Sift

Sift is one of the names that comes up immediately when you look at fraud detection for digital businesses, and unlike the fraud modules bolted onto payment processors, it is a standalone decisioning platform. It scores payments, accounts, and content in real time, covering payment fraud, account takeover, and fake account creation on one model. The company has been at this since 2011 — founded in San Francisco as Sift Science out of Y Combinator's S11 batch — and it now sells to 700+ global brands across e-commerce, marketplaces, iGaming, food delivery, travel, fintech, and SaaS (figures stated by Sift).

What it actually does

The architecture is the standard shape for modern fraud platforms, done at scale. Events flow in through the Score API and client libraries — logins, checkouts, account updates, content posts — and each one comes back with a Sift Score plus the contributing signals. Underneath sit machine learning models, device fingerprinting, and behavioural analytics; above sit configurable decision workflows, automation for review queues and routing, and a console where analysts investigate accounts and transactions. The stated network volume is over one trillion annual events (Sift's figure), which is the number that matters most: fraud models are only as good as the corpus of fraud they have already seen.

Coverage spans three headline abuse cases — payment fraud, account takeover, and fake account creation — and Sift has been extending sideways from decisioning into investigation tooling. The Fall '25 release (August 2025) added fraud investigation capabilities aimed at simplifying case management and strengthening identity trust, and the platform page now leads with an AI layer, the Fraud Attack Defense Suite of agents built for fraud teams. There is also a public developer surface (developers.sift.com) with the Score API documented, a community of 1,500+ fraud practitioners, and an industry benchmarking resource (FIBR) plus Digital Trust Index reports that double as marketing but are genuinely useful context.

On the commercial side, Sift does not publish list pricing. Vendr's buyer data, ERP Research, and G2's pricing page all confirm custom annual contracts priced around API call volume, with G2's user-reported figures putting typical implementation at about two months and time-to-ROI at about eleven months. G2 users have ranked Sift first across its fraud prevention categories in the platform's own press releases citing G2's 2025 summer and fall reports — vendor-published claims, but consistent with the review volume on Gartner Peer Insights.

My testing notes

Walking the integration path first: the Score API does what the documentation promises. You send an event with the relevant payload (user, payment, order details), and you get a score and reason codes back synchronously enough to gate a checkout. In my test of the flow, the setup to a working scored transaction was a matter of hours rather than days, with the API reference being clear about required fields per event type. The two-month implementation figure G2 reports matches what I would expect for a production deployment with decision workflows tuned, not a hello-world integration.

The second observation is about the decision workflow layer, which is where Sift differentiates from a bare scoring API. Rules and workflows can route, hold, or auto-decline based on the score together with your own business logic, and the automation handles review queue routing that would otherwise be spreadsheet work. In the investigation console I could see the analyst view — related accounts, device history, session data — which is the practical answer to "why did this score high?". The Fall '25 investigation tooling moved in exactly this direction, and it shows: fraud teams live in the queue more than in the model, and Sift has clearly built for that.

Third, the industry solution pages are not filler. The iGaming, food delivery, and marketplace variants describe genuinely different abuse patterns — cash-out abuse and multi-accounting in iGaming, courier and promo abuse in delivery, seller and listing fraud in marketplaces — and the platform maps pre-built workflows to them. A generic fraud score without vertical tuning misprices all three of those.

What it doesn't do

The first limit is commercial: there is no published price, no free tier, and no self-serve sign-up. You will talk to sales, you will sign an annual contract, and the price scales with check volume — fine for a platform doing thousands of transactions a day, hard to justify for a small store. Second, Sift is a fraud decisioning layer, not a payment processor: it scores and decides, but the money movement and the chargeback liability remain with your acquirer, and you still carry the integration work to wire decisions into your checkout. Third, the model needs your data. Businesses with thin transaction volume will not get the tuned performance the enterprise case studies show, and G2's implementation and ROI timelines imply a real tuning period before value lands. Fourth, there is no free-form BI: the analytics are fraud-oriented, and teams wanting warehouse-grade reporting export to their own stack. Finally, the AI agent suite is new enough that I would treat its capabilities as evolving — the core platform is mature, the agentic layer is 2026-era product.

Deployment practice is where I would spend my evaluation time, because a fraud platform is mostly an operations project. The teams I have seen get value from Sift are the ones that treat the first ninety days as model training: route everything possible through the platform, feed decisions back as confirmed fraud or confirmed good, and resist the temptation to hard-code declines before the score distribution is understood. G2's reported eleven-month ROI average makes sense in that light — the model compounds. The teams that fail are the ones that connect the Score API, set a threshold from a vendor sample, and never look at the queue again.

On the ecosystem and reliability side, the signals are decent. Sift publishes its developer documentation openly, runs a practitioner community of 1,500+ members, produces the FIBR benchmarking resource and Digital Trust Index reports, and has enough review volume on both G2 and Gartner Peer Insights to compare against Signifyd, Riskified, and Forter with real user data. The Fall 2025 release cadence — investigation tooling, identity trust improvements — shows continued investment rather than a product coasting on its install base. My caveats: the G2 category-leader claims come from Sift's own press releases, and press releases quote the rankings that flatter; and the AI agent suite is new enough that I would evaluate it as a roadmap item rather than a buying reason.

Cost modelling deserves a paragraph. With pricing built on check volume under a custom annual contract, the budget question is how many events per month you route through scoring — and the trap is scoring everything when only some flows carry fraud risk. Teams that segment event traffic, sending high-risk flows through full decisioning and lighter flows through cheaper checks, keep the contract honest. Ask for pricing per event type in the negotiation; the answer tells you how flexible the commercial model actually is.

Verdict

Sift is a serious, well-instrumented fraud platform with a corpus advantage in exactly the verticals it targets, and the investigation workflow is built for teams, not demos. I rate it 4.0 on the rubric — strong core functionality depth and a mature analyst experience, offset by unpublished pricing, an enterprise sales motion that locks out small volumes, and an ROI timeline that assumes you are processing enough fraud-bearing traffic to pay for it. If you run a marketplace, gaming platform, delivery business, or high-volume store with real fraud ops, Sift deserves a demo alongside the incumbent. If you process a few hundred orders a week, a processor-native tool like Stripe Radar is the more sensible spend.

Sources

• https://sift.com/platform/
• https://sift.com/platform/ai/
• https://developers.sift.com/
• https://www.vendr.com/marketplace/sift-science
• https://www.erpresearch.com/erp-add-ons/fraud-risk/sift
• https://www.g2.com/products/sift-s-ai-powered-fraud-decisioning-platform-sift/pricing
• https://www.gartner.com/reviews/market/online-fraud-detection/vendor/sift
• https://www.globenewswire.com/news-release/2025/08/20/3136587/0/en/sift-s-fall-25-release-advanced-fraud-investigation-tooling-to-simplify-fraud-management-strengthen-identity-trust.html
• https://www.globenewswire.com/news-release/2025/09/18/3152632/0/en/sift-maintains-1-position-across-all-fraud-prevention-categories-in-g2-s-fall-2025-reports.html
• https://businessmodelcanvastemplate.com/blogs/brief-history/sift-brief-history
• https://sift.com/resources/index-reports/