AI & Agentic Workflow

No. r4 doesn't train or fine-tune any language model, and no customer data is ever used to train a model — by r4 or by any AI provider. This is enforced both contractually and architecturally.

No. r4's AI layer is model-agnostic. The default setup uses Google Gemini or a locally-run open-source model (Llama), both inside r4's secure environment. You can also connect your own AI model via API key — in that setup, all AI processing happens inside your own environment and no data ever reaches r4's AI systems.

No. The AI only recommends — it never acts autonomously. Every workflow has human-owned approval points the AI cannot skip, reorder, or alter. Anything that would activate a campaign, export data, or modify records requires a person's sign-off first.

Because the AI reasons over real, retrieved data rather than its training memory, outputs are grounded in verified facts. Low-confidence results are automatically routed to a human for review instead of being delivered automatically, and all structured outputs are validated against expected formats before delivery.

That risk is designed out at multiple layers. The AI only pulls from r4's internally validated data stores — not the open web — which removes the most common attack path. All inputs are sanitized before reaching the AI, and any incoming data, including public reference sources, is validated and quarantined if it looks suspicious.

Yes. Every AI request is logged — what data was used, which model and version ran, and what the output was. Logs are tamper-resistant, retained for a minimum of 12 months, and available for audit on request; customers can be granted read-only access to their own AI activity logs.

Each customer's AI workloads run in an isolated environment with no shared infrastructure between customers. All AI processing happens inside r4's secure perimeter — data never leaves that boundary during inference.

r4 XEM is built to work well with anonymized data, and we encourage customers to send de-identified inputs — hashed or tokenized customer IDs, aggregated fields — rather than raw PII wherever possible. The platform's segmentation, scoring, and AI outputs don't require raw identifiers to function, so anonymized data flows through the same pipeline with no loss of analytical value. Where PII is present, configurable detection and masking is applied to AI outputs, and data access is scoped so the AI can only see the specific data partition it's authorized for. Together, that's three layers of protection — anonymization at the source, masking on output, and scoped access in between — so personal data exposure is minimized at every stage of processing, not just one.

Your data always wins. r4 maintains a strict priority order — customer data first, r4's curated reference data second, public sources (like Census or weather data) last — and every output is traceable back to which source was used.

Because r4 isn't locked to a single provider, there's a built-in fallback — Gemini, a locally hosted model, or your own model can all serve the same architecture. Any model update is tested and scored against expected quality before it's rolled into production.

No. Workflow sequencing is controlled by a separate orchestration layer, not the AI itself. The AI performs its analytical task within a bounded step; it never decides on its own to call another system or expand its own scope.