Permitted actions
Each agent receives only the capabilities required for its approved use case.
Safety architecture
Safety begins with what an agent is allowed to do.
Healthcare agents do more than generate text. They speak to patients, access information and take actions. We design safety around that operational reality.
Our principle
Control capability—not only conversation.
A careful prompt is useful, but it is not a sufficient safety boundary. We also govern access, available actions, confirmation requirements and the point at which human judgement takes over.
An agent should be able to explain its limits because the system itself enforces them.
Escalation rules
Conditions for clarification, refusal and human handover are defined before deployment.
Data boundaries
Access, retention and disclosure follow provider policy and the needs of the task.
Human authority
Clinical judgement and higher-risk actions remain with appropriately authorised people.
Research signal
1 in 10Patients experience harm in healthcare.
WHO estimates that more than half of this harm is preventable. The figure concerns healthcare overall—not AI—but it sets the standard any agent must respect.
WHO · Patient safetyIn practice
When the system should stop
If identity, consent, scope or confidence is insufficient, the agent does not improvise. It pauses or hands the interaction to an authorised person with context intact.
Safety through the lifecycle
Designed in. Observed in production.
Safety is a continuous operating discipline—from defining scope through monitoring real interactions and approving each expansion of capability.
01
Before deployment
Map protocols, define prohibited actions, agree escalation thresholds and test representative failure cases.
02
During every interaction
Apply identity, consent, data and action controls in the context of the current patient and task.
03
After every interaction
Retain the evidence needed to review what the agent understood, decided, did and handed over.
04
Across every release
Re-evaluate behaviour before expanding scope or changing models, tools, policies or integrations.
Evidence, not assurance
Every important decision should be reviewable.
Teams need to understand not only the final response, but the context, policy and action path that produced it.
Interaction evidence
Review the conversation, relevant inputs, decisions, tool use and escalation path.
Policy adherence
Evaluate whether the agent remained within approved clinical and operational boundaries.
Outcome monitoring
Connect agent behaviour to patient access, care-team capacity and continuity outcomes.
Make safety part of the system—not a claim around it.
We can map your protocols and show how they translate into agent behaviour.