Riskified

A dilemma solved in a fraction of a second, thousands of times a day

In an online transaction there is no physical card, no signature and no seller looking you in the eye. There is only data: a card number, an address, a device, a time. Based on these alone, the merchant has to decide right away, with no time to think.

The merchant faces two kinds of mistakes, and only one of them is obvious. Approving a fake order costs the goods, the money and the chargeback fee. Declining a legitimate customer costs the sale, and often the customer too, who will go to a competitor and not come back. The common intuition is that the danger lies on the fraud side, but in many categories a wrong decline is actually more expensive. A merchant who tightens screening to feel safe ends up blocking real customers, and does not see it in any report.

This is the problem Riskified was built around, and it also explains its size: in 2025, transactions worth about $155 billion passed through the system. A decision that looks trivial for a single order becomes, at that volume, a major business variable.

What sets it apart: who absorbs the mistake

Most vendors in the field provide a risk score and leave the decision to the merchant. Riskified does something else. It decides on its own whether to approve, and commits to refunding the merchant the full amount if an order it approved turns out to be fraud. This product is called Chargeback Guarantee, and it is the core of the company.

The difference is not technical but economic. When a vendor sells a score, accuracy is a marketing claim. When a vendor is willing to absorb the cost of a mistake itself, accuracy is a direct expense on its own books. That is a far more credible signal than any statement about success rates, because it is measured in money, not in a slide deck. The commitment also changes what the company has to manage: it must estimate in advance how much it will pay for future mistakes, just as an insurer estimates claims.

The pricing follows from this too. Riskified charges a percentage of the volume of transactions that pass through it, not a fixed subscription fee. This means its revenue rises and falls with the customer's sales, and the interests of the two are almost fully aligned.

Beyond approving orders: returns, accounts and shopping agents

Deciding on the order is only one layer. A product called Policy Protect deals with a different kind of problem: not a thief using a stolen card, but a real customer who systematically abuses the store's rules, with repeated refund requests, reuse of promotions, and claims that a delivery never arrived. This is an important distinction, because the person's identity is entirely real. Justt, a Tel Aviv company founded in 2020, works on another side of the same problem. It uses AI to draft, case by case, the arguments merchants submit when they dispute a chargeback.

Within the same product, a capability called Dynamic Returns was launched. It adjusts the refund and exchange options in real time to the customer's risk level, instead of applying one rule to everyone or holding back a refund until the item is checked at the warehouse.

The engine itself relies on behavior analysis, detection of proxy servers meant to hide a location, and "elastic linking": spotting repeated details that connect accounts that seem unrelated. The new frontier concerns the question of who is actually doing the buying. Many merchants now run their own conversational shopping assistants, and in 2026 the company expanded a product line called AI Agent Intelligence, meant to protect these touchpoints.

The 2021 IPO: the exact numbers

Riskified began trading on the New York Stock Exchange on 29 July 2021, under the ticker RSKD. A day earlier, 17.5 million shares were priced at $21 per share, above the range shown earlier of $18 to $20. Of these, 17.3 million were offered by the company, which raised $363 million gross, with a possible additional $55.1 million if the underwriters exercised their option. The offering was led by Goldman Sachs, J.P. Morgan and Credit Suisse.

The IPO price implied a valuation of about $3.3 billion. It was the tenth IPO of an Israeli tech company on Wall Street that year, not counting mergers with special purpose acquisition companies. After the IPO, CEO Eido Gal held shares worth about $287 million, and co-founder and CTO Assaf Feldman held shares worth about $283 million.

Among the large shareholders at the time were Genesis, General Atlantic, Qumra, Pitango, Fidelity and Entrée Capital. Before this came a long series of funding rounds, including a $165 million round in November 2019 led by General Atlantic.

2025: the first profitable quarter since founding

For 2025 the company reported revenue of $344.6 million, up 5% from 2024, on transaction volume of $155.1 billion, up 10%. The gap between the two growth rates is not a mistake: when revenue is a percentage of volume, changes in the customer mix and in the fee level change the ratio between them.

The fourth quarter was a turning point. Revenue reached $99.3 million, the highest ever, gross margin rose from 52% to 57%, and the company recorded its first profitable quarter under U.S. accounting rules (GAAP), with net income of $5.8 million. Adjusted EBITDA for that quarter came to $17.7 million, more than in all of 2024. This is earnings before interest, taxes, depreciation and amortization, so it is not the same as operating profit. For the full year the company still recorded a loss of $27.6 million.

At the end of 2025 the company held about $297.6 million in cash and investments, with no debt, and an authorization to buy back shares worth about $85 million. Most of it depends on completing regulatory procedures in Israel; only about $10 million of it is unconditional. Its guidance for 2026 at the time was revenue of $372 million to $384 million. In August 2026, after a second quarter in which revenue grew 22% to $98.7 million, the company raised its full-year guidance to $400 million to $410 million.

What the company teaches about the AI wave before this one

Riskified and the Israeli company Forter together defined an entire niche in payments: automatic, real-time decisions on transactions. Both grew from the same local background of expertise in fraud detection. Both also did so many years before the term "generative AI" entered public conversation.

This point is worth holding on to. Much of the economic value that machine learning has actually created comes from classification problems that look boring, such as fraud or not, unusual or routine, and not from producing text or images. These problems are simply invisible to the end user, so they get almost no coverage. It is also a field where the model is never "done": it faces a human adversary who keeps changing methods, so it needs constant retraining.

The practical takeaway for the reader is a simple test. When any vendor offers a prediction capability, ask what happens to the vendor if it is wrong. An answer that includes a real cost for the vendor says more than any accuracy figure shown in a slide deck.