AI Governance

We are model-agnostic and eval-driven: for each use case we benchmark candidate models, including Arabic-dialect performance, on a golden set before selection, and we re-run evaluations before any model swap. Model choice, version, and change history are recorded in your runbook.

We use commercial API tiers of model providers under terms that exclude training on your inputs and outputs, enable zero-data-retention options where the provider offers them, and never use your data to train, fine-tune, or evaluate systems for any other client. Fine-tuning on your data happens only with a written agreement, and the resulting artifacts are yours.

Every deployment defines, in writing at the Pilot stage, which actions the AI may complete autonomously, which require human confirmation, and which are always human-only. Healthcare documentation is always reviewed and signed by the clinician. Escalation to a human is a designed path in every agent, never an afterthought.

Agents answer from grounded sources — your documents, catalog, and policies — using retrieval, and are configured to say "I don't know, let me connect you to the team" rather than guess. High-risk topics such as pricing commitments, medical, or legal questions are gated or escalated. Golden-set evaluations include adversarial and dialect cases, and production outputs are sampled and scored continuously on the monitoring dashboard.

A harmful-output incident — a wrong commitment made to a customer, a privacy leak in a response, or systematic bias — is treated as a P1 or P2 under our SLA: we contain it by disabling the affected path, notify the client per the DPA's timelines, investigate root cause via logs and eval replay, fix the issue, and add the failure case to the permanent evaluation set so it cannot silently recur.

Agents identify themselves as AI assistants; we do not build agents that pretend to be human.