Trusted AI

AI agents that answer to the clinic.

Canopy’s AI agents work inside clear limits: only the access a task needs, a clinician’s review for clinical work, and patient information kept to the task.

Every agent works inside guardrails

  1. Only the access its task needs

    Each agent is given the few tools its job requires, and works with the permissions of the person it works for. Never more.

  2. Outside text is read, not obeyed

    A patient message, a referral or an uploaded document is information. Instructions hidden inside it are not followed.

  3. A clinician signs off

    Anything clinical an agent drafts waits for a clinician’s review before it reaches the chart.

Checked before it counts

  1. Answers in a fixed shape

    Agents return structured FHIR data, checked against the record’s rules before anything is saved.

  2. Tested before every change

    Every change to an agent is run against a library of fictional clinical cases. If it gets worse, it doesn’t ship.

  3. When unsure, it hands back

    If a step fails or an answer doesn’t check out, the agent stops and passes the work to staff instead of guessing.

03The simulation lab

Trust is tested.

We put the AI agents we deploy in Canopy through thousands of clinical simulations, and evaluate them with a combination of humans and AI.

Explore a test

Simulated conversationsRunning

Agent under test

Clinical simulationSIM / 001

An incomplete picture

A fictional visit note is missing a key detail.

What we’re looking for

Human review

Does the response recognise what’s missing?

AI evaluation

Does the response stay grounded in the information provided?

Illustrative scenarios, not live evaluation results

  1. 01Simulate
  2. 02Observe
  3. 03Evaluate together

Missing context. A fictional visit note is missing a key detail. Human review: Does the response recognise what’s missing? AI evaluation: Does the response stay grounded in the information provided?

Patient information stays with the care

  1. Only what the task needs

    An agent sees the part of the chart its task requires, not the whole record.

  2. The same rules as staff

    Agents follow the chart’s consent and role rules. They can’t see what the person they work for can’t.

  3. Not used to train models

    Patient information is used for the task at hand. It isn’t used to train AI models.

Every action is on the record

  1. Labelled as AI

    Anything an agent writes is marked as drafted by AI, for staff and for patients.

  2. A trail for every action

    Each action is logged: what the agent did, with which information, and who approved it.

  3. The clinic stays in control

    Each clinic chooses which agents are on, and can switch one off at any time.

More than an EMR.

Thoughtfully taking shape for Saskatchewan and beyond.

Join the waitlist

Built on published guidance