Responsible AI and AI Disclosure
1. What runs autonomously
Food Line AI operates eight named Routines that run on a schedule: morning briefing, reorder analysis, A/R collection drafts, expiration scanning, route optimization, purchase order drafts, customer re-engagement campaigns, and margin analysis. Each Routine reads operational data, performs analysis, and produces a draft or recommendation.
Everything that runs unattended is read-only, a notification, or a task. The system can scan, analyze, draft, and recommend without human involvement. It cannot execute.
2. What never runs without a human
Nothing that moves money or stock runs unattended. That gate is structural, not configurable. A named person must approve any action that posts an invoice, sends a collection email, creates a purchase order, adjusts inventory, modifies pricing, or initiates a shipment. This is a product rule, not a setting.
3. How the Copilot answers
The Copilot answers questions by reading the customer's own operational data. It shows its sources at the record level, so you can verify every answer against the underlying data. The Copilot does not hallucinate inventory counts or fabricate order histories because it is grounded in the live database, not in a general knowledge model.
4. Model providers and data boundaries
Customer operational data is sent to our model provider for inference when the Copilot or a Routine processes a request. The current list of sub-processors, including the model provider, is maintained at /subprocessors.
5. Whether customer data is used to train models
We do not use customer data to train models. We are confirming upstream retention terms with our model provider to ensure this commitment extends through the full processing chain.
6. The immutable audit log
Every tool call and decision made by the AI system is recorded in an immutable audit log. Posted invoices and inventory movements cannot be edited. Corrections are made by linked reversal, preserving the complete record. Role-scoped visibility ensures that cost and margin data is accessible only to authorized roles.
7. AI usage metering
AI usage is metered and visible to the customer. The platform tracks the number of Routine runs, Copilot queries, and inference calls, so customers can understand and manage their AI consumption.
8. Known limitations
AI outputs should be reviewed by a qualified person before acting on them. The system may occasionally produce recommendations that do not account for context only a human operator would know, such as a customer's verbal agreement, a supplier relationship nuance, or a seasonal pattern not yet reflected in the data. The approval gate exists precisely because of these limitations.
9. Reporting an AI output that was wrong
If you encounter an AI output that is incorrect, misleading, or harmful, report it to hello@foodlineai.com with the details of what was produced and what was expected. We review every report and use them to improve the system.
10. Compliance posture
Every layer that holds customer data is SOC 2 Type II attested. Supabase, which also holds ISO 27001. Clerk. The cloud and edge providers. Food Line AI's own Type II examination is the next milestone, with a Type I available sooner as an interim artifact.