AI can propose, only people decide.
Updated 2026-09-07
Bernard Deffarges
Day nine, two in the morning. A participant in a six-week surgical study wakes with a pain that is different from the pain she was told to expect. The study nurse is asleep. The next visit is in five days.
Her Patient Digital Twin is awake. It asks a few questions, in her language and in her words. It does not diagnose, and it does not prescribe. It recognises that this conversation has reached the edge of its role, and it wakes the on-call nurse with a short summary and its sources. By twenty past two, a human has made the call.
This is what we build. We have taken to calling it the patient guardian: someone watching, who knows when to fetch a person.
Three rules a guardian lives by
It proposes while people decide. The guardian can ask, explain, notice and escalate. It cannot act on a patient. Every path from a proposal to a patient runs through a decision by a named person, and this is built into the structure of the platform, not into a button.
Everything is on the record. What the guardian saw, what it asked, what it answered, whom it woke and what that person decided: all of it is written to a journal in which each entry is immutable and chained to the one before.
It only sees what it is cleared to see. Which model may see which data is a setting made when the software is installed, and that setting has its own record. The model that talks to her never learns her name; the link to who she is stays inside the hospital's walls, and every use of it is recorded.
Why a study team should care
A clinical study asks for evidence about everything, including the software that talks to participants. With a guardian on T2R2, every interaction is attributable, time-stamped and reproducible, the properties inspectors summarise as ALCOA. A monitor can replay any night of any participant's episode. When the guardian's model or wording changes, it changes the way a protocol changes, through a version with a record, never silently. And when a participant withdraws consent, her data becomes unreadable while the trail of what happened stays whole, which is what the regulation and the biostatistics both need.
We wrote earlier about what a companion twin changes for participants and about consent as a conversation. The guardian is what those two ideas become when they run on one engine.
What changed recently
T2R2 was born in pharmaceutical research in 2020, where a result you cannot reproduce is worth nothing. From 2024 we rebuilt it as a second generation, for the patient's side and for the demands of medical device software.
The easiest way to picture it is a flight recorder that also recorded every dice roll. The AI is not deterministic; the frame around it is. Every answer a model gave is stored with the question and the version that produced it, so the whole flight can be flown again, years later, exactly.
Most AI agent toolkits were built to demonstrate autonomy. Agents talk to one another, and the conversation decides what happens next. It does not work in an hospital. On T2R2, agents coordinate through the patient's record, the steps are designed in advance, and what the system can do can be written down and inspected. Engineers call this kind of structure a harness. Whatever the word, it is the part of the system a regulator will ask about first.
Care (still) means presence
A guardian does not replace a nurse. It makes sure the nurse is woken at two instead of finding out at eight. That is our idea of care: presence, with a person always at the end of the line.