Right answer, wrong messenger: trust in the source decides whether advice lands

A recommendation can be clinically correct and still fail, because the person receiving it does not find the source credible. I keep coming back to this, and it does not get enough attention in how teams plan patient-facing AI.

The Before You Build method sets out five uncertainties that decide whether a health product creates value in practice: need and adoption, workflow and accountability, action logic, commercial viability, and regulation and governance. Trust in the source does not sit inside any one of them. It sits underneath workflow and action logic both. You can specify the right action, put it at the right point in the pathway, give it to the right person to deliver, and still lose the patient at the moment of delivery because the message did not come from someone they believed.

There is good evidence for the general shape of this from outside AI. Saul Weiner and colleagues sent unannounced actors, trained to present as patients, into consultations with 111 internal medicine physicians. Each presented a clinical red flag and a contextual one, something about their circumstances that would make the standard plan fail. The physicians probed the contextual flags about half the time. In the encounters where context complicated the picture, only 22% produced an error-free plan of care.

That study is about clinicians rather than AI, and about attention to context rather than trust in the messenger. The transferable finding is the one that matters here. A plan that is correct in general and wrong for this person is not a correct plan, and a system optimised for the first kind of correctness will produce a great deal of the second.

The other half of this is delivery. How something is framed, the order it arrives in, and how much of it arrives at once all change whether a person moves towards engagement or towards avoidance. That is a different problem from credibility and it is just as determinative.

Take side effects in obesity care. Someone starting a GLP-1 medicine who is told, accurately and in full, about the range of gastrointestinal side effects, at the moment they are already anxious about starting, may decide not to start. The same information, ordered differently, with what to expect in the first fortnight separated from what is rare and serious, and with an obvious route to ask a question, produces a different decision. Nothing about the clinical content changed. The person's decision did.

The useful distinction is that credibility is largely inherited and delivery is designed. Whether someone trusts a brand, an app, or a clinician is mostly settled before your product arrives. Framing, sequencing, and how much information you give at once are choices your team makes, and they can be tested before launch rather than discovered after it.

Which makes them questions an Evidence Phase can answer, rather than opinions to argue about in a workshop. Does this population find this source credible for this kind of advice. Does this way of delivering it move people towards the action or away from it. Both can be tested with simulated interactions and small studies, before anything is built at scale.

Predictive accuracy sets the ceiling on what a system could achieve. Trust in the source and the design of delivery decide how much of that ceiling you ever reach. Most teams I meet are spending nearly all their effort on the ceiling.