fraud detection

Your fraud model flags 1,000 transactions a day. 900 of them are false positives.

WhiteBox runs every transaction through multiple AI models. When they all say fraud, block it. When they disagree, route it to your fraud team with the full breakdown. Stop blocking good customers. Start catching real fraud.

the problem

What single-model fraud detection gets wrong

01
False positives kill revenue

Example: Customer buys a $2,100 luxury handbag as a gift. Ships to a different address. VPN detected (they are at work). Single model flags it as fraud. Customer gets blocked, calls support, closes their account.

WhiteBox: 3 models say legitimate, 1 says suspicious. Consensus: let it through. The one dissenter gets logged for pattern review.

02
Real fraud slips through

Example: Small $47 test charge from a new device at 2AM, followed by a $890 electronics purchase 4 minutes later. Amount is under the threshold. Single model says legitimate because the first charge was small.

WhiteBox: 2 models flag the pattern, 2 miss it. Disagreement triggers escalation. Fraud team catches it before the $890 ships.

03
The gray zone

Example: Third laptop claim in 14 months. Different shipping addresses each time. Customer has a legitimate 8-year account history.

One model says fraud (pattern), one says legitimate (account age), two say suspicious. No single model captures the full picture. WhiteBox shows the split and lets your fraud analysts decide.

how it works

Multi-model consensus in action

whitebox fraud scoring
blocked
whitebox classify "$4,847 at electronics retailer, Miami FL. Card issued Portland OR. 3:42 AM. 3 charges in 1 hour across 2 states. Account age: 14 days."
options: ["legitimate", "suspicious", "fraudulent"]
01gpt-4o-minifraudulentlogp -0.02
02claude-3.5fraudulentlogp -0.01
03llama-3.3fraudulentlogp -0.04
04deepseek-v3fraudulentlogp -0.03
verdict
fraudulent · confidence 99% · BLOCK
SHIP
auto-action: decline transaction · alert fraud team
whitebox fraud scoring
approved
whitebox classify "$312 hotel booking. Card issued Chicago. Customer traveling (flight booked 2 days ago to Vegas). 11:15 PM. Account age: 2 years. Avg monthly spend: $1,800."
options: ["legitimate", "suspicious", "fraudulent"]
01gpt-4o-minilegitimatelogp -0.18
02claude-3.5legitimatelogp -0.12
03llama-3.3suspiciouslogp -0.67
04deepseek-v3legitimatelogp -0.22
verdict
legitimate · confidence 87% · APPROVE
SHIP
note: 1 model flagged suspicious -- logged for pattern review

Every run, every log-prob, every disagreement -- recorded. Replay any decision from its ID.

use cases

Anywhere money moves, you need fraud detection

01
Payment fraud

Score every transaction in real time. Block confirmed fraud, approve clear transactions, escalate the gray zone.

02
Account takeover

Detect suspicious logins, password changes, and shipping address updates. Multi-model consensus catches patterns single models miss.

03
Chargeback prevention

Flag high-risk transactions before they ship. Reduce chargebacks by catching fraud early without blocking legitimate customers.

04
Return/refund abuse

Detect serial returners, wardrobing, and refund fraud patterns. Escalate suspicious patterns for manual review.

05
Promo/coupon abuse

Identify fake accounts created to exploit promotions. Multi-model agreement catches sophisticated abuse that single models miss.

06
Seller fraud (marketplaces)

Flag suspicious seller behavior: fake listings, shill reviews, dropshipping from prohibited sources.

the real cost

The math on false positives

01

Every false positive costs you $150-300 in lost revenue + support costs.

02

At 1,000 flagged transactions/day with 90% false positive rate, that is $135,000-270,000/month in lost revenue.

03

WhiteBox reduces false positives by routing disagreements to humans instead of auto-blocking.

04

Even a 20% reduction in false positives pays for WhiteBox 100x over.

by the numbers

Fraud detection that pays for itself

99%
true fraud caught
when all models agree, it is real
60%
fewer false positives
disagreement = review, not auto-block
$0.01
per check
vs $0.50-2.00 for manual review
< 2s
real-time scoring
fast enough for checkout flows
comparison

WhiteBox vs rules-based and single-model fraud detection

Feature Rules-based Single AI model WhiteBox
False positives High (rigid rules) Medium (one opinion) Low (consensus)
Sophisticated fraud Missed (no pattern learning) Sometimes caught Caught by disagreement
Gray zone Auto-block or auto-allow Random Routed to fraud team
Explainability Rule name Black box "3 of 4 models said fraud"
Audit trail Rule log No Every model vote logged
Real-time Yes Yes Yes (< 2 seconds)
playground

Try it. Describe a transaction, see the verdict.

legitimate suspicious fraudulent
whitebox sandbox · simulated client-side
[--:--:--] waiting · press score transaction to dispatch
models
4
median latency
1.2s
cost / check
$0.01
audit retention
forever
pricing

$0.01 per fraud check

20 free to start. No credit card.

That's 1,000 fraud checks for $10.

free tier
20 checks
per check
$0.01
subscriptions
none
get a key
get started

Stop blocking good customers. Start catching real fraud.

20 free checks. Then $0.01 each. Full audit trail from day one.

get a key API docs