AI & Model Data Use Standard
AGRO-AI uses intelligent systems to reason over agricultural context. This standard defines the boundaries around inference, retrieval, training, provider access, provenance and human control.
1. Separate the activities
AGRO-AI distinguishes model inference, retrieval-augmented generation, embeddings, evaluation, product analytics, fine-tuning and model training. Data being processed by a model for an authorized request does not by itself make that data an unrestricted training asset.
2. Tenant and authorization boundaries come first
Identity, organization, project and connector authorization should be evaluated before customer evidence is retrieved for an intelligent workflow. A language model is not relied upon to enforce tenant isolation or access policy.
3. Customer Data and shared-model training
Customer Data is not automatically designated for training a shared AGRO-AI or third-party model merely because the data is processed by an intelligent feature. Any customer-data use for training or generalized model improvement must have an approved purpose, a valid legal and contractual basis, appropriate provider controls, defined retention and an evidence-backed customer disclosure.
4. Provider governance
Before a model provider is used with sensitive production data, AGRO-AI's governance process should evaluate the provider's retention, training/use terms, security controls, subprocessor position, data location where relevant, contractual protections and available enterprise configuration. Provider settings must match the public and contractual statements AGRO-AI makes to customers.
5. Secrets do not belong in model context
Passwords, API keys, signing keys and connector credentials should not be inserted into prompts or model context. Tools and connectors should enforce authorization outside the model and accept only the parameters needed for the approved action.
6. Evidence is not instruction
Documents, transcripts, files, webpages and other retrieved evidence may contain malicious or misleading instructions. AGRO-AI's architecture treats customer evidence as data, not as authority to override system policy, permissions or tool controls.
7. Derived intelligence
Recommendations, summaries, classifications, alerts and other model outputs may contain errors. Where outputs can materially affect an agricultural operation, product design should preserve source context, confidence or evidence where available and require appropriate human authorization before consequential execution.
8. Provenance
Where technically supported, AGRO-AI records enough source and workflow context to distinguish observation from inference and to show whether a recommendation was proposed, scheduled, applied or verified.
9. Evaluation
Models and prompts may be evaluated using synthetic, public, licensed, consented or appropriately governed data. Evaluation datasets containing Customer Data require an approved purpose and access controls. Benchmark results should not be represented more broadly than the tested conditions support.
10. De-identification
De-identified or aggregated datasets may be used where permitted, but the strength of de-identification must be evaluated against agricultural context. A combination of geography, crop, timing and operational attributes can sometimes re-identify an operation even when obvious account identifiers have been removed.
11. Human control and action tools
Consequential actions should be gated by explicit authorization, typed tool interfaces and non-model controls. The model may help prepare or recommend an action; product controls determine whether the action is allowed to execute.
12. Changes to AI data use
Material changes to model providers, training use, retention or customer-data treatment should trigger governance and legal review before updated claims are published.
13. Questions
Questions about AI data use may be sent to contact@agroai-pilot.com.
Back to Trust Center