Implementation Guide

    Inventory which departments use which LLMs without sending prompts to a vendor

    Build the inventory from metadata your organization already controls: identity logs, DNS and egress telemetry, API key and service-account records, procurement records, and enforcement-point events. Attribute usage to departments through directory attributes and ownership fields, store a minimal record with no prompt or response content, and keep it current with events from enforcement inside your own cloud.

    1. Record the policy decision on embedded AI features: inventoried or exempted.
    2. Create the approved-provider list and a maintained list of known AI destinations.
    3. Add owner and team fields to API key and service-account records.
    4. Add AI fields to your existing software-request and system-inventory processes, with owner attestation.
    5. Adopt the minimal record schema with validation that rejects free-text and content fields.
    6. Route high-risk traffic through an enforcement point in your own cloud and compare its events against egress telemetry.

    Treat no prompts to a vendor as a design constraint

    A department-level LLM inventory answers one question: which organizational units use which models, AI sites, coding assistants and API-based models. Many discovery approaches answer it by routing prompt content to a third party. For a regulated enterprise, that creates a new exposure and can conflict with data residency and confidentiality obligations. One agency compliance plan illustrates the concern: its policy prohibits exposing agency data to generative AI models that may train on user prompts containing sensitive data. That is an agency policy rather than a general legal requirement, but it shows why the discovery method itself must not export content.

    The practical consequence is that the inventory is built from metadata, meaning facts about who connected to what and when, not what was said. NIST AI 600-1 supports this kind of practice. GOVERN 1.6 expects mechanisms to inventory AI systems, resourced according to organizational risk priorities. GV-1.6-001 asks you to enumerate generative AI systems, and GV-6.1-007 asks you to inventory third-party entities with access to organizational content and to keep approved provider lists. NIST supports inventory practices but does not prescribe a department-to-LLM schema, so the record design below is proposed guidance. NIST also suggests considering provenance information in inventory entries. This guide does not cover model or dataset provenance, and your own scope decision should say so explicitly.

    Metadata sources compared by coverage and blind spots

    No single source produces a complete inventory. Each one sees part of the picture and misses another part, so the method is to correlate several and record which source supports each mapping. The assessments below are general technical guidance and will vary with your environment.

    • Identity provider logs: Coverage: users who sign in to AI services through single sign-on, and OAuth grants to AI applications. Blind spots: personal accounts and tools reached without corporate sign-in.
    • DNS and proxy or egress telemetry: Coverage: which users, hosts or networks contact known AI destinations, including unsanctioned sites. Blind spots: devices outside the egress path, and the destination is visible but not always the specific model. Needs a maintained destination list.
    • API key and service-account registries: Coverage: named owners and teams for keys that call model APIs. Blind spots: keys created outside the registry and keys shared across teams.
    • SaaS and procurement records: Coverage: approved contracts, purchases and software requests. Blind spots: free tiers, personal subscriptions and AI features embedded in already-approved applications.
    • Cloud billing and usage metadata: Coverage: model services consumed per cloud account or project. Blind spots: models hosted or called outside your cloud accounts.
    • Enforcement-point events: Coverage: every interaction that passes through the enforcement point, with the decision taken. Blind spots: traffic that bypasses the point, so pair it with network telemetry.

    Attribute usage to departments and handle the gaps

    Attribution joins an observed event to an owner. For interactive users, join the identity to directory attributes such as department, cost center and group membership. For service accounts and API keys, require an owner and owning team field at creation and treat a missing owner as a finding. For applications that call models, take the owning department from your application or system-ownership records.

    Three gaps recur. Shared credentials attribute usage to the credential's owner, not the actual users, so record the identity type as shared and move to per-team or per-service credentials over time. Personal accounts can often be seen as network destinations but not as corporate identities. Attribute them by device or network user where possible and mark the confidence as low. Embedded AI features inside other applications may not appear as distinct traffic at all. NIST GV-1.6-002 treats this as a policy decision: define whether such features are inventoried or exempted, and record the decision.

    Telemetry will never be complete, so add a human process. The FTC example builds its AI use case inventory on its existing annual system inventory and software-request process, requires system owners to attest to completeness, and assesses inventory fields in an existing audit cycle. Reusing established processes this way covers use that telemetry cannot see and makes the inventory auditable.

    Define a minimal inventory record that excludes prompt content

    The following fields are author-proposed design choices, not regulatory requirements. The principle is that every field is an enumeration, identifier, timestamp or count. Do not include free-text fields, prompt text, response text, file contents or message excerpts. Enforce this with schema validation that rejects unknown fields, so you can show an auditor that content cannot enter the record. Store data at department level where individual attribution is not needed, and set retention by policy.

    • Subject: Department, attribution method and attribution confidence.
    • Model or tool: Provider, model or product name, and category: chat site, coding assistant, API model or embedded feature.
    • Access path: Identity type (user, service account or shared), credential or application owner, and destination.
    • Time and volume: First seen, last seen and request count per period, with no content.
    • Governance status: Approved-provider list status, exemption decision and risk tier.
    • Evidence reference: Data source type, retention class and record owner.

    Keep the inventory current with enforcement in the execution path

    An inventory assembled from periodic telemetry is a point-in-time view. Models, sites and tools change faster than review cycles, and the record drifts. Enforcement in the execution path addresses this because the decision point sees each interaction as it happens. It can record which model, tool or site was requested, by which team, and what the policy decided, without that data leaving your environment. Its limit is coverage: it only records traffic routed through it, so continue to use egress telemetry to find paths that bypass it.

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with actions to allow, redact, block, rate limit or escalate for human approval. Enforcement is at runtime, not reviewed after the fact, and its scope is individual interactions and tool calls. It is deployed self-managed in the customer cloud, with configurable retention and data residency, so enforcement and evidence data stay in your environment and prompts are not sent to a vendor. The audit-ready evidence it generates is mapped to frameworks including NIST and the EU AI Act, and supports the inventory and approved-provider practices described above. Trussed does not provide model provenance, model lineage or end-to-end tracing of agentic workflows, and this inventory method does not depend on them.

    Metadata-first inventory flow

    Identity and directory

    Who signs in, and which department they belong to.

    Network and egress

    Which AI destinations are contacted, and from where.

    Registries and records

    API keys, service accounts, SaaS and procurement entries.

    Enforcement events

    Per-interaction decisions recorded inside your own cloud.

    See how enforcement events become audit-ready evidence

    Review how policy enforcement in the execution path, deployed in your own cloud, records which teams use which models and tools without sending prompts to a vendor.

    Request a Demo