AI & Translation Policy
MPBxChange uses AI for specific, bounded tasks in the procurement workflow. This policy discloses where, what, and the limits of those outputs in compliance with EU AI Act Article 50 (transparency obligations for limited-risk AI systems) and the spirit of analogous regimes in other jurisdictions.
1. Where AI is used on the platform
The platform uses AI in seven scoped surfaces:
- Counsel contract review19-layer deterministic detector pipeline plus the Anthropic /review-contract skill as Layer 11. Flags spec drift, missing schedules, entity mismatch, bilingual divergence, predatory clauses, BOI incentive applicability, etc.
- AI clause bias analysisPer-clause review surfaced on contract pages. Indicates whether a clause favors buyer / supplier / neutral.
- Holistic MatchmakerSupplier ranking for an RFQ. 5-signal composite (spec fit + cert stack + CBAM + lead time + capacity) × risk-band modifier.
- Predictive risk scoring6-signal composite (drift + counsel + violations + sanctions + payment + concentration) producing a 0-100 score with green / amber / red band.
- Tariff arbitrageLanded-cost scenario engine across alternative origins for the same HS-6 product.
- TranslationAI translation of user-generated content surfaces (RFQ titles, supplier descriptions, contract clauses) when reader and writer use different languages.
- Float underwriting (advisory)Indicative quote on /buyer/float; final lending decision is made by a partner lender, not by MPBxChange AI.
2. What the outputs are — and are not
Every AI output is advisory. None of the following is true:
- Not a legal opinionCounsel flags items for review. It does not render binding legal opinions or replace counsel review by a qualified attorney.
- Not a credit decisionThe Float quote is indicative. The actual lending decision is made by a partner lender per their own underwriting.
- Not a tax recommendationTariff arbitrage shows landed-cost scenarios. Tax classification and customs filing remain the buyer’s responsibility.
- Not a guarantee of complianceCBAM and other regulatory flags are pattern detection. Final compliance is the importer’s legal duty.
3. Disclosure surface
Every AI-generated output on the platform carries a visible disclosure: either an inline AI · NOT WARRANTED badge or a block-level banner labelled AI DISCLOSURE · EU AI ACT ARTICLE 50. Where evidence supports the output, the disclosure includes a link to that evidence (e.g. cited contract clauses, source documents, calibration dates).
4. Models, providers, retention
- Primary providerAnthropic Claude (frontier model for Counsel + clause analysis). Fallback chain: OpenAI · DeepSeek when the primary is unavailable.
- Data sent to providersOnly the content needed for the specific task (e.g. the contract clauses being reviewed). User account identifiers are NOT sent to the AI provider; redaction occurs at the API gateway.
- RetentionProvider-side per their contract (Anthropic zero-retention enterprise plan). MPBx-side: AI outputs are cached for replay + audit but redacted PII is never stored beyond the contract record.
- TrainingCustomer data is never used to train any AI model. We use API endpoints with explicit no-training data clauses.
5. Your rights
- Right to human reviewYou can request human review of any AI-generated flag or recommendation. Contact hello@mpbxchange.com.
- Right to opt outYou can opt out of AI-driven translation of your user-generated content. The default is opt-in for the convenience of cross-language buyer-supplier communication.
- Right to evidenceYou can request the evidence cited by any AI output via the link in the disclosure or by ticket.
- Right to deletionYou can request deletion of stored AI outputs related to your account via /privacy/my-data.
6. Updates
This policy reflects the platform’s state as of 22 May 2026. Material changes will be announced via in-app banner and through the email on file. The most recent version always lives at /legal/ai-disclosure.