Government agencies worldwide face growing pressure to tighten controls as fraud cases climb across welfare, tax and procurement systems. In the United States alone, federal agencies reported over 60 billion dollars in improper payments in 2023, a figure that has risen for five consecutive years. European Union auditors recently flagged similar concerns, citing weaknesses in member states’ ability to spot irregularities before funds are lost. The challenge is no longer theoretical—it is a daily operational burden that demands faster, sharper tools than manual reviews can provide.
Agencies are now testing AI models trained on historical fraud patterns to flag suspicious transactions in real time. The Internal Revenue Service in the U.S. has deployed machine learning systems that analyze tax returns for anomalies linked to identity theft and refund fraud. Norway’s tax authority, Skatteetaten, uses similar software to cross-check welfare claims against bank records and property data, cutting false payments by 18 percent in two years. These systems do not replace human investigators but shift their focus toward the most complex cases.
The technical hurdles remain steep. Agencies must integrate AI with legacy databases that were never designed for such workloads. Data privacy rules in the EU and Norway require strict controls on how personal information is processed and stored. Security officials also worry about adversarial attacks—fraudsters who deliberately feed false data to mislead detection systems. Norway’s digitalization directorate recently published guidelines requiring agencies to encrypt AI inputs and allow external audits of algorithmic decisions.
Cost is another barrier. Developing in-house AI teams demands specialized skills and computing power that many agencies lack. Some turn to commercial platforms like Databricks, which offer pre-built fraud detection modules. The Norwegian Labour and Welfare Administration signed a contract in 2024 to use such a system for its child benefit program, aiming to cut processing time from three weeks to under 48 hours. The shift is part of a broader push to modernize public services while keeping fraud losses in check.
Source: databricks.com