Every request that flows through a governed AI platform passes through several distinct layers, each with a narrow responsibility. Understanding the boundaries between them makes it much easier to reason about where a given control — routing, redaction, approval — actually lives, and why.

LayerResponsibility
Identity & accessAuthenticates the requester via SSO, resolves their role and team via SCIM, and attaches that identity to every downstream decision.
Policy & guardrailsEvaluates the request against content, security, and budget policies before anything is sent to a model — this is where PII redaction and prompt-injection screening happen.
Routing & gatewayClassifies the task and selects the best available model for it, handling failover if the primary choice is unavailable.
Knowledge & connectorsRetrieves relevant content from connected sources, tags it with provenance and trust level, and attaches citations to the final response.
Approval workflowIntercepts any action classified as consequential and routes it to a named human approver before it executes.
Audit & loggingRecords every step above — request, retrieved content, model used, output, and approval decision — as an immutable record.

Why the order matters

Policy evaluation happens before routing, not after — a request that violates a content policy should never reach a model at all, regardless of which one would have been selected. Similarly, retrieval and provenance tagging happen before the model generates a response, so the model always knows which parts of its context are trusted instructions versus untrusted retrieved data.

A request, end to end

  1. An employee sends a request through the chat interface or an API call. Their identity, role, and team are resolved immediately.
  2. The policy layer checks the request against content and budget rules. If it's within a team's budget and doesn't violate a content policy, it proceeds; if not, it's blocked or flagged before any model sees it.
  3. If the request needs external knowledge, the connector layer retrieves relevant documents, tags each with its source and trust level, and passes them to the model as clearly-marked data, not instructions.
  4. The routing layer classifies the task and selects the appropriate model, considering cost, quality requirements, and current provider availability.
  5. The model generates a response. If the response includes a proposed write action — sending an email, updating a record, executing code — it's intercepted by the approval workflow instead of executing immediately.
  6. A named human approver reviews the proposed action in context and approves, edits, or rejects it.
  7. Every step — the request, retrieved sources, model used, generated output, and approval decision — is written to the audit log as an immutable record.

Why this is a platform, not a wrapper

Each of these layers could, in principle, be built and maintained separately — a gateway product for routing, a security product for redaction, a workflow tool for approvals. The reason to run them as one platform instead of four integrated point products is that the layers need to share context: the approval workflow needs to know which policy rule flagged an action, and the audit log needs a complete record across every layer, not four separate logs that have to be reconciled after the fact during an incident review.