AI transparency

EU AI ACT · NYC LL 144

FreshCut uses a single Claude Haiku 4.5 model across 16 distinct sub-roles. Every AI invocation falls into one of these surfaces:

• Support triage — categorising new support tickets by topic + urgency
• Review moderation — screening new customer reviews for profanity, PII, threats
• Listing-copy assistance — generating product descriptions for vendor onboarding (vendor opt-in)
• Ad-copy generation — suggesting headline + body for sponsored listings (vendor opt-in)
• Onboarding coach — answering vendor questions during their first 30 days
• Revenue insights / AI conversational analytics — answering vendor questions about their own LTV / churn / cohort data
• Document review — pre-screening compliance document uploads (e.g. food handler license)
• Driver ETA + fraud anomaly + dispute mediator + HPAI monitor + churn predictor + compliance auditor — operational ML scoring

Every AI invocation sends aggregate statistics and the specific user-supplied text being scored (a review body, a support ticket, the vendor's typed question). We never send:

• Customer payment card information
• Driver background-check details
• Customer addresses or geo-coordinates
• Individual customer-PII columns (email, phone, name) — the conversational analytics endpoint scrubs these before calling the model
• Photos or attached media

All payloads sent to the model are logged to our internal audit log under AI_INVOCATION actions and retained for 30 days for incident review only.

No AI decision affects pricing, order eligibility, payouts, or driver pay. Those are all rule-based and human-auditable.

AI is used as a recommendation layer for human reviewers (moderation queue, support triage, fraud flags). Every negative AI decision can be appealed via the support flow at /support and a human will respond within 1 business day.

Provider terms: anthropic.com/legal

You can opt your account out of optional AI surfaces (review moderation suggestions, listing copy assistance, AI insights) at any time:

1. Go to /profile → Privacy
2. Toggle "Allow AI-assisted features" to OFF
3. Save

Note: a small set of safety-critical AI uses (review moderation for hate speech / PII leakage) cannot be opted out of because they protect other users. The full list is disclosed above.