WellBridge Companion
The desktop presence, with its Ready, Working, Needs approval, Blocked, Complete and Failed states.
AI agent
Systems that act, and stop where they should.
A governed desktop agent that accepts a bounded business outcome, plans and executes through approved tools, verifies its own result, and returns a deliverable with the evidence behind it.
Experience
Five surfaces, each with one job.
Task Console
Brief, plan, active tools, elapsed time, checkpoints, output preview and the stop control.
Approval Tray
The exact proposed action, the system and data it affects, its risk class, and whether it can be reversed.
Evidence Viewer
The deliverable, the tests, the source record, the decisions, the changes, the permissions used and the rollback instructions.
WB Visual Model / Preview
A disclosed fictional synthetic persona for demonstrations and guided workflows. It is not an employee, a clinician or a licensed adviser.
Approval policy
What runs alone, and what waits for a person.
| Risk class | What it covers | What happens |
|---|---|---|
| Low | Read-only retrieval and work inside the task sandbox | Runs automatically |
| Medium | Draft changes to internal systems | Approval required before any write |
| High | External communication, production deployment, credential changes and regulated outputs | Explicit approval and secondary verification |
| Prohibited | Irreversible financial, legal, clinical or identity actions | The authorised human executes the final step directly |
Execution
The task state machine.
- Queued
- Planning
- Awaiting approval
- Executing
- Verifying
- Packaging
- Completed
- Failed
- Cancelled
First jobs
Three jobs, chosen because they can be checked.
Website QA and release pack
A release pack assembled from an approved checklist, with the checks and their results attached.
Client portal permission audit
A report of who can reach what. No permission is changed.
Quantitative report assembly
A report built only from approved data sources, with the sources named.
What the first release will not do
Autonomous fund transfers, market orders, medical decisions, account deletion and external publication are out of scope for the first release. The final step of an action like that is left to an authorised person.
Security
The controls behind it.
- Least privilege, short-lived credentials and per-tool scopes.
- Sandboxed execution with controlled network egress.
- Secret redaction in prompts, logs and screenshots.
- Prompt-injection detection on retrieved content.
- File malware scanning and safe rendering.
- Tenant isolation and private object storage.
- A human-readable audit trail and a signed artifact manifest.
- Emergency stop, task cancellation and tested rollback.
Completion
How it finishes.
- The companion stays in a neutral Ready state while nothing is running.
- During execution it shows a subtle activity pulse and the task title, and nothing more.
- An approval request takes focus, and never simulates urgency for low-risk work.
- On success it moves to Complete and presents the deliverable, the evidence and the next decision.
- On failure it explains what failed, what changed, what did not change, and how to resume safely.
Next step
Start with a conversation.
A consultation is a scoped conversation about what you need, not a sales call. You will leave it knowing whether this is the right practice for the work.