Responsible AI Policy

Responsible AI Policy

Closed-beta legal boundary: Founder-prepared draft for closed-beta readiness. Lawyer/CA review required before paid/public production. PGI Hub has no approved live payments, GST invoicing, vendor payouts, escrow, public marketplace launch, automated supplier award, or production legal approval under this pack.

1. Master Overview and Boundary

This Responsible AI Policy governs PGI's use of AI-assisted and rule-based systems for procurement data extraction, BOM normalization, supplier matching, ranking explanations, quote comparison, total landed cost assumptions, review flags, procurement reports, and user-facing decision-support outputs. It binds PGI operators, buyers, vendors, and implementation teams, and it does not approve autonomous supplier awards, automated legal/tax/compliance conclusions, contact disclosure, payment actions, or production-grade AI claims.

2. Operational Definitions

• AI-Assisted Output means a classification, extraction, summary, mapping, score, label, warning, recommendation, or explanation generated using AI, rules, heuristics, retrieval, provider/API data, internal taxonomy, or structured procurement logic.
• Human Review Queue means a workflow where low-confidence, high-impact, conflicting, stale, or disputed outputs are routed for review before reliance or action.
• Confidence Score means a platform indicator of estimated output reliability; it is not a guarantee.
• Review Flag means a warning attached to ambiguous part names, uncertain MPN aliases, missing specifications, stale vendor data, unsupported certifications, or conflicting quote/TLC fields.
• AI Governance Record means retained evidence of model/prompt/version, input class, output class, confidence, review decision, and correction history where technically feasible.

Platform Informational Inferences means outputs generated, derived, normalized, classified, ranked, estimated, compared, summarized, or otherwise presented by PGI using platform logic, user-provided data, vendor-provided data, public-source data, third-party data, provider/API data, AI-assisted analysis, rule-based logic, or internal procurement intelligence methods.

Platform Informational Inferences are not professional, engineering, legal, tax, logistics, insurance, customs, financial, or procurement advice. They are decision-support outputs only. The buyer, vendor, or contracting party remains exclusively responsible for independent validation before relying on such outputs for commercial, technical, legal, tax, logistics, compliance, or contractual decisions.

3. Core AI Compliance Architecture

PGI AI must operate as a decision-support layer. It should support extraction, classification, normalization, comparison, summarization, and workflow prioritization while preserving human validation for procurement decisions. AI outputs must be versioned or traceable where feasible, confidence-scored, visibly caveated, auditable in sensitive workflows, and connected to correction routes.

AI must not be represented as replacing engineering review, supplier due diligence, legal review, tax/GST review, customs classification, quality inspection, logistics planning, commercial negotiation, contract review, or final procurement approval.

4. High-Risk Output Boundaries

PGI must not allow AI outputs to automatically send RFQs, award suppliers, expose hidden contacts, create purchase orders, activate payments, trigger vendor payouts, confirm tax treatment, guarantee stock, guarantee price, guarantee delivery, guarantee certification validity, or make legally binding procurement decisions without reviewed human approval workflows.

5. Required Controls by AI Use Case

6. Bias, Fairness, and Organic Ranking

Supplier ranking should evaluate procurement-relevant signals such as category fit, capability fit, geography, quantity fit, certifications, freshness, responsiveness, and data confidence. Organic ranking must not use commission, paid subscription tier, credits, sponsored status, payout potential, or PGI commercial relationship.

Where ranking is shown, PGI should disclose material factors and warnings without exposing proprietary logic or enabling manipulation.

7. Data Quality and Correction

AI outputs must be correctable. A vendor can challenge profile categories, certification labels, capability matches, rates/offers, stock, lead time, and contact status. A buyer can correct BOM fields, MPN aliases, quantities, specifications, and quote comparison assumptions. Corrections should feed data-quality records, not silent unsupported overwrites.

8. Product UI and Acceptance Mapping

9. Monitoring and Audit

PGI should retain AI governance records for high-impact workflows, including model/prompt version where feasible, source data class, confidence, warning flags, human review decision, correction history, and user acceptance context. Internal review should prioritize low-confidence outputs that could influence supplier selection, RFQ sending, quote comparison, contact access, certification status, or vendor visibility.

Related Documents

• terms-and-conditions
• privacy-policy
• closed-beta-terms
• acceptable-use-policy

Review Status

Founder-prepared draft. Lawyer/CA review required before paid/public production.

Paid/public production readiness: Not approved. Lawyer/CA/payment-provider/founder review required.