AI

Cityblock’s test for healthcare AI trust starts with a problem the patient needs solved

Toyin Ajayi described what patients would use. Novartis and GE HealthCare executives explained how institutions could build evidence for tools that guide care.

Author
HDM Staff
Save
Share
Katie Mazuk, Toyin Ajayi, Jan Beger and moderator Roy Perlis seated on the Main Stage during the panel on AI and trust
Katie Mazuk, Toyin Ajayi, Jan Beger and moderator Roy Perlis discuss AI, trust and misinformation in Las Vegas. Photo credit: HLTH USA 2026.

    For a Cityblock Health member with a broken wheelchair, a useful AI tool would help get the repair paperwork done during the call. A less useful one would take a message and add another callback to the queue. Cityblock co-founder and CEO Dr. Toyin Ajayi used that example to explain why a patient might reject the idea of AI in healthcare and still welcome a service that gets a specific problem resolved. What the clinic does after the call will shape whether that patient wants to use the service again.

    Ajayi made the case during a Nov. 17 Main Stage discussion in Las Vegas with Katie Mazuk, who leads strategy, platforms and transformation at Novartis, and Jan Beger, GE HealthCare’s global head of AI advocacy. Dr. Roy Perlis, editor in chief of JAMA+ AI, moderated. He opened with a finding from a 50-state survey he led. The share of respondents reporting a lot of trust in physicians and hospitals fell from 71.5% in April 2020 to 40.1% in January 2024. The survey covered the pandemic years. The panel’s question was how healthcare organizations would earn trust as more people brought information from chatbots into appointments.

    Ajayi works with people on Medicaid and Medicare, including those covered by both programs. Cityblock sees patients whose care depends on coordination across physical health, mental health and social services. In that setting, she said, “that trust gap existed” before many patients used generative AI. People form an impression of a clinician from whether they feel heard and taken seriously. They form an impression of the organization when they try to fill a prescription, receive an unexpected lab result or learn whether one specialist has spoken with another.

    A shorter wait has to lead somewhere

    Mazuk said a common question for Novartis is whether a patient can find out sooner if a newly prescribed medicine is covered and where to get it. The company is testing whether AI can give patients that information in minutes in situations where they can now wait days or weeks. Mazuk described the work as an attempt to “prove value for patients” over time. The information takes care of one obstacle. The patient still needs a pharmacy that can fill the prescription and a price they can afford.

    She offered a separate example involving a hard-to-diagnose blood disorder. Mazuk said Novartis worked with about 50 hospitals to test an algorithm that flagged roughly 121 patients who might need additional testing. She said 55% of those flagged were confirmed to have the disorder. She did not name the condition or describe how the hospitals counted patients the tool missed. Her account illustrated how a flag could prompt a clinician to pursue a diagnosis that might otherwise take years. The effectiveness of the screening effort and what happened to patients after diagnosis would require further evidence.

    Perlis asked how an institution should respond when a patient, a physician and a decision-support tool reach different conclusions. Mazuk called that an “accountability and a design challenge.” She put responsibility with the organization that adopts the tool and sets the workflow around it. A hospital deciding when to page a clinician, a pharmacy deciding how to route a question and a drugmaker deciding how to present evidence are all making choices about what happens after the software produces an answer.

    Models have to work with a new population

    Perlis raised a familiar objection from physicians considering a model trained at another institution. A tool that performed well in one setting may behave differently with the patients they see. Ajayi said Cityblock has built and validated many of its models with data from the communities it serves. Its records can include questions about housing, food, safety at home and loneliness. Such information can help a care team understand whether a proposed next step is realistic for a particular member. Ajayi gave no performance comparisons between Cityblock’s models and tools developed elsewhere.

    Beger said developers of large models should start with diverse, high-quality training data. He also envisioned engineers working directly with hospitals to adapt models to local populations. The point emerged after a discussion of who should teach clinicians to use the tools and how frequently that training would need to change. Beger said “education should be evidence based” and drew a distinction between documenting what a tool can do and teaching people how to recognize when its answer may be wrong.

    Beger also argued that AI could help restore the “human touch” that he believes suffered when a screen came between patient and clinician. He envisioned tools that sort through different kinds of medical information before the visit so the clinician can arrive better prepared. Ajayi brought the question back to the daily experience of using the software. A tool that sends physicians to more screens, asks them to repeat work or adds steps for which they are given no extra time has to compete with everything else they are expected to finish.

    What happens when patients bring a chatbot’s answer?

    Ajayi said patients who count on a care team will return with what they read online. In her words, “they’ll bring their AI.” A physician can ask where a patient found an answer, what about it seemed plausible and what worries them. The resulting conversation can establish whether the information is relevant to that patient and give the clinician a chance to explain a different reading of the evidence. Ajayi said patients would be more willing to have that conversation if the practice had taken their questions seriously before a chatbot entered the picture.

    She also warned against treating misinformation as a problem that better messaging can solve on its own. A patient who receives a $5,000 MRI bill without understanding it, or who knows a medicine exists but cannot afford the copay, has a concrete reason to question a healthcare company’s motives. Beger agreed that cost mattered and argued that access and quality presented opportunities for technology to help. Mazuk said companies also have to follow where patients now look for answers and make reliable evidence available in those places.

    The test suggested by the panel starts with a single promise. If a chatbot gives a patient information about a treatment, the patient needs a way to ask a clinician whether it applies. If a practice promises to call with a result, someone has to make the call. If an agent helps a member file wheelchair repair paperwork, the member needs to know what happens to the request. A practice can check whether each of those things happened. The patient will have a clearer reason to decide whether to call it again.