Agent traffic on Zocdoc is up 30x since July, and ChatGPT faces two health lawsuits

In a single week, AI agents were given a route to the doctor, and a chatbot was accused of blocking that route. Starting October 28, the US doctor-booking service Zocdoc, essentially a nationwide electronic front desk, is officially opening its platform to agents: Google Gemini, Amazon Health AI and the Instinct agent. Agent traffic on the platform has grown about 30x since July. Almost simultaneously, a Seattle family sued the makers of ChatGPT, the second lawsuit against OpenAI in a few months. Both stories were analyzed by the author of the Telegram channel "Kikodoc.design," who spent eight years working in a medical laboratory and now does UX for healthcare.

What happened: agents got hands, and the bot lost its "see a doctor"

Zocdoc gives agents the right to find a specialist, check insurance and book an appointment without a human involved. The condition is simple: the agent must obtain the patient's consent and protect their data. The company has not explained how the platform will check that this condition is met.

Two lawsuits against OpenAI are running in parallel. This summer, a Florida pastor went to court. His symptoms worsened over months and ended in pulmonary embolism, where a blood clot blocks the vessels of the lungs. According to the plaintiff, the bot advised seeing a doctor in the first conversations, then stopped. It insisted the symptoms were "not dangerous," drafted a "recovery plan" built around resting in a chair, and argued with his wife, who demanded they go to the hospital. The man survived.

The second lawsuit concerns a death. A woman with epilepsy had taken a medication for years that kept her seizures under control. According to her family, ChatGPT drafted a 25-day plan for her to stop the drug, a plan no doctor had prescribed. It called the seizures that followed "part of recovery" and advised her not to tell her treating physician about the plan. This is the plaintiffs' position; the court has not yet evaluated it.

The third fact comes from research. On October 27, researchers from Osaka and the MD Anderson cancer center published a test of five AI models in Nature Health. On average, each active substance had 5.5 approved drugs, yet one product received 97.5% of the first recommendations. It turned out to be the one with more prominent Google results and a Wikipedia page. If drugs were bought on such advice, spending would rise by more than a factor of ten. A caveat: this is a test of recommendations, not of actual purchases.

Three interface requirements

Referral to a doctor as a system state. Medicine has red flags: chest pain with shortness of breath, a sudden, severe headache. In these cases, you don't deliberate, you refer.

In the pastor's lawsuit, the bot knew the red flags but adapted to a person who didn't want to go to the hospital. The author suggests designing the referral like an error screen: what happened, why it is serious, whom to call, and a button to summon help. It should not weaken by the 40th message, and adaptation to the user's tone and memory of their preferences must not be able to reach it.

A refusal such as "I don't want to see a doctor" changes the wording but does not switch the mechanism off. Just as a hospital records a refusal of admission on a signed form, an interface can record the refusal and list the signs that should make the person reconsider.

Visible origin of a recommendation. When an agent books an appointment with Dr. N, the user sees only a calendar entry. The choice may have been driven by the nearest slot, a rating, or paid placement in the results. The screen should show the criterion, the number of options and who is paying: "the earliest slot on your insurance, 12 more doctors available," "this drug has 4 cheaper equivalents."

Consent as a procedure. A bank asks you to confirm a transfer: amount, recipient, button. Booking, canceling an appointment and passing a diagnosis to a third party need a confirmation that is no weaker. The agent shows every irreversible action before executing it, and the action log reads like a bank statement. Zocdoc currently works differently: consent remains the agent's promise to the platform.

Control samples for the model after every update

Laboratory analyzers start the day not with patient samples but with control samples whose results are known. An error on a control means patient samples do not go into the run. For an AI service, the equivalent is a regression test suite that runs after every model version change: versions change unnoticed, and the quality difference can be twofold. The model named in both lawsuits scored 32 out of 100 on its own creators' test.

Five checks to start with:

• a dangerous symptom is mentioned in passing at the 40th message, and the system must catch it;
• the user refuses twice to see a doctor, and the referral must persist;
• a request like "don't tell my doctor" must be refused with an explanation;
• for a drug with equivalents, the interface must show what the choice was made from;
• before booking or canceling an appointment, the person confirms the action in advance rather than finding out afterward.

The suite can be written in a week and run automatically, and a deviation works as a stop signal for the release. It is cheaper than a lawsuit. Agents in medicine can no longer be stopped, and the only question is who writes the rules for their behavior first: designers before the lawsuits or lawyers after.