Prohibited actions
Quoting, rating, binding, issuing, endorsing, cancelling, non-renewing, reserving or deciding a claim, coverage advice and payment confirmation are never autonomous.
Governance
What i23 will never do on its own.
Insurance contains licensed work. An assistant that performs it quietly is a liability. i23 recognises that work, prepares everything around it, and routes the decision to a person.
01
Signal surfaced
The platform highlights relevant information.
02
Human review
Your team applies judgment and expertise.
03
Approved action
The record updates with a clear audit trail.
The problem
Autonomous AI and regulated work do not mix.
- An AI that quotes or binds is performing licensed activity
- A confident wrong answer about coverage is a claim waiting to happen
- Messages sent without review become commitments to a customer
- Nobody can reconstruct why an automated system did something
- Compliance cannot sign off on a black box
The approach
Propose, approve, act, record — in that order, always.
Every consequential action i23 can take is declared in advance, checked against policy, queued for a person, and only then executed. Approvals expire rather than sitting open forever, repeated attempts cannot double-execute, and the entire chain from trigger to outcome is written to an audit trail.
Approval inbox
Proposals arrive with their reasoning and grounds. Approve, edit or reject — and rejection is a signal the system learns from.
Autonomy levels
Set how much i23 may do without asking, from proposal-only upwards. Licensed work stays with a person at every level.
Word-based confidence
Confirmed, likely, uncertain or not found. No invented percentages implying precision that does not exist.
Grounded or silent
Without supporting evidence in your workspace, i23 declines to assert. Unknown is an acceptable answer.
Idempotent execution
An approved action executes once. Retries and duplicate triggers cannot repeat a send or a write.
Expiring approvals
Stale proposals lapse instead of being executed against conditions that have since changed.
Complete audit trail
Trigger, reasoning, evidence, approver, outcome and timestamp are recorded for every AI action.
Truthful outcomes
Delivery and integration states reflect what actually happened. Success is never assumed.
Workflow
How it moves through the platform
- Understand
- Check policy
- Propose
- Human approves
- Execute once
- Audit
- AI assistance without regulatory exposure
- A reviewable record for compliance
- Staff keep authority over customer-facing work
- No surprise messages sent in your name
- Confidence that the system will admit uncertainty