AI Data Use Policy ================== Last updated: 2026-08-05 Legal AI Data Use Policy ------------------ Last updated: August 5, 2026 · Effective: August 5, 2026 This AI Data Use Policy ("Policy") explains how Owndevice, Inc. ("Grounds") handles the inputs and outputs of the AI features of Grounds (the "Service"). It supplements the Terms of Service, the Privacy Policy, and our Data Processing Addendum (available on request from legal@grounds.ai). Capitalized terms not defined here have the meaning given in the Terms. 1. Core promise We do not use Customer Data to train, fine-tune, distill, re-rank, or build evaluation sets for any AI model — neither ours nor any third party's. This applies to documents, redlines, comments, prompts, AI outputs, embeddings, and any derivative of those materials. Our inference partner contracts are configured for zero retention: the inputs we send leave our infrastructure only to obtain an inference result and are not stored, logged for product improvement, or used to train the partner's models. As of the effective date, our sole inference provider is AWS Bedrock under terms that explicitly prohibit training on, or retention of, prompts and outputs. 2. Models and providers All AI inference runs exclusively through AWS Bedrock in United States regions. We use two model families: Anthropic Claude models for primary drafting, review, and chat features, and Moonshot AI Kimi models for cost-efficient background tasks. Because every request is served by AWS Bedrock under zero-retention terms, your prompts and outputs are not shared with Anthropic or Moonshot AI (the model developers) and are not used to train their models. Firms may request Claude-only processing, which routes every task to Claude models; contact privacy@grounds.ai. 3. Tenant isolation AI inputs and outputs are scoped to the customer tenant (Clerk organization) that produced them. Embeddings, vector indexes, prompt logs, and any cached intermediate state are partitioned by tenant. No cross-tenant retrieval-augmented generation, fine-tuning, or analytics is performed against Customer Data. 4. Human access Grounds personnel do not review AI inputs or outputs in the normal course of operating the Service. Human access to a tenant's AI data occurs only under one of the following, each of which is logged to an audit trail that the customer can request: • Least-privilege incident response (an active production issue or a confirmed security event that the engineer must inspect to resolve). • A specific customer support request initiated by the customer where access is necessary to fulfill the request. • Compliance with a lawful order, in which case we will notify the customer unless legally prohibited. Access is limited to the minimum data necessary, time-bounded, and revoked when the work is complete. 5. Model improvement Any future program to use Customer Data to improve Grounds-operated models will be opt-in only, separately disclosed, controllable at the tenant level by a customer administrator, and reversible. As of the effective date, no such program is active and the relevant controls default to off. Aggregated, deidentified operational metrics (e.g., latency percentiles, tokens-per-request distributions, feature usage counts) are permitted and used to operate and improve the Service. These metrics are constructed so they cannot be reverse-engineered to recover Customer Data. 6. Prompt and response logs Operational logs of prompts and responses are scoped to the originating customer tenant and retained while the account is active. They are deleted when the account is deleted, and on a verified deletion request to privacy@grounds.ai. These logs exist solely for debugging, rate-limit enforcement, abuse prevention, and incident response. They are not used to train models and are not shared outside Grounds (other than to our infrastructure sub-processors as listed in the Subprocessors page). 7. Embeddings and vector stores Where the Service generates embeddings (for retrieval, similarity search, or playbook matching), the resulting vectors are stored encrypted at rest in our managed Postgres instance, partitioned by tenant, and deleted when the source content is deleted or when the tenant itself is deleted. Vectors are derived from your content and we treat them with the same confidentiality as the underlying Customer Data. 8. Output disclaimer and verification AI outputs — including redlines, summaries, citations, statutory references, and risk flags — can be incorrect, incomplete, or misleading. They are not legal advice. You must independently verify any output before relying on it, and a qualified attorney must verify any citation used in a legal proceeding or transaction. The cross- reference rules in the Terms and the AUP apply. 9. Provenance and transparency AI tasks are routed to a specific model version by a server-side configuration (see Section 2 for the model families in use). On request, we will identify the model family and version used for a given feature and time period. Contact privacy@grounds.ai. 10. Agent actions Where the Service acts as an agent on behalf of the customer (for example, drafting and sending a counter-offer to a counterparty portal), additional rules apply under the Agent Authorization Terms. 11. Changes We will post any material change to this Policy here and notify active customer administrators at least 30 days before the change takes effect. Reductions to a customer's protections require explicit customer consent. 12. Contact Questions about how AI data is handled? privacy@grounds.ai.