Real-time fraud scoring
Every donation, scored in under 200ms
Anthropic's Claude evaluates every Stripe transaction with two hours of behavioural context and returns a risk level, a 0–100 score, and a plain-English reasoning string your finance team can read.
Recurring donor, 11-month run. Established cadence; charge fits historical pattern.
High-entropy local-part, free-mail domain. Sixth A$1 attempt from this card BIN in 14 minutes. Decline codes alternating do_not_honor / incorrect_cvc. Card-testing pattern.
First-time donor, unusual round amount, country mismatch (card AU, IP DE). Worth a human review before refund/dispute.
Two hours of context, not just one charge
Stripe Radar scores transactions in isolation. The Cause Shield classifier sees the donor's preceding two hours of transactions on the same IP, card BIN, and fingerprint when it makes the call. That's the difference between "this charge looks fine on its own" and "this is the seventh A$1.00 attempt on this card in twelve minutes."
Recurring donations re-scored on every renewal
Stripe Radar only ML-scores the first charge of a subscription. The Cause Shield classifier re-evaluates every recurring invoice — which is exactly where friendly-fraud chargebacks slip past Radar months after the original gift. A donor who set up A$50/mo in 2023 and disputes a 2025 renewal looks like a clean recurring renewal to Radar; the classifier surfaces the cadence break.
Plain-English reasoning, not a black-box score
Every flag carries a one-paragraph explanation in the language a finance manager can act on. "High-entropy local-part on a free-mail domain, six donations under A$5 in the last hour from the same Cloudflare-hosted IP, decline codes alternating between do_not_honor and incorrect_cvc — pattern matches card-testing." Your operations team makes the decision; you stay out of compliance scope.
Bounded cost, bounded blast radius
Every plan includes a monthly allowance of AI fraud analyses (Starter 2,000, Growth 10,000, Partner 50,000, Enterprise unmetered). When you exceed your allowance, monitoring continues uninterrupted on a deterministic rule-based fallback — card-testing patterns, round-number amounts, velocity spikes — at no extra charge. AI scoring resumes on your next billing period or via a one-time top-up pack.
How customers use it
Three scenarios where this lands
Scenario
Stopping a card-testing wave before sunrise
A bot starts a card-testing campaign against your donation form at 02:00. The classifier flags the third decline as critical; the alert lands in your on-call Slack channel. By the time your team wakes up, the attack is documented, the donor pattern is logged, and your finance lead has a paragraph they can quote in the Stripe support ticket.
Scenario
Catching a friendly-fraud recurring chargeback early
A donor disputes their A$25/mo renewal in month 18. The classifier sees the cadence break against a previously-clean subscription and flags it as recurring_renewal_anomaly. Your finance team has the context to refund before Stripe formally opens the dispute — protecting your win rate.
Scenario
Justifying a fraud-related refund to the board
A flagged charge needs board sign-off to refund. The classifier's reasoning string goes directly into the board pack. Your treasurer doesn't have to translate Radar score 87 into English — they can read the paragraph and approve.
Try Real-time fraud scoring on a 14-day free trial.
Available on every plan, including Starter.