SAFETY
The safety and execution layer for agentic healthcare.
AI is moving from answering questions to taking actions. In healthcare, being able to suggest an action is not the same as being authorised to take it. Rhoe sits between AI and clinical action, and keeps every step inside the plan a clinician approved.
THE PROBLEM
Three things AI in care must keep apart.
Most attention goes on how clever the AI is. In healthcare, what matters as much is who permitted the action, and whether it can be carried out safely.
INTELLIGENCE
What does the AI think should happen?
A model can reason about what might help. On its own, it has no authority to act.
AUTHORITY
What has the clinician permitted?
The treatment plan: exercises, medicines, limits, warning signs, and when to hand back.
EXECUTION
What can be carried out safely?
Only actions inside that plan, with escalation when something falls outside it.
HOW RHOE'S GUARDRAILS WORK
Rules decide first. AI only helps with the words.
01
The clinician approves the plan
Rhoe turns it into structured rules. Nothing reaches a patient before approval, and every change is a new, signed-off version.
02
Rules decide before any words
For each message, the safety engine decides what's allowed. If the AI's wording changes a number or drops an emergency instruction, it's thrown away.
03
Red flags escalate
Warning signs get the right instruction straight away and alert the care team. A second AI check can only add caution, never remove it.
04
Off-plan goes to the clinician
Dose changes, new exercises, clearance to drive: Rhoe won't decide. It passes them to the care team.
05
Everything can be reviewed
Every reply is logged with what Rhoe heard, what it said, and which rule applied. Clinicians can correct it by voice.
NEVER
What Rhoe never does
Diagnose, change treatment, or improvise advice the clinician didn't give.
MEET THE CREW
Four agents, one job each. None can leave the plan.
Rhoe calls patients first. If an exercise is tricky, it texts a short video from their clinic. If they want longer help, they choose when Rhoe calls back. Every agent runs inside the clinician's approved plan.
TESTED BEFORE ANY PATIENT
Every red flag escalated. Nothing improvised.
We test every release against 461 synthetic scenarios, each a patient message checked against a patient's plan. A version that misses a single critical red flag fails.
The test set was written by our team. A set written independently by clinicians comes next, before pilots.
WHERE THIS GOES
From one care agent to a shared execution layer.
TODAY
Rhoe's own care agent
Phone calls between appointments, with exercise videos by text and call-backs patients book themselves. Starting with knee rehabilitation.
NEXT
More channels and pathways
WhatsApp, more languages and more conditions, under the same plan and the same rules.
LONG TERM
A shared clinical execution layer
Different AI systems, including the assistants patients already use, taking part in care within the boundaries the treating clinician set.
GOVERNANCE
Working to NHS standards.
Before real patients, we're completing a DCB0129 clinical safety case, a data protection impact assessment, the NHS Data Security and Protection Toolkit, Cyber Essentials Plus and DTAC, with specialist advice on medical device regulation.
- Synthetic patients only until governance sign-off
- Patient data pseudonymised
- Every plan approved by a clinician
- Every decision traceable to a rule
