Praxi AI & Actions
Praxi can suggest and execute actions from the same registry used by the Command Palette and workflows. Every AI-initiated action requires your explicit approval before it takes effect -- no exceptions.
One Assistant Across Your Practice
You always work with one assistant called Praxi. Praxi can help across clinical documentation, scheduling, billing, research, reporting, team wellbeing, and workflow automation. It automatically chooses the right capability and model for each request, so you do not need to select a specialist.
Student and Essentials plans include Praxi's efficient general assistance. Advanced specialist capabilities require a Pro plan or higher.
Model Routing
Praxi routes approximately 70% of AI interactions to Haiku (fast, efficient), 25% to Sonnet (balanced), and 5% to Opus (advanced clinical reasoning). The system automatically escalates to a more capable model when it detects clinical urgency, sensitive content, or high complexity.
Your subscription tier sets the model ceiling:
| Plan | Maximum model class |
|---|---|
| Student | Efficient |
| Essentials | Efficient |
| Pro | Balanced |
| Enterprise | Most capable |
How Praxi Uses Actions
When you ask Praxi to do something, it translates your request into one or more actions from the Universal Action Registry. For example:
- You ask Praxi: "Reschedule Sarah's Thursday appointment to next Monday at 2pm"
- Praxi identifies the matching action — rescheduling the appointment
- Praxi fills in the parameters: client, original time, new time
- Praxi presents the suggestion for your approval
- You approve, and the action executes
All AI-suggested actions use the .suggest.* pattern, meaning they always produce a suggestion rather than executing directly.
Domain Routing
Praxi routes each request to a capability restricted to the appropriate domain. This prevents clinical assistance from modifying billing records or scheduling assistance from altering clinical notes.
| Capability | Can Access | Cannot Access |
|---|---|---|
| Clinical | Clinical notes, risk assessment, documents, care plans | Billing, scheduling, staff HR |
| Operations | Appointments, billing, Medicare, navigation, help | Clinical notes, risk assessment |
| Research | Research databases, MBS schedule, guidelines | Practice data modification |
| Analytics | Analytics, reports, revenue data (read-only) | Client records, clinical notes |
| Team wellbeing | Staff wellness indicators, HR data, leave | Client records, billing |
| Automation | Workflows, automation rules, integrations | Clinical notes, direct client data |
You can ask Praxi any supported question. It automatically routes the request to the appropriate capability.
PsyBA 2025 Compliance
The Psychology Board of Australia's Code of Conduct (effective 1 December 2025) requires that practitioners maintain direct responsibility for all clinical decisions. Praxi enforces this by requiring human approval for every AI-initiated action, regardless of the action's normal approval tier.
Even actions that normally execute immediately (approval tier: NONE) require your confirmation when initiated by Praxi. This applies to every AI capability, including clinical assistance.
Reviewing AI Suggestions
AI suggestions appear in three places:
- Inline -- When chatting with Praxi, suggestions appear as actionable cards in the conversation
- Approval inbox -- Pending suggestions appear in the Command Palette under Pending Approvals
- Notifications -- You receive a notification for each suggestion requiring your attention
What a Suggestion Looks Like
Each suggestion card shows:
- The action name and description
- The parameters the agent has filled in
- Praxi as the source of the suggestion
- A confidence indicator (how certain the agent is about the parameters)
- Approve and Reject buttons
Approving or Rejecting
- Approve: The action executes with the suggested parameters. You can edit parameters before approving.
- Reject: The suggestion is discarded. You can optionally provide a reason, which helps Praxi learn your preferences.
- Edit and Approve: Modify the parameters, then approve. Useful when Praxi got the action right but a detail wrong.
You can reject a suggestion with a note like "Use 30-minute sessions, not 50" and the agent will remember this preference for future suggestions in the same session.
AI Permission Ceiling
Praxi inherits your permissions and can never exceed them. If you do not have permission to delete a client record, Praxi cannot suggest that action to you.
| Your Role | Praxi Can Suggest |
|---|---|
| Practice Owner | Any action in the registry |
| Practitioner | Actions scoped to own clients and records |
| Intern | Read-only suggestions, note drafts (require supervisor co-sign) |
| Receptionist | Scheduling, client intake, communication actions |
Audit Trail
Every AI interaction is logged in the audit trail:
- The query you sent to Praxi
- The suggestion Praxi produced
- Whether you approved, rejected, or edited the suggestion
- The resulting action execution (if approved)
These records are retained for 7 years in compliance with the Privacy Act 1988 and AHPRA record-keeping requirements.
Configuring AI Access
Practice owners can configure Praxi at Settings > AI. Individual practitioners can adjust their preferences at Settings > AI Preferences, including:
- Enabling or disabling Praxi
- Setting suggestion verbosity (Minimal, Standard, Comprehensive)
- Configuring custom instructions for more relevant suggestions
Related
- Command Palette -- Execute actions directly without AI assistance
- Approval Workflows -- How approval tiers work
- Undo & Recovery -- Reversing actions after approval
- Workflow Builder -- Using Praxi nodes in workflows