02 / 04
An AI-native CRM that shows its uncertainty
Northstar is a working prototype of a goal-driven CRM for retention teams. I designed the journey from a stated goal to an audited action, then built it end to end as a React and TypeScript front end on synthetic data: 7 flows, 3 roles and 10 designed failure scenarios.

A CRM that stays honest about what its AI doesn't know
Retention teams reactivating dormant enterprise clients work in tools built around modules, not outcomes. The judgment that decides who gets contacted lives outside the system, in exports and spreadsheets, where nobody can audit it.
Northstar starts from a business outcome. The AI investigates and drafts, people keep authority over consequential decisions, and the handoffs between them are designed rather than left to chance.
the implementation
The safety rules are in the system, not only in the UI
A disabled button is not a control. The rules from the Think section (versioning, approval scope, role checks, idempotent retries) are enforced below the interface, and the interface reflects them.
architecture
The AI provider sits behind a boundary. Nothing it returns is trusted by default
Layer 1
Interface
Requests carry the user's role.
- Workspace
- Investigation
- Opportunities
- Plan refinement
- Campaigns
- Activity
Layer 2
Application
- Permission checks
- Validation
- State persistence
Layer 3
Domain services
- Investigation
- Planning and versioning
- Approval (exact version)
- Execution (safe retries)
- Audit log, written by every service
Boundary
AI provider
- Structured responses only
- Checked against business rules
- Unavailable: follow-ups disabled
Layer 4
Data and channels
- Synthetic CRM records
- Simulated execution channel
mechanisms
How the rules hold
Plan versioning
Plans are never edited in place. A refinement writes version n+1 that records which version it came from. History is append-only.
Exact-version approval
An approval stores the version it covers. Execution only runs the version that was approved, so a newer version needs its own approval.
Structured AI output
The model returns structured data, not prose. Each response is checked against business rules before the user sees it, and fixed or rejected if it fails.
Safe retries
Execution safeguards track each action, so a retry skips the ones that already completed. Unknown outcomes are reconciled before they can be retried.
Role permissions
Restricted roles can't review, approve or execute. The check guards the action itself, so hiding a button is never the only safeguard.
Audit log
Every service writes events with a plain-language summary and a technical record. People and tools read the same history.
validation
What the end-to-end prototype run verified
Fresh Start run
- First-time workspace and goal creation
- Investigation and supporting evidence review
- Uncertainty resolution
- Plan revision and impact preview
- Handoff to the compliance reviewer and approval of the exact version
- Controlled, simulated execution
- Recovery from failed operations
- Returning workspace with investigation history
Results
- 11 clients initially eligible
- 14 to 13 uncertain records after one resolution
- 10 clients in the revised target audience
- 8 completed, 1 failed, 1 unknown execution outcomes
- 0 duplicated actions after retry
Prototype verification on synthetic data. These numbers confirm the system behaved correctly. They are not customer results or business impact.
limits & next
What this is, and what it isn't yet
Still a prototype. The data is 36 invented clients built to exercise eligible, uncertain and excluded paths, and execution is simulated: no real email was sent and no live CRM or provider was connected. The design reasoning comes from the problem framing, not from studies with retention or compliance teams, and nothing here measures reactivation, conversion or time saved.
Next: test the journey with real retention managers and compliance reviewers and watch where they hesitate at each seam. Measure time to a trusted decision, not clicks saved. Learn which conflicts teams resolve fastest and which they escalate. Then connect a real data source behind the same boundary, keeping execution simulated until approvals hold under load.
The AI does the investigating. People keep the authority. The space between them is designed, not left to chance.
Informative Nudge
Next