In retail, an AI support agent can load a customer's whole history to be helpful. In healthcare, HIPAA's minimum-necessary rule forbids that. The agent may only access the smallest slice of patient information the task in front of it requires, and that single rule reshapes the whole design.
Healthcare organisations watch a retailer deploy an impressive AI support agent and ask for the same thing. The technology transfers. The design does not, and the reason is one specific rule that most teams discover halfway through the build.
This article explains the minimum-necessary rule in plain terms, what it stops an AI agent from doing, and how teams design around it.
How a retail AI agent works, and why healthcare cannot copy it
A strong retail AI agent works by having context. When you contact it, it loads your full profile: every order, every past ticket, your preferences, your loyalty status. The more it knows, the more helpful it is. There is no legal limit on how much of your own data the retailer's agent may see to help you.
Blake Morgan's interview with American Express describes that model working well: an AI that knows the customer deeply and personalises every interaction. Healthcare organisations see it and want to copy it. They cannot, and the blocker is not technical.
The minimum-necessary rule, in plain terms
HIPAA includes a rule called the minimum-necessary standard. In plain terms: when anyone accesses a patient's protected health information, they may only access the smallest amount needed for the specific task at hand. A billing clerk handling a billing question does not get to see the clinical notes. The rule applies to systems as much as to people.
For an AI agent, this rule lands directly on the part that makes a retail agent good: the context window, the bundle of information the AI loads to handle a contact. A retail agent loads everything. A healthcare agent handling a question about a co-pay may not load the patient's diagnoses, medications, or visit history, because that task does not need them. Loading them anyway is not just untidy. It is a compliance violation.
What this stops the AI from doing
The practical effect is that a healthcare AI agent cannot be designed as one all-knowing assistant. It has to be designed as a set of narrow roles, each allowed to see only what its task requires.
So the agent cannot freely "remember" a patient across topics the way a retail agent does. It cannot proactively mention something it noticed in the clinical record while handling a scheduling call. It cannot load the whole patient to be conversationally smooth. Each of those moves, natural for a retail agent, is a minimum-necessary breach for a healthcare one.
How teams design around it
The teams that get this right design the constraint in from the start rather than discovering it in a compliance review. Three patterns recur.
They scope the agent by task. Instead of one agent with full access, they build task-bounded roles: a billing agent that sees billing data only, a scheduling agent that sees the calendar and contact details only. They gate access at the point of retrieval, so the system enforces what each role may load rather than trusting a prompt to ask politely. And they log every access, because minimum-necessary compliance has to be auditable: you must be able to show, later, exactly what the AI saw and why the task justified it.
The model is the same, the design is not
Kevin Davis, who writes KD Be Schemin, has described AI doing serious clinical work, reading X-rays and catching findings humans miss. The capability in healthcare is real and large. The customer service layer can use it. It just cannot be wired the way retail wires it.
For a healthcare CX leader, the retail demo is the wrong target. The same helpfulness has to be delivered through task-scoped, access-gated, audited roles, and that design has to be in place from the first day rather than added at the compliance review.