Best AI Fraud Detection Tools 2026: Sift vs Forter vs Signifyd vs Riskified — Buy the Coverage Schedule, Not the Model

Most AI fraud detection tools are sold on detection accuracy. That is the wrong thing to compare them on.

LexisNexis published its 2026 True Cost of Fraud study for retail and ecommerce in North America, and the fraud multiplier crossed $5 for the first time in both markets — $5.13 in the US, $5.23 in Canada, up from roughly $2.40 a decade ago. The number that should worry you more is still the one quoted least: false declines are estimated at around $443 billion a year globally against roughly $48 billion in actual ecommerce fraud losses. Both figures come from merchant-funded research, so treat them as directional rather than gospel. The direction is not in dispute.

Here is the thesis. The four platforms buyers actually shortlist — Sift, Forter, Signifyd and Riskified — no longer compete on model quality. They compete on which slice of your loss they are willing to own financially. That is a contract question, not a data-science question, and it is the one buying committees skip.

The two business models behind AI fraud detection tools in 2026

There are exactly two commercial models in this category, and confusing them is the most expensive mistake in the stack.

Risk scoring. You buy scores and decisioning tools; you keep the financial risk. Sift is the purest example, along with Kount (now inside Equifax) and Stripe Radar. You get a risk score, a rule engine, a manual review queue, and analysts you still have to staff. Sift runs on a network of more than 1 trillion annual events and says it protects $325 billion in gross annual merchandise value across 34,000+ sites, with a median customer preventing $4.2 million in losses annually. None of that changes the fact that when a bad order gets approved, the loss is yours.

Liability transfer. You buy a promise: approve this order and we eat the chargeback. Forter, Riskified, Signifyd, ClearSale and NoFraud sell some version of it. Pricing follows the promise — a percentage of approved or protected GMV rather than a per-event fee, because the vendor is now underwriting you.

That distinction explains most of the price spread. Guarantee vendors cost more per transaction because they sell insurance. Scoring vendors cost less because they sell a model. Neither is wrong. Buying one while expecting the other is.

What the four platforms actually charge

Public list pricing barely exists here. Sift pricing is quote-only, with no tiers page. Vendr's transaction data across 47 tracked Sift deals puts the median annual contract at $144,600, ranging from $28,400 to $600,000, and mid-market buyers processing 100,000 to 500,000 events a month budgeting $40,000 to $100,000 a year.

The sift vs forter question therefore is not a rate comparison at all. Forter, Signifyd and Riskified all price as a percentage of GMV, which turns Forter vs Signifyd into a coverage question rather than a discount hunt. Vendr's mid-market ranges land at $60,000–$180,000 for Forter, $50,000–$150,000 for Signifyd and $70,000–$200,000 for Riskified, with Riskified typically 15–30% above Sift because the chargeback guarantee is bundled rather than negotiable. Riskified pricing is quoted as a percentage of approved GMV plus an annual floor, which is why the same platform can cost $70,000 for one merchant and $200,000 for another at similar volumes. Signifyd pricing and Forter pricing both tack on implementation fees of $2,000–$30,000 depending on integration depth.

Stripe radar pricing is the exception that proves the rule, because Stripe publishes it. Radar's machine learning fraud detection lists at $0.05 per screened transaction, free for accounts on standard payments pricing, with the legacy Fraud Teams tier at $0.07 and $0.02 for standard-pricing accounts. Stripe has been migrating that structure to four tiers — Lite (free with Payments), Standard, Plus and Pro — with reported US pay-as-you-go rates of $0.05, $0.07 and $0.09 per screened transaction. One caution: Stripe's Radar pricing page renders in local currency by region, so what you see from a European or Asian IP is not US pricing.

Now the part nobody puts in a vendor comparison. Multiply those per-transaction rates by a plausible volume and the arithmetic looks cheap: one third-party comparison puts Sift at $1,800–$9,000 a year for a merchant doing 15,000 orders a month, and the guarantee vendors at $9,000–$27,000. Signed contracts, per Vendr, land 10 to 30 times higher. Both numbers are real. The published rate is what you pay above the floor; the floor is a platform fee with an annual minimum. Budget from the contract range, not the rate card. The same logic applies to overages, which run 20–50% above base per-event rates, and to the 3–7% annual escalation clauses that are standard unless you strike them.

The coverage schedule is the real product

Riskified's own documentation states it plainly: the Chargeback Guarantee covers "all chargebacks related to unauthorized card usage or stolen financial information, but do not cover chargebacks for reasons such as 'item not received' or 'item not as described', as these are not related to fraud." Refund abuse, return abuse, promo abuse and reseller abuse sit in a separate product, Policy Protect.

Signifyd sells the same idea as a ladder. Guarantee Fraud Protection covers fraud or unauthorized reason codes. Chargeback Recovery Service adds managed representment on those same codes. CBR+INR adds item-not-received. Only Complete Chargeback Protection covers all card-not-present reason codes, including the non-fraud ones — and even that tier declines claims where goods were returned, where the shipment was rerouted, or where the merchant cannot prove fulfillment.

Then run the arithmetic against your actual dispute mix. Chargebacks911's 2026 field report, based on more than 250 merchants, found merchants attribute 43.8% of chargebacks to friendly fraud, while its own case data puts it at 86%. Juniper Research models a far lower 22% of chargebacks in 2026, rising to 28% by 2031, because it works from network data instead of merchant self-report. The spread is definitional, not sloppy research. Use whichever number your acquirer uses.

Either way, the point holds: the fastest-growing slice of disputes is the slice a base-tier guarantee excludes. Anyone shopping for chargeback prevention software to stop friendly fraud needs to price the upgrade, not the entry tier.

The number nobody guarantees: false decline rate

Every vendor shows a detection curve. Almost none will contractually guarantee a false decline rate, because rejecting legitimate orders costs the vendor nothing and costs you everything.

Riskified argues that a one-percentage-point improvement in false decline rate is worth more than a five-point improvement in fraud detection for most merchants — a candid framing from a company selling a guarantee. The macro version: false declines estimated at $443 billion a year against $48 billion in fraud losses. LexisNexis found 56% of US retailers and 54% of US ecommerce merchants reported increased customer churn tied to anti-fraud measures. Customers who get wrongly blocked rarely complain; they just do not return.

This is where approval-lift claims deserve stress-testing. Signifyd says its merchants approve 5–9% more orders than they otherwise would. Forter claims false declines down 46% and authorization rates up 1–3% for switchers. Sift claims 55% lower block rates. All vendor-reported. Ask for your category, your ticket sizes, in writing.

Why 2026 changed the math

Agents started buying. Ravelin's Agentic Commerce and Fraud Report 2026, covering 1,504 fraud and payments professionals across 10 countries at merchants above $50 million in revenue, found 44% already integrating agentic commerce protocols and another 32% planning to within six months. Only 29% felt very prepared for the fraud and security implications. Ravelin also found 53% of merchants now trust AI agents more than human shoppers. That trust has a complication: risk engines tuned to human behavioral signals generate more friction on legitimate agent traffic, and liability when an agent-driven purchase is disputed remains unresolved. UK research from The Payments Association found 58% of online retailers believed AI-initiated transactions had already reached their platforms, while only 41% felt confident in existing liability frameworks. A widely reported Riskified briefing from April 2026 put agent-attributed chargebacks up 34% in Q1 2026 — mostly consumers who authorized an agent, received the goods, then disputed the charge on the grounds that they never personally authorized that specific transaction. That is friendly fraud wearing a new hat, and it sits even further outside a fraud-code-only guarantee.

Stripe repriced screens on October 21, 2026. Stripe's own support notice states that from that date Radar pricing applies to all supported payment methods — wallets, bank debits, BNPL, stablecoins — and to every transaction in a subscription, not just the first payment. Per-transaction rates did not change; the billable surface did. A subscription business screening 10,000 renewals a month at $0.05 is looking at roughly $6,000 a year it was not paying before, and it can opt out of the newly covered payment methods before the effective date. Merchants migrated to Radar Standard are also reported to be on a trial ending January 22, 2027, after which per-screen billing begins unless they move to Lite — giving up custom rules, 0–99 risk scores, adaptive 3D Secure and manual review queues. If you run recurring revenue, check this line item against your AI subscription billing tools recovery stack, because dunning and fraud screening now bill on the same transaction volume.

How to run the numbers before you sign

Model total loss, not fraud loss. Add fraud losses, chargeback fees, the loaded cost of manual review hours, estimated false decline revenue, and the lifetime value of customers you insulted. That last line is usually the largest and always the least measured. Operators on an AI ecommerce platform can pull order-level data for this in an afternoon. Most never do.

Pull last year's reason codes. If more than a third are non-fraud codes, a fraud-only guarantee is buying the smaller half of your problem.

Price the upgrade, not the platform. Ask each vendor for the all-codes number explicitly. Then compare the premium against the disputes it would reimburse. At a 0.1% chargeback rate the guarantee rarely pays; at 0.6% on high-ticket electronics it usually does.

Screenshot the exclusion list. Claim windows, proof-of-shipment rules, signature thresholds above $1,000, pickup carve-outs, returned-merchandise exclusions. One Shopify brand scaled approved revenue 4x in twelve months on Riskified while holding a 0.06% chargeback rate — that came from calibrated thresholds, not from the guarantee clause.

Decide where support sits. Refund and return abuse now surfaces in the support queue more often than the risk queue, which is why Riskified integrated its identity risk intelligence into Zendesk in September 2026. If your agents approve appeasements, your AI customer service stack is part of the fraud perimeter whether or not procurement treats it that way.

Frequently Asked Questions

Is Sift or Forter cheaper for a mid-market store?

They are different purchases, so the comparison collapses. Sift is a subscription plus event volume with no guarantee; Forter prices a percentage of GMV and reimburses covered chargebacks. Vendr's contract data puts mid-market Sift at $40,000–$100,000 and Forter at $60,000–$180,000. Sift wins on cost if your team owns decisions and your fraud rate is comfortable. Forter only wins the economics if reimbursements exceed the premium, which takes a real dispute history to test.

Does a chargeback guarantee cover friendly fraud?

Usually not on the tier you are quoted. Riskified's guarantee covers unauthorized card usage and stolen financial information, and explicitly excludes item-not-received and not-as-described disputes; Policy Protect handles refund and return abuse separately. Signifyd's Complete Chargeback Protection covers all card-not-present reason codes, but it is a higher tier and still declines claims where goods were returned or shipment cannot be proven.

How much does Stripe Radar cost compared with Sift?

Two orders of magnitude apart, and not the same product. Stripe Radar lists at $0.05 per screened transaction for the machine learning tier in the US, free on standard payments pricing. Sift's median signed contract is $144,600 a year. Radar screens transactions on Stripe; Sift scores identity, account, content and payment risk across your own stack and leaves the loss with you. If you are Stripe-only with an unremarkable fraud rate, Radar is close to free money. If you run multiple processors or a marketplace, Radar does not cover the whole surface.

What false decline rate should I accept?

No public benchmark is worth copying, and vendors deliberately avoid guaranteeing one. Test it revenue-weighted: compare the conversion rate of orders routed to manual review against orders that auto-approve, then decide whether the threshold is converting fraud prevention into customer acquisition damage. LexisNexis found 56% of US retailers reporting churn tied to anti-fraud controls, which suggests many thresholds are set too tight.

Do I need AI fraud detection tools at all if I use Stripe?

Depends on volume and merchandise. Under roughly $1 million in annual GMV with a sub-0.3% dispute rate, Stripe Radar plus disciplined descriptor hygiene is usually enough, and statement-descriptor mismatches remain one of the largest avoidable sources of disputes. Above that, or in resale-prone verticals like electronics, sneakers and gift cards, the approval lift from a dedicated platform tends to pay for itself.

Bottom line

AI fraud detection tools are a coverage purchase in 2026, not an accuracy purchase. The model is table stakes; every serious vendor has one. What differs is the coverage schedule — which dispute codes get reimbursed, what proof is required, and whether the fastest-growing category of disputes sits inside or outside the contract. Rank your options by what they will pay you, not by what they claim to catch. Budget from signed-contract ranges rather than rate cards, read the exclusion list like an insurance policy, and re-check your Stripe bill after October 21. The switch that saves money is rarely the model swap. It is usually the tier change.