Responsible AI.

The Platform operates a two-tier oversight model.

• Tier 1 — Transactional oversight. Every AI-generated suggestion is reviewed, edited, or approved by a named user before it has operational effect. This is enforced at the data layer, not by policy alone.
• Tier 2 — Systemic oversight. A Compliance Office reviews model performance, override rates, and flag distribution quarterly. Material anomalies are escalated before customer rollout of updated models.

Bias mitigation is treated as a measurable engineering requirement, not a principle statement.

• Classification models are evaluated against stratified test sets across matter type, claimant demographic, and fee-earner grade.
• Confidence calibration is monitored per stratum. Divergence above the configured tolerance triggers re-training and a model hold.
• Override rate by user role is tracked. Systematic over-rides of AI outputs by senior users signal model miscalibration — not user error — and drive corrective action.
• No protected characteristic is used as a model feature. Proxy features are reviewed for indirect discrimination during each model release.

Every AI output, every user override, and every state change is written to a tamper-evident audit log with actor type, source records, model version, confidence score, timestamp, and before/after values.

• Audit logs are read-only after write and retained for the full tenancy plus seven years.
• Audit logs support regulator queries, client instruction, and panel insurer review without requiring vendor involvement.
• Customers can export their full audit log at any time in a structured, signed format.

Every AI-generated suggestion surfaces:

• A confidence score between 0 and 1.
• The top three contributing factors (in plain language).
• The underlying source documents or events consulted.
• The model version and the date of that version’s last evaluation.

No suggestion is surfaced without these four components present.

Continuous monitoring covers:

• Performance drift. Rolling accuracy against held-out labels and user-approved outcomes.
• Calibration drift. Whether a 0.85 confidence suggestion is correct 85% of the time in production.
• Distribution drift. Whether incoming data has shifted from the training distribution.
• Override patterns. Systematic user rejection of suggestions, analysed by cohort.

New models are rolled out under a staged deployment protocol:

• Shadow mode first — new model runs alongside the production model; outputs are compared but not acted on.
• Pilot tenant approval before general availability.
• Customer-visible change notes with evaluation metrics.
• Rollback path available for at least one prior model version.

Users and customers can report an AI output they believe is incorrect, biased, or unsafe via a single in-product channel. Reports are triaged within 24 hours by the Compliance Office. Material reports pause the relevant model on the affected tenant pending review.