Legal
LLM Sub-Processor Data Flow
1. Purpose
This page provides data-flow transparency for the LLM sub-processors referenced in our Privacy Policy and Data Processing Agreement. It describes what data Intended may transmit to those sub-processors during intent compilation, how that data is constrained, the contractual safeguards Intended has executed, and how customers can opt out of LLM-assisted processing entirely. Publishing this page satisfies a commitment made in both the Privacy Policy and the Security Addendum; Intended will update it whenever the set of LLM sub-processors, the data categories transmitted, or the contractual posture changes.
2. Current LLM sub-processors
As of the Last Updated date of this page, Intended uses the following LLM sub-processors for the bounded purpose described in Section 3 below:
- OpenAI, L.L.C. — https://openai.com — United States — purpose: intent compilation and classification where the natural-language intent cannot be deterministically parsed by Intended's rule-based backend.
- Anthropic, PBC — https://anthropic.com — United States — purpose: intent compilation fallback and confidence repair when OpenAI is unavailable or its output fails Intended's confidence threshold.
3. What data is transmitted
Only the natural-language intent string submitted by the customer's agent is transmitted to the LLM sub-processor, together with a fixed system prompt that constrains output to Intended's structured intent schema. The following data categories are NEVER transmitted to LLM sub-processors:
- Authority tokens, credentials, API keys, or any other shared secret.
- Audit ledger entries, decision records, risk scores, or policy evaluation traces.
- Personally Identifiable Information (PII) beyond what is present in the natural-language intent string itself (names, email addresses, or ticket identifiers inside a customer-authored prompt).
- Connector credentials, third-party API tokens, or any material stored in Intended's Secrets Manager or KMS-encrypted stores.
- Historical intents, prior decisions, or any cross-customer data.
- Customer billing, contract, or account metadata.
- Any data labelled 'Sensitive' by a deployed domain policy pack unless the policy pack explicitly opts into LLM processing for that data category.
4. Processing boundary
LLM sub-processor calls occur only inside the Intent Compilation stage of the authority runtime and only when the deterministic (rules-based) backend fails to produce a confident classification. The LLM response is parsed into Intended's structured intent schema, validated against a confidence threshold, and discarded from memory at the end of the request. No LLM response material is persisted to the audit ledger; only the structured, post-compilation intent record is persisted. If the LLM response fails the confidence threshold, the request is fail-closed (see docs/security/threat-model.md) and escalated to human review rather than executed.
5. Contractual safeguards
Intended has executed a Data Processing Addendum with each LLM sub-processor. Both DPAs incorporate the following commitments:
- No training, fine-tuning, evaluation, or model improvement on Customer Data. Both providers are contractually prohibited from using data submitted via Intended's API to train or improve any model, whether by default or as an opt-in.
- Retention limited to the processing window. Data is retained only for the duration necessary to return a response and is then discarded. Provider-side abuse-monitoring windows do not exceed 30 days.
- Sub-sub-processor flow-down. Both providers require any downstream sub-processors to meet equivalent data-handling obligations.
- Breach notification. Sub-processor must notify Intended within 72 hours of confirmed breach affecting Customer Data.
- Deletion on termination. Sub-processor must delete all Customer Data within 90 days of contract termination, subject to any legal-hold obligations.
6. Opt-out mechanism
Customers may opt out of LLM-assisted intent compilation in two ways. Both are supported without contract amendment; the opt-out applies to the entire tenant and may be enabled or disabled by any tenant Owner or Admin role.
- Per-tenant feature flag: in the Operator Console, navigate to Settings → Privacy → 'LLM-assisted intent compilation' and toggle off. When off, intent compilation falls back to rules-only mode. Intents that the deterministic backend cannot classify are fail-closed and escalated for human review rather than routed to OpenAI or Anthropic.
- Per-request header: inbound API requests may include the header `X-Intended-Disable-LLM: 1` to disable LLM fallback for that specific request. Useful for testing or for a tenant that wants to batch-process without any LLM traffic while keeping the feature enabled for interactive use.
7. Changes and customer notice
Intended treats changes to the LLM sub-processor list, the data categories transmitted, or the opt-out mechanism as material changes to the Subprocessor List, subject to the 30-day advance-notice and objection-rights procedure defined in the Subprocessor List (/legal/subprocessors, Section 2). Customers subscribed to subprocessor notifications will be notified by email; material changes are also logged in the Change Log section of the Subprocessor List.
8. Contact
Questions about this page, about LLM sub-processor data handling, or about exercising the opt-out may be directed to dpa@intended.so. Data-subject rights requests (access, deletion, rectification) follow the process documented at /legal/privacy and are fulfilled within the statutory windows described there.