Databricks Launches Real-Time Fraud Detection Tool on One Platform

Card fraud is not a slow-moving problem. A stolen card number can trigger dozens of purchases in minutes, and once money moves, getting it back is far harder than stopping the transaction in the first place. Global card fraud losses are projected to have hit $37 billion in 2024, up 19% from the prior year. In the UK, remote purchase fraud cases alone jumped 22% last year. The FTC reported that US consumers lost more than $12.5 billion to fraud in 2024, a 25% rise year over year. For any organisation that touches payments, whether that is a bank, an insurer, a retailer, or a travel company processing bookings, the pressure to catch fraud before it settles is growing fast. The technical bottleneck has always been speed. Most fraud models work well. The problem is running them quickly enough. Traditional data systems process transactions in batches: collect a group of transactions, run them through a model, flag the bad ones. By the time that cycle completes, the fraudster has already moved on. Catching fraud before a transaction clears means scoring it in under a second, sometimes in under 300 milliseconds. Until recently, achieving that speed required organisations to run two separate technology stacks side by side: one for standard data processing and analytics, and a specialist real-time engine for low-latency scoring. The result was duplicated infrastructure, split governance, and engineering teams spending more time keeping systems in sync than actually improving fraud detection. Databricks is now offering a direct answer to that problem. The company has published a ready-to-deploy fraud detection reference system built on two products it launched in the past year. The first is Real-Time Mode, an upgrade to its existing Spark data processing engine. Spark is already widely used for data pipelines and machine learning. Real-Time Mode, which entered public preview in August 2025 and reached general availability in early 2026, lets that same engine process events in milliseconds instead of batches. Coinbase, for instance, is using it to run risk checks on blockchain transactions with sub-100ms response times and has cut end-to-end latency by over 80%. The second product is Lakebase, a managed database that Databricks launched in June 2025, built on technology from its $1 billion acquisition of the database company Neon. Lakebase acts as a fast-access store for the data a fraud model needs during scoring: account history, card behaviour, location patterns. It is designed to respond in under 10 milliseconds and handle more than 10,000 simultaneous queries, which matters when payment volumes spike during a busy weekend or a major retail event. The reference system Databricks has released shows how the two products work together end-to-end. Transactions come in from a payment stream. Each one is scored against a fraud model within 300 milliseconds. The decision, approved, flagged, or blocked, is logged and immediately visible to fraud analysts through a live monitoring dashboard. Internal testing showed median latency under 40 milliseconds and worst-case latency between 215 and 392 milliseconds, which is within the window required before a transaction settles. For businesses outside financial services, this matters too. Insurers processing claims, retailers running loyalty programmes, travel companies handling booking payments, and logistics firms dealing with supplier invoicing all face some version of the same fraud exposure. The pattern Databricks is demonstrating, score every event in real time, keep all the data and the model in one governed place, log every decision for audit, applies across those industries. The broader shift worth watching is what Databricks is doing strategically. By making Real-Time Mode part of the same engine customers already use for batch analytics and machine learning, it removes one of the main reasons organisations kept separate specialist streaming systems. And by adding Lakebase, a transactional database, directly into that same platform, it is trying to eliminate the data movement problem that has historically caused delays between where data lives and where decisions need to be made. Adoption since launch has been fast. Since Lakebase entered public preview, thousands of companies have started running production workloads on it, growing at more than twice the rate of Databricks' earlier data warehousing product. That rate of uptake suggests the market was genuinely waiting for something that reduced the number of systems to manage. For organisations evaluating whether to modernise fraud infrastructure, the practical question is not whether the technology works. Early production deployments suggest it does. The question is whether the cost and effort of migrating onto a consolidated platform outweighs maintaining existing setups. For teams already on Databricks, that answer is straightforward. For those not yet there, the reference system at least gives a concrete benchmark to test against their current performance.