AI Data Governance Control Plane

The problem it solves

Shadow AI discovery tells you which AI tools are in use. It does not tell you what went into them. The control plane closes that gap: it examines the content of AI interactions for sensitive data — personal information, health data, payment details, credentials, source code, and customer records — and applies your organization's rules before or after that content reaches the provider.

How content is classified

Classification runs in two layers. A fast pattern layer catches well-structured identifiers such as card numbers, national IDs, and API keys. On top of that, an optional Named Entity Recognition service reads unstructured prose and identifies people, organizations, locations, and other entities that no pattern can reliably match. Each finding carries an entity type and a confidence score.

What happens on a detection

The detector never decides the outcome. It reports what it found and how confident it is; your policy decides what to do. That separation matters, because the same detection can be acceptable in one workflow and unacceptable in another.

Allow

The flow proceeds unchanged and is recorded for reporting and trend analysis.

Mask

Detected values are replaced before the content reaches the AI provider. The user's request still works; the sensitive value never leaves.

Block

The request is refused, the user is told why, and the attempt is recorded.

Policy packs

Rules are grouped into policy packs so you can express intent once and apply it broadly: which entity types matter, what confidence is high enough to act on, which destinations are trusted, and what action each combination produces. Every organization starts with a sensible baseline pack and can add its own on top. Changing a pack takes effect on the next flow — no redeployment, no engineering ticket.

Connecting a system: two paths

There are two ways to bring an AI system under the control plane, and they are not mutually exclusive.

• Inline enforcement — route AI requests through the RAIC proxy endpoint. Classification and policy run before the content reaches the provider, so masking and blocking actually prevent exposure rather than just reporting it.
• Flow reporting — have an application report its AI interactions after the fact. You get classification, policy evaluation, and the full record, but enforcement is advisory because the content has already been sent.

Agentless collection from managed AI platforms

For organizations running Microsoft 365 Copilot, ChatGPT Enterprise, or Claude (Enterprise/API), RAIC connects read-only to the vendor tenant and sweeps administrative and usage telemetry on a schedule. Nothing is installed on a laptop, and no prompt or file content leaves the vendor — these connections are detect-only. Each sweep runs through the same policy and reporting pipeline as directly connected systems, so you see who is using which sanctioned tool, at what volume, and when behavior changes.

• Read-only credentials are stored encrypted and are never displayed again after they are saved.
• Each sweep resumes from where the previous one stopped and deliberately overlaps, so late-arriving vendor records are still picked up.
• If a vendor starts refusing requests, that connection pauses itself for a cooling-off period and resumes automatically instead of retrying in a loop.
• Claude is collected as two independent views — administrative activity and usage analytics — each with its own schedule and its own pause behavior, so a slowdown on one never blinds the other.
• Unusual jumps in usage volume are flagged for review against your own recent baseline rather than a fixed threshold.
• Every sweep writes a run record to your organization's activity log and to the platform log, whether it found anything or not.

The payload vault

When a policy needs the evidence — not just the verdict — a sample of the offending content can be retained in the encrypted payload vault so a reviewer can see what actually triggered the decision. Samples are encrypted at rest, visible only to authorized reviewers in your organization, and expire automatically within a retention window of at most 180 days. Expired samples are purged on a daily cycle; the decision record survives, the sensitive content does not.

What gets recorded

Every flow, classification result, policy decision, pack change, vault access, and purge is written to your organization's activity log and mirrored to the platform log. That record is what turns "we control what our staff send to AI tools" from an assertion into something you can demonstrate to an assessor.

Availability

The AI Data Governance control plane is available to Enterprise organizations and is enabled per organization by an administrator. Named Entity Recognition is an optional add-on within the module and can be toggled independently.