Data Privacy, Isolation, and LLM Provider Controls

This article summarizes how Reindeer protects customer data when agents use large language models, enterprise tools, and customer-specific workflow context. It is written for security reviewers, architects, and implementation teams that need to understand where data lives, how it is isolated, and how LLM provider usage is controlled.

Privacy Commitments

• Customer data is not used to train third-party foundation models. Reindeer uses enterprise/API provider configurations designed to prevent customer prompts, completions, files, and workflow data from being used for provider model training.
• Zero data retention controls for LLM providers. Where third-party LLMs are used, Reindeer routes requests through provider configurations that support zero data retention or equivalent enterprise no-retention controls.
• Minimum necessary data is sent to models. Agents send only the workflow context needed for the current task, rather than broad workspace or system-level data.
• Secrets are not exposed to the model. Credentials are injected into approved tools at runtime and are not included in prompts or visible to the agent as plain text.

Deployment Isolation

Reindeer supports deployment models that isolate customer data, workflows, and runtime resources according to enterprise requirements. For the builder quick start and harness setup path, see Getting Started: Building and Deploying Agents on Reindeer.

• Single-tenant environment. Customer workloads can run in a dedicated tenant boundary with isolated data stores, runtime configuration, and access controls.
• Private cloud option. For customers that require stricter residency or governance boundaries, Reindeer can run in a dedicated private cloud environment.
• Workspace-level separation. Agents, artifacts, contexts, connectors, runs, and audit records are scoped to the relevant workspace.
• Role-based access control. Users and services receive only the permissions required for their role and workflow responsibilities.

Identity and Access Controls

• Bring your own IdP and SSO. Reindeer can integrate with the customer's identity provider so users are provisioned and governed through existing identity processes.
• Scoped connector access. Connectors are configured with explicit permissions for the systems and actions required by the workflow.
• Per-workflow secrets. Secrets can be scoped to a workflow, connector, or tool so an agent cannot access credentials it does not need.
• Runtime-only secret injection. Secrets are resolved at execution time by the platform and are not stored in prompts, agent instructions, or model-visible context.

What May Be Sent to an LLM

When an agent step requires LLM reasoning, classification, extraction, summarization, or drafting, Reindeer constructs a task-specific request. Depending on the workflow, that request may include:

• The current case or run input, such as an email, document text, ticket, invoice, shipment request, or customer message.
• Relevant workflow instructions, policy rules, examples, or playbook excerpts.
• Relevant artifact or context snippets selected for the task.
• Structured outputs from previous tools in the same run.

Reindeer avoids sending unrelated workspace data, unrelated historical cases, connector credentials, or broad customer data that is not required for the current task.

LLM Provider Controls

• No training on customer data. LLM provider usage is configured so customer data is not used to train or improve third-party foundation models.
• No provider-side retention where supported. Requests are routed through enterprise controls intended to disable provider-side prompt and completion retention.
• Provider abstraction. Reindeer can integrate with approved model providers and customer-approved endpoints while keeping workflow governance, logging, and access controls in Reindeer.
• Model outputs remain governed artifacts. LLM responses are treated as workflow outputs subject to Reindeer audit, review, retention, and access policies.

Auditability and Traceability

Privacy controls are paired with traceability. Reindeer records the execution path of each run so teams can understand what happened without relying on a black box.

• Agent steps, tool calls, decisions, confidence scores, escalations, and human interventions are logged.
• Reviewers can inspect the case context that led to a decision, subject to workspace permissions.
• Audit records support internal review, compliance evidence, and root cause analysis.
• Human-in-the-loop checkpoints can be added where policy or confidence thresholds require approval before action.

Retention and Customer Control

• Reindeer retention is governed by customer and workspace policy. Run history, case records, artifacts, and audit logs are retained according to the applicable configuration and contractual requirements.
• Provider retention is separately controlled. LLM provider interactions use enterprise no-training and zero-retention controls where applicable.
• Artifacts and contexts are versioned. Reference documents, prompts, and workflow context can be promoted, reviewed, rolled back, and audited across versions.

Operational Best Practices

• Use dedicated workspaces for development, staging, and production workflows.
• Keep connector and tool permissions narrow.
• Review which artifacts and context are attached to each agent revision before promotion.
• Use human-in-the-loop review for sensitive actions, regulated decisions, or low-confidence outputs.
• Confirm customer-specific LLM provider and retention settings during security review and implementation planning.

Summary

Reindeer is designed to let teams run enterprise agents while maintaining data isolation, scoped access, auditability, and LLM provider controls. Customer data is protected through tenant isolation, workspace scoping, runtime secret injection, enterprise identity controls, zero-retention/no-training LLM provider configurations, and complete run-level traceability.