Every fraud decision is a bet placed with incomplete information. Approve an order and you risk letting a fraudster through, but decline it and you risk losing a real customer. Good fraud risk management comes down to making better bets with the information that you have at your disposal.
Fraud risk is a bet made with limited information
Raj Jain, Sift’s Chief Product Officer, spent about three years playing poker semi-professionally, and he sees the same discipline at work in fraud prevention. The best players didn’t win the most hands: they just made better decisions with what they knew and understood when to take a calculated risk, when to fold, and when to change strategy as the table shifted.
Fraud teams work under the same conditions. Consumers reported losing about $16 billion to fraud in 2025, setting a record and roughly 25% above the 2024 figure, according to the Federal Trade Commission. Those numbers describe outcomes after the fact. Your team has to make the call before the outcome is known, by using various signals to essentially make an educated guess. A large purchase, a login from a new device, and a shipping address the account has never used before are all fraud risk indicators, but none of them is proof.
The decision is also wider than simply approving or declining. You can approve, add verification, send the case to manual review, or decline. Each action has a different cost, and the right one depends on more than a fraud risk score. The same high-value order deserves a different response from a five-year customer than from an account created five minutes ago. A score is an input. The workflow picks the action, based on your risk tolerance, your customer experience goals, and your operational priorities.
Fraud risk management starts with the workflow, not the model
Much of the current conversation about AI in fraud centers on which model is fastest, smartest, or newest. Those qualities matter, but the leading general-purpose models are available to any company willing to pay for them, and new versions keep arriving. A vendor that leads with “we use the best model” is describing cards you should expect your competitors to hold soon. Raj’s advice for what to do with the cards: build the workflow, rent the intelligence, and own the governance.
Strong poker players also size their bets to what is at stake. They don’t push in the biggest bet on every hand, and a recent test from Fireworks AI and Harvey, a legal AI company, shows the same principle in AI systems. Across 100 complex legal tasks, one setup used a top frontier model, Claude Opus 4.7, for every step. It fully passed 14 tasks at a total cost of $954. A second setup used an open-weight model, GLM 5.1, as the worker and called Opus only when it needed a second opinion, an average of 0.83 times per task. It fully passed 18 tasks at a total cost of $368.
Fireworks reports a standard error of about 2.5 percentage points on the pass rate, so the quality gap sits within the margin of error. The cost gap does not. Using the most powerful model on every step did not win. Knowing when to call on it did. This was a legal benchmark, not a fraud one, so treat it as a design principle rather than a fraud result.
For a fraud team, the takeaway is practical. Route routine events through fast, automated decisions, and save deeper analysis and human judgment for the ambiguous, high-value cases where a wrong call costs the most.
Read fraud risk indicators across the whole table
A strong hand can turn weak depending on what is happening around it. Fraud works the same way. A single transaction or login rarely tells the whole story.
Take an account that changes its password, adds a new payment method, and places several high-value orders. Each action can look legitimate on its own, but together, they can point to account takeover. If similar behavior appears across dozens of different accounts and are all linked by shared devices or IP addresses, what looked like unrelated events may be one coordinated attack.
The data supports looking wider. FinTech Global reports that Federal Reserve Financial Services put U.S. account takeover losses at more than $15.6 billion in 2024, up from $12.7 billion in 2023. The Federal Reserve’s 2026 Risk Officer Report found that 23% of surveyed financial institutions had experienced account takeover activity, an increase of seven percentage points year over year. The same coverage notes that criminals increasingly operate through legitimate customer credentials, which makes account takeover detection a broader exercise than spotting a suspicious login.
Fraud risk indicators earn their value when viewed in combination. Signals worth connecting include:
- A new or unrecognized device, especially right after a credential change
- A new payment method or shipping address added shortly before a purchase
- Unusual order value or velocity compared with the account’s history
- Shared devices, IP addresses, or payment details across multiple accounts
That’s why fraud risk assessment has to move beyond scoring individual events. Teams need a view across the whole customer journey, a way to find connections that are invisible one transaction at a time, and the ability to act before losses escalate. A powerful model can help explain what it is seeing. The value still comes from the quality of the underlying signals, the context available to the system, and what your team can do with it.
Judge the decision, not just the outcome
Poker teaches a lesson that is hard to accept: a correct decision doesn’t guarantee a win. You can calculate the odds, make a smart bet, and still lose. You can make a poor one and get lucky. Serious players review their decisions, not just their results.
Fraud teams need the same habit. An approved order that becomes a chargeback does not automatically mean the decision was wrong. A decline does not prove fraud was prevented, because the customer may have been legitimate. If you only judge by the outcomes you can see, you’ll draw the wrong conclusions.
Reviewing a decision means knowing why it was made. What was known at the time? Which signals mattered? Which thresholds and workflows applied? With what you know now, should anything change? An AI-based system should not hand analysts a recommendation with no context. They need to know why an action was suggested, when to trust it, and when human judgment should take over.
Analysts also need to be able to fix what they find. Maybe a workflow is sending too many legitimate customers to manual review. Maybe new fraud patterns are slipping past your thresholds. Maybe extra verification is adding friction without reducing risk. In practice, this is what governance means: not only documenting decisions for compliance, but giving teams the visibility and control to understand, evaluate, and improve how decisions get made. While you can’t control the outcome of every hand, you can build a process that improves with every decision.
Questions to ask when evaluating AI for fraud risk
General-purpose models are the part you can rent. Fraud-specific machine learning trained on transaction and behavioral data, backed by network intelligence and expertise, is a different asset, and in Sift’s view a general-purpose model cannot replicate it on its own. The strongest systems combine both, using general-purpose AI to help analysts investigate, summarize patterns, and work faster.
Models will keep improving, and the best one today may not be the best one next year. Ask whether a vendor’s value depends on any single model. If it does, you’re exposed when that model changes or gets overtaken. If the value comes from the signals, workflow, and governance around the model, improvements can arrive without you needing to rebuild your rules, review queues, or reporting.
Four questions help separate the two:
1. Does it give my analysts the context to investigate suspicious activity, and connect signals across accounts and the full customer journey?
2. Does it help me apply the right level of friction for the risk?
3. Can I see why a decision was made and adjust the process as fraud changes?
4. Do my results depend on one specific model, or on the signals and workflow around it?
Today’s most advanced technology will become tomorrow’s baseline. The fundamentals of good decision-making change far more slowly. Over time, the teams that come out ahead are the ones that understand the odds, watch the whole table, and learn from every decision.
If your fraud team would benefit from investing in advanced fraud detection technology, then you should consider using Sift. Request a free demo today.
Frequently asked questions
What is fraud risk management?
Fraud risk management is the practice of identifying, assessing, and responding to the chance that a transaction, account, or action is fraudulent. It combines signals, a workflow that decides whether to approve, verify, review, or decline, and ongoing review of past decisions so the process improves.
What are common fraud risk indicators?
Common fraud risk indicators include a new or unrecognized device, a change to credentials, a newly added payment method or shipping address, unusual order value or velocity, and shared devices or IP addresses across accounts. No single indicator proves fraud. The combination, and the context around the customer, is what matters.
How do fraud teams decide whether to approve, verify, review, or decline?
The decision depends on the combined risk signals, the customer’s history, the value at stake, and the cost of added friction. A workflow applies those factors consistently, so a long-standing customer and a brand-new account placing the same order can receive different responses.






