AI-Powered Label Compliance Tool

Overview

Mark developed a Claude-based AI tool to evaluate food product labels against US regulatory requirements. The tool ingests large regulatory documents as context and compares a given label against them, surfacing potential compliance issues quickly and systematically.

The tool was first used in production during the ^[1] US market compliance review, where it identified multiple issues with the longevity (superfood) coffee label.

How It Works

The tool is built as a project inside Claude (referred to in the meeting as "Clawed"). The workflow is:

1. Load regulations — Relevant US food regulations are added to the Claude project as context documents. These can be very large (e.g., 18,000-line and 7,000-line documents).
2. Submit a label — A product label image or text is provided to the tool.
3. Run compliance check — The tool is prompted to evaluate whether the label is compliant, specifying the product category (e.g., food product, not supplement).
4. Review findings — The tool returns a structured list of issues, which may include:
   - Critical non-compliance issues
   - Fortification policy concerns
   - Structure/function claims issues
   - Other category-specific flags

Regulations Loaded (US Context)

As of the La Natura review, the project included four US regulatory documents covering areas such as:

• General food labeling requirements
• Allergen declarations
• Supplement and health claims regulations
• Additives, vitamins, and minerals

Multiple regulations can apply to a single product. For example, a flavored coffee with added vitamins may fall under regulations for: coffee products, nutritional declarations, vitamin/mineral additives, and flavoring agents.

Strengths

• Speed and breadth — Processes thousands of lines of regulation against a label in seconds, far faster than manual review.
• Multi-regulation coverage — Surfaces issues across several applicable regulatory frameworks simultaneously.
• Iterative refinement — Prompts can be refined mid-session (e.g., clarifying that a product is a food, not a supplement) to narrow results.

Limitations and Caveats

• May be overly sensitive — The tool can flag issues that a human expert (e.g., a specialist consultant) would consider acceptable or inapplicable. Dubravka Kukic (La Natura) noted this concern, preferring to cross-reference with consultant Christian's review.
• Not a substitute for expert review — Results should be treated as a first-pass screen, not a definitive compliance opinion. Human regulatory consultants remain the authoritative source.
• Dependent on regulation currency — The tool is only as current as the regulatory documents loaded into it. Outdated documents will produce unreliable results (a discrepancy in daily allowance percentages was encountered during the La Natura review due to differing regulatory versions).
• Label format matters — The tool works best when the label content is clearly provided; ambiguous or incomplete label data may produce incomplete analysis.

Usage Notes

• When submitting a label, explicitly state the product category and registration intent (e.g., "food product, not a supplement") to avoid irrelevant flags.
• Cross-reference AI findings against a qualified consultant's review before acting on results.
• Flavored products with no allergens and minimal additives are likely to have fewer flags than fortified or superfood-positioned products.

Evidence from Client Work

During the ^[2] call:

• The tool was run against the longevity coffee (superfood) label and returned multiple compliance issues across several categories.
• The flavored coffee label had not yet been run through the tool, but was expected to be lower-risk (flavor-only addition, no allergens).
• Dubravka acknowledged the tool's utility while cautioning that it may be over-sensitive compared to consultant Christian's manual review.

Related

• ^[3]
• ^[4]
• ^[5]