Principle — The 70/30 Rule
Scope of AI
AI inside the Platform is used for three bounded purposes:
- Classification. Categorising operational events (billing events, evidence items, medical findings) into structured types with confidence scores.
- Extraction. Pulling structured fields from unstructured source documents (medical reports, disclosure, correspondence).
- Suggestion. Proposing draft outputs (time entries, liability positions, settlement ranges) for human review and approval.
AI is not used for: autonomous decision-making, external communication, financial commitments, or any action that would bind the firm or insurer.
What AI cannot do
AI inside the Platform is architecturally incapable of the following actions:
- Send communications to clients, counterparties, experts, courts, or insurers.
- Finalise, sign, or dispatch invoices, settlement offers, or pleadings.
- Move matter status forward past an approval gate without a human approver.
- Change rates, reserves, authority levels, or billing terms.
- Access data outside the tenancy in which it is operating.
- Overwrite or delete audit logs, approvals, or signed-off outputs.
Notice — Architectural, not policy
actor_type — and actor_type = ‘ai’ cannot enter a finalising state.Human approval is required
Every AI-suggested action flows through a defined approval chain before it has effect:
Explainability
Every AI output is accompanied by: (a) a confidence score between 0 and 1; (b) the top contributing factors to the suggestion; (c) the source records consulted; and (d) the model version and timestamp. Suggestions below a configurable confidence threshold are automatically routed for expanded human review.
Training data
- Customer Matter Data is not used to train third-party foundation models.
- Where model fine-tuning is performed, it uses aggregated, anonymised signals only and remains within the UK / EU perimeter.
- Customer tenants are isolated at the model-inference level. No prompt or output from one tenant is visible to another.
User override
Users may override, reject, or edit any AI suggestion at any time. Every override is captured in the audit trail with a reason and timestamp. The Platform does not reward AI concurrence or penalise human disagreement — both are equally valid outcomes.
Monitoring
Model performance (accuracy, calibration, drift) is monitored continuously. Material changes in model behaviour trigger review by the Compliance Office before customer rollout. No silent model swap occurs; every model version change is logged and notifiable.
Principle — Summary
Questions or Requests
Contact the GAAS LAW Compliance Office at compliance@gaaslaw.com.
Data Subject Requests
Submit a GDPR request to privacy@gaaslaw.com. Responses within 30 days.