Read-mostly
Default permission boundary
Roadmap architecture control
Roadmap — Stage 3
Not live. Cases the deterministic workflow cannot resolve may be routed to a planner only within explicit tool, permission, step and approval boundaries.
Roadmap — Stage 3
Agentic-Roadmap
Agentic Exception Resolution
A governed future stage for cases the deterministic workflow cannot resolve: a planner may prepare bounded, reversible next steps, while a person approves every irreversible action. This capability is not live.
GOVERNANCE
Confidence-gated intelligence
The planner must cite the evidence behind each proposed step. Missing evidence, policy conflicts, low confidence or a reached step limit stops the run and returns the case to a person.
The problem
Disconnected evidence creates avoidable review work.
Some exception queues may eventually contain enough high-variety cases that fixed playbooks cannot prepare every useful next step. Ralixar will bring this stage forward only when operating evidence and governance maturity justify it—not because agents are fashionable.
Hard limit
Maximum tool steps per plan
Roadmap architecture control
Human approval
Required before irreversible action
Non-negotiable governance rule
The AI under the hood
A named technique behind an explicit control boundary.
The model has a bounded job. It does not own the workflow or decide what a person must do.
Agentic-Roadmap
Technique in plain English
A planning model may decompose one unresolved exception into bounded steps and call only an allow-listed tool set under workflow supervision.
Inputs
- The exception record and linked shipment evidence
- Applicable rules, approvals and prior confirmed actions
- A tenant-specific allow-list of read-mostly tools
- Explicit step, time and permission limits
Confidence gate
The planner must cite the evidence behind each proposed step. Missing evidence, policy conflicts, low confidence or a reached step limit stops the run and returns the case to a person.
Rules fallback
The deterministic exception queue, named owner, due date, escalation path and manual playbooks continue unchanged. No case depends on an agent to remain operable.
What the human decides
A named, authorised person reviews the complete plan and approves every irreversible action, including filings, payments, bookings, partner instructions and external communications.
Bounded operating controls
- Bounded, allow-listed tool set
- Read-mostly permissions by default
- Hard step limit and timeout
- Evidence citations and complete action trace
- Mandatory human approval before irreversible action
Triggers that bring Stage 3 forward
- Sustained exception volume beyond deterministic playbook capacity
- Enough exception variety to justify bounded planning
- Confirmed deal losses attributable to unresolved exceptions
- Maturing internal and external governance standards
Evidence and expected outcomes
Industry evidence, not invented product proof.
Each item states its source class and boundary. Ralixar is pre-launch; these figures do not claim customer outcomes.
Govern
NIST's AI Risk Management Framework places governance across the AI lifecycle rather than treating it as a final check.
NIST AI Risk Management FrameworkHuman oversight
The EU AI Act identifies human oversight as a control for applicable high-risk AI systems; Ralixar uses approval boundaries as a broader product-design principle.
European Commission — AI ActNot live
This page documents a Stage 3 roadmap direction, not current product availability or measured performance.
Ralixar architecture roadmapModules involved
One workflow across a shared shipment record.
Each module works on the same evidence, event and decision history.
Related solutions
Continue through the same operating record.
Demurrage and Detention Prevention
See a deterministic exception workflow with embedded prediction.
ML Prediction
Learn moreConversational Analytics
Ask read-only questions through a governed semantic layer.
LLM Single-shot
Learn moreRegulatory Intelligence
Review grounded summaries before releasing rule changes.
LLM Single-shot
Learn more