Answers / AI & data

    What are the 4 risk levels in AI receipt fraud detection?

    Written by PEKO Team.Last updated: 05/24/2026.

    Low (flagged, auto-credited), Medium (flagged, queued for review), High (blocked pending review), Critical (auto-blocked, member warned). Thresholds combine duplicate-image score, scan frequency per hour, receipt age, and total amount.

    Published: 05/24/2026

    Receipt fraud is real — a small share of members try to upload the same receipt twice, an old receipt, or a screenshot from a friend. A 4-level risk model lets you stop the abuse without inserting friction for the 95% of honest guests.

    The signals that matter: perceptual-hash similarity to recent uploads (duplicate detection), uploads per hour per member, receipt age in days, and total amount relative to the member's normal ticket.

    Low risk: auto-credit, log

    Single weak signal (e.g. receipt 6 days old). Credit points, mark for batch review.

    Medium risk: queue

    Two weak signals or one moderate (e.g. 4 uploads in 1 hour). Queue for operator approval — usually resolved within the hour.

    High risk: block + review

    Strong duplicate-hash match or receipt over 14 days old with high amount. Block pending operator review; member sees 'verifying' state.

    Critical: auto-block, notify

    Confirmed duplicate of a previously credited receipt. Auto-block, log, notify owner via email and in-app.

    FAQ

    Won't this annoy honest members?

    If thresholds are tuned to the venue, fewer than 1 in 200 honest uploads land in Medium or higher. The Low band absorbs most edge cases silently.

    Can the thresholds be configured?

    Yes. PEKO exposes per-brand thresholds for uploads-per-hour, max receipt age, and amount caps. Defaults work for most cafés and casual restaurants out of the box.

    Calculate your PEKO ROI

    Related

    People also read