Skip to main content
Last reviewed: March 5, 2026 Owner: Security + Engineering Review cadence: Quarterly Status: Implemented This page explains how AI provider integrations are configured and what controls apply to AI-related data paths.

What this page answers

  • What data is sent to AI providers for classification and analysis workflows
  • Which provider and key-ownership modes are supported
  • How retention and training controls differ by provider and configuration

Current state (as of March 5, 2026)

Tero supports three AI deployment modes: hosted default provider path, bring-your-own provider credentials, and customer-controlled self-hosted provider routing.

AI workflow data scope

AI workflows use data needed for classification and analysis tasks. By default, this includes control-plane-relevant context and telemetry samples required for the task, not unrestricted ingestion of all customer telemetry.
TopicBaseline approach
Prompt scopeTask-scoped inputs for classification and analysis workflows
Data minimizationInputs are bounded to workflow requirements
Model output handlingOutputs are used for product classification and recommendation workflows
GovernanceProvider path and key ownership can be customer-controlled by deployment model

Provider and key ownership modes

ModeDescriptionTypical use case
Tero-managed provider pathTero-hosted default provider configurationFastest onboarding
Bring-your-own provider credentialsCustomer API credentials used for provider callsCustomer-controlled commercial relationship with OpenAI/Anthropic
Self-hosted provider pathRuntime and provider routing controlled in customer environmentMaximum infrastructure and network control

Provider data controls (external policy alignment)

Provider pathTraining usage baselineRetention baselineZero-retention option
OpenAI APIAPI data is not used for model training by default (unless customer opts in)Abuse-monitoring and application-state retention depends on endpoint/configuration (default policies apply)Available for approved organizations on eligible configurations/endpoints
Anthropic APICommercial/API data is not used for model training by defaultStandard API retention baseline applies (commercial policy default)Available by agreement for eligible enterprise API use cases
AWS BedrockPrompts/outputs are not used to train base models and are not shared with model providersBedrock service data-protection model appliesPrivate connectivity and customer-managed encryption controls available

Deployment boundary

Control areaTero-hostedSelf-hosted
Provider account boundaryTero-managed or customer-provided credentialsCustomer-controlled
Runtime/network boundaryTero-hosted boundaryCustomer infrastructure boundary
Provider policy configurationConfigured per provider pathCustomer-configured
Compliance postureTero controls + selected provider controlsCustomer environment + selected provider controls

Model and provider coverage

  • Recommended frontier providers for production quality are Anthropic and OpenAI.
  • Bedrock and additional provider paths are supported as deployment-dependent integration options.
  • Additional open-source/local model paths (for example, Ollama-compatible) are possible, with quality validated case-by-case.
ControlRecommended setting
Provider account boundaryUse customer-provided credentials or self-hosted provider routing for strict boundary control
Data retention modeUse provider zero-retention mode where eligible and contractually enabled
Training controlsKeep provider training opt-in disabled for API data
Prompt scope controlLimit prompts to minimum task-relevant context and avoid unrestricted raw payload submission
Secret managementStore provider credentials in managed secret stores with scoped runtime access
Key lifecycleRotate provider credentials on a defined cadence and on any incident trigger
Network controlsRestrict outbound destinations to approved AI provider endpoints
AuditabilityLog AI request metadata and outcome events without exposing sensitive prompt text in operational logs

External references

Evidence you can request

TopicPrimary evidence
High-level data scopeOverview, Data Handling
Ownership splitShared Responsibility
Identity and credential controlsIdentity and Access
Encryption and key controlsEncryption and Key Management

Exceptions and governance

Any AI data-path exception requires documented approval, compensating controls, and a time-bound remediation plan. Evidence requests: