AI in e-prescriptions should slow things down, not prescribe
At MedConnect, we tied electronic prescriptions to pharmacy systems, medication interaction checks, and an escalation protocol for chest pain.
AI in an e-prescription should say “stop” more often than “prescribe.”
That’s easy to argue with. If you’re going to add AI to a medical product, the instinct is to make it help the doctor make decisions faster: suggest a drug, a dosage, a substitute, the next step for the patient. But at MedConnect we ended up with a narrower and less exciting role: AI checks whether the prescribed drugs are compatible through a drug interaction database and raises a flag if the combination looks dangerous.
It’s less flashy, but it fits the real workflow better. The prescription is still written by the doctor. The pharmacy needs to receive the e-prescription in a format its system can understand. The patient needs to get the medication, not an explanation of why “the model decided otherwise.”
A prescription is an integration problem, not a screen with a button
The main challenge here was not designing a prescription form. An e-prescription sits between several parties at once: the doctor, the patient, the pharmacy system, and the drug interaction database.
We connected pharmacy systems so that the prescription would not remain an internal MedConnect record. After the consultation, the prescription has to be usable at the pharmacy. That means medication data cannot be stored as arbitrary text like “the thing the patient took in the spring.” It needs to be structured in a way that can be matched against the pharmacy side and checked against the drug interaction database.
And here’s the unpleasant part: medical meaning and integration format rarely match one to one. The doctor thinks in terms of a prescription. The pharmacy system thinks in terms of an item, availability, and dispensing rules. The interaction database thinks in terms of substances and combinations. AI ends up not as a “smart doctor,” but as a bridging filter between different representations of the same prescription.
Compatibility checks need to be intrusive
The most controversial question was not “do we need AI,” but where it is allowed to intervene. If the medication interaction check is hidden somewhere after the prescription is saved, it is useless: the doctor has already finished the workflow, the patient has already heard the plan, and the pharmacy is already waiting for the data.
So the check has to be placed where the prescription can still be changed. Yes, that adds friction. Yes, the doctor may get a warning in the middle of an already busy consultation. But the alternative is worse: miss a drug conflict and hope someone catches it later in the chain.
A concrete detail from the project: we argued not about the “quality of the AI,” but about the tone of the warning. If it is too soft, it becomes background noise. If it is too harsh, the doctor sees the system as blocking bureaucracy. In the end, what mattered more than the wording was when it appeared and how clear the reason was: which combination triggered it and where the check came from.
At the same time, we do not think AI should replace a clinical decision. The drug interaction database is a source of checks, not a license for autopilot. The model can help spot a risk, but responsibility for the prescription stays with the doctor.
Emergency scenarios should not be attached to the same UX
There is a separate case for consultations where a symptom like chest pain comes up. At that point, this is no longer about a careful hint regarding drug compatibility. In MedConnect, that case triggers an escalation protocol: calling an ambulance and sending the patient’s geolocation.
This is a different class of logic. A prescription has time for clarification, correction, and confirmation. In an emergency scenario, time works against the patient. That is why it should not be designed like a normal warning in the prescription UI. It should move the system into a separate branch: record the situation, trigger the ambulance call, send the coordinates.
This is the one place where we may be wrong: sometimes it is tempting to bundle everything into a single “medical AI assistant” that checks the prescription, recognizes a red-flag symptom, and tells the doctor the next step. On paper, that’s convenient. In practice, these scenarios are better separated by risk and by the consequences of a mistake.
A medication interaction check can stop a prescription and ask the doctor to reconsider it. An emergency protocol for chest pain should switch the system to escalation. These are two different brakes, and each has its own reason to trigger.