Supplier metrics and validated business rules are processed before information reaches the language model. The AI receives structured supplier context and uses it to explain risk drivers, surface evidence, and recommend investigation actions.
<aside> 💡
Core principle: Use deterministic logic for what can be calculated reliably, and AI for interpretation where context and reasoning add value.
</aside>
1. Supplier Data - Supplier quality, delivery, inventory, and related operational information serve as the source of truth. The AI does not create or modify these underlying supplier facts.
2. Deterministic Layer - Metrics that can be objectively calculated are processed outside the language model. Examples include defect rates, lead-time deviations, inventory conditions, and supplier exposure. Validated business rules can also be applied here.
3. Context Builder - Relevant supplier signals are converted into a structured context object before being sent to the model. This keeps the model focused on the information required for the current supplier investigation rather than sending unnecessary data.
4. AI / LLM - The model interprets the structured context to produce the AI Risk Brief. Its responsibilities include:
Risk drivers → Evidence → Healthy signals → Recommended investigation actions → Data limitations
The model interprets supplier information. It does not become the source of truth.
5. Output Validation - Before the response reaches the user, the system checks whether the expected output structure is present and whether the response respects available evidence and missing information. Invalid or unsupported outputs should be flagged rather than presented as reliable conclusions.
6. Human Review - The procurement user reviews the evidence and recommended actions before deciding what to do next.
<aside> 💡
The Copilot supports the decision. The human owns the decision.
</aside>
| Layer | Primary responsibility |
|---|---|
| Data | Provide authoritative supplier information |
| Deterministic Logic | Calculate objective metrics and validated rules |
| Context Builder | Select and structure relevant information |
| AI / LLM | Interpret signals and recommend investigation actions |
| Validation | Check grounding, structure, and missing-data behavior |
| Human | Review evidence and make consequential decisions |
| Evaluation | Monitor quality, failures, feedback, latency, and cost |