It is happening because the people building the new cancer care economy don't see us as a clinical capability worth integrating, and the systems they are building have no native way to detect the gap.

The fix is not a single intervention. It requires that we be in the room when accreditation standards are written for virtual care models. It requires that we be loud and specific about what specialty palliative care is, distinct from primary palliative care and supportive oncology, in language buyers can verify. It requires that we partner with — and where necessary, build — the AI tools that make our work scalable in a virtual format, before the buyers conclude they don't need us. None of these are quick. All of them require the field to understand that we are in a different fight than we were in twenty years ago, and that the moves that worked then will not work now.

If specialty palliative care wants to be in the cancer care economy of 2030, the next two years are when we have to start making the case loudly enough that it can be heard over the marketing.

The Disappearing Specialty has been a three-part R&R series on how specialty palliative care is being structurally disappeared from the next generation of cancer care.

Part I — When the Floor Disappears
Part II — The Vocabulary Heist
Part III — No Signal for the Work (you are here)

I am a palliative care physician, educator, and professional strategery expert. Known for turning rounds into rants and rants into teaching points. Rounds & Rants represents my views — not those of any institution or professional membership organization where I hold a role. I don't write on their behalf and they don't vet what I publish.