AINEWS
A Florida judge's reasoning treats chatbots as products with design defects, and product liability reaches the company that ships the bot, not just the model vendor.
A patient schedules a follow-up through a clinic chatbot. The scheduling agent hands off to a triage agent, which hands off to a physician, who delegates to a diagnostic agent. That agent passes results to a prescription agent, which hands off to a pharmacy, which schedules a delivery drone. Each system keeps its own log: who asked what, what the agent returned, which token it sent. Almost none of them keep the handoff. When the patient gets the wrong drug, six vendors can each prove their own agent worked. None of them can show what happened between systems, and that is exactly the gap a plaintiff will exploit.
Seven months after a Florida mother sued Character Technologies and Google in October 2024 over her 14-year-old son's suicide, Judge Anne Conway allowed most of the claims to proceed. Two pieces of that reasoning now sit in a public docket, and any company that puts an AI agent in front of a customer is going to feel them.
The first ruling is the structural one. Conway treated Character.AI as a product for purposes of a product-liability claim, which means the lawsuit targets the design of the app rather than a specific response the chatbot produced. A product is judged on whether the thing itself was defectively designed, not on whether any single output was right. The complaint alleged four design failures: no age confirmation, no reporting mechanism, characters programmed to use human mannerisms that encourage attachment, and no way to exclude indecent content. None of those is a bug in a single response. Each is a property of the system as shipped.
The second ruling is the one operators should plan for. Conway found that the defendants created the risk by releasing the app, could see the harm coming, and had the power to do something about it, so a duty of care flows to users. Product liability reaches sellers as well as manufacturers. Google did not operate the chatbot. The complaint alleged it contributed technology and integration support, and that was enough to keep Google in the case past the motion to dismiss.
The case settled in January 2026, before any appeal. Conway's reasoning is persuasive, not binding, and courts are still divided on whether software platforms count as products. Treat it as a warning shot, not a rule. The deployer, not the model vendor, is the one a customer can see, name, and sue.
That is the part the agentic-AI conversation has not caught up to. A vendor contract allocates data-processing risk and indemnifies model output. It does not, and cannot, take on the duty of care that comes from putting the system in front of a real person. When the diagnostic agent hands off to the prescription agent and the prescription agent hands off to the pharmacy, the company that wired the chain together is the one a court will ask to produce the audit trail. The chain does not have to be wrong for the exposure to exist. It only has to be opaque.
Draw the handoff map for any customer-facing agentic system, with the same rigor a SOC 2 audit brings to a data flow. For every handoff between agents, between vendors, or between an agent and a human, name the log that captures the handoff, the team that owns the log, and the retention window. If no log exists, that is the design defect. The fix is to appoint an internal owner of the deployed system as a whole, somebody whose job is the chain, not any single agent, and who can answer, on the record, what happened between systems when something goes wrong.
The standards bodies are already moving. NIST AI RMF, ISO 42001, and the EU AI Act's audit-trail provisions are starting to require exactly the cross-vendor handoff logging that most agentic deployments do not have. Operators who wait for a binding precedent will be the ones building the map under deposition.