I spent about three years playing poker semi-professionally. One of the most important lessons I took from it had very little to do with the cards I was dealt.
The best players weren’t the ones who won the most hands. They were the ones who consistently made better decisions with the information they had. They knew when to take a calculated risk, when to walk away, and when to change strategy based on what was happening around them.
I see the same thing in fraud prevention. Every fraud decision is a bet made with limited information. Approve a transaction and you risk letting fraud through. Decline it and you risk losing a real customer. Add friction and you might stop an attacker, or frustrate someone who was ready to buy.
You don’t get to pick your cards
Right now, much of the conversation about AI in fraud is about which model is fastest, smartest, or newest. Those things matter, but they’re the deck, and everyone plays from the same deck. The leading general-purpose models are available to any company willing to pay for them, and new versions arrive every few months. A vendor that leads with “we use the best model” is describing cards you can expect everyone else to hold soon.
What separates winners is what they do with the cards. My version of the advice: build the workflow, rent the intelligence, own the governance. Poker taught me what each one looks like in practice.
1. Play the odds, not just the cards
In poker, you rarely have complete information. You decide based on probabilities, the bets already made, the behavior you’ve observed, and the cost of being wrong.
Fraud teams work under the same uncertainty. A customer might be making an unusually large purchase, logging in from a new device, or shipping to an unfamiliar address. Each signal tells you something. None of them tells you everything. The real question is what to do: approve, add verification, send to manual review, or decline. Those actions have different costs, and the right choice depends on more than a risk score. The same high-value order deserves a different response from a five-year customer than from an account created five minutes ago.
AI can help by analyzing more signals and surfacing patterns an analyst would take far longer to piece together. But intelligence alone doesn’t pick the action. A workflow does, based on your risk tolerance, your customer experience goals, and your operational priorities.
Strong players also size their bets to what’s at stake. They don’t put in the biggest bet on every hand. A recent test by Fireworks AI and Harvey, a legal AI company, shows the same principle. On 100 complex legal tasks, one setup used a top frontier model, Claude Opus 4.7, for every step. Another used a cheaper open-source model (GLM 5.1) and called in Opus only when it needed help. The Opus-only setup fully passed 14 tasks at an estimated $954. The combined setup passed 18 at an estimated $368. The difference in passes is within the margin of error, but the cost difference is not. Using the most powerful model on every step didn’t win. Knowing when to call on it did.
2. Read the whole table, not just your hand
A strong hand can turn weak depending on what’s happening at the table: how other players behave, how betting changes, what new information appears as the hand unfolds.
A single transaction or login rarely tells the whole story in fraud either. Take an account that changes its password, adds a new payment method, and makes several high-value purchases. Each action might look legitimate alone. Together, they can point to account takeover. And if similar behavior shows up across dozens of accounts, linked by shared devices or IP addresses, what looked like unrelated events may be one coordinated attack.
That’s why the next phase of AI in fraud prevention has to go beyond scoring individual events. Teams need a view across the whole customer journey, the ability to find connections that aren’t visible one transaction at a time, and a way to act before losses escalate. A powerful model can help explain what it’s seeing. But the value comes from the quality of the underlying fraud signals, the context available to the system, and what your team can do with it.
In poker, reading the table helps you see beyond your cards. In fraud prevention, connecting signals helps you see beyond the transaction.
3. A good decision can still lose. Learn from it anyway.
Poker teaches something that’s hard to accept: making the right decision doesn’t guarantee you’ll win the hand. You can calculate the odds correctly, make a smart bet, and lose. You can make a terrible one and get lucky. That’s why serious players review their decisions, not just their results.
The same applies to fraud. An approved order that becomes a chargeback doesn’t automatically mean the decision was wrong. A decline doesn’t prove fraud was prevented, because the customer may have been legitimate. If you judge only by the outcomes you can see, you’ll draw the wrong conclusions.
Instead, you need to know why each decision was made. What was known at the time? Which signals mattered? Which thresholds and workflows applied? Knowing what we know now, should anything change? An AI-powered system shouldn’t be a black box that hands you a recommendation with no context. Analysts need to know why an action was suggested, when to trust it, and when human judgment should take over.
They also need to be able to fix things. 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. This is what governance means to me in practice: not just documenting decisions for compliance, but giving teams the visibility and control to understand, evaluate, and improve how decisions get made.
You can’t control the outcome of every hand. You can build a process that gets better with every decision.
Know what you’re betting on
Not all AI is interchangeable. General-purpose models are the part you can rent. Fraud-specific machine learning trained on transaction and behavioral data, and the network intelligence and expertise behind it, is different. A general-purpose model can’t 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. So ask whether a vendor’s value depends on any one of them. 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 rebuilding your rules, review queues, or reporting.
If you’re evaluating AI-powered fraud technology, look past the model and ask:
- Does it give my analysts the context to investigate suspicious activity, and connect signals across accounts and the full customer journey?
- Does it help me apply the right level of friction for the risk?
- Can I see why a decision was made and adjust the process as fraud changes?
- 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 won’t change nearly as fast. In poker, the cards matter, and so does luck. But over time, the players who come out ahead are the ones who understand the odds, pay attention to the table, and learn from their decisions. I believe the same is true in fraud prevention.
The winning strategy isn’t having the best AI. It’s knowing how to use it.
