How Artificial Intelligence is Reshaping Fraud Detection and Prevention Across the Public Sector

Heather Adams, Global Head of Risk and Compliance Consulting at Accenture, took to the stage at Counter Fraud 2026 to set out how artificial intelligence is reshaping fraud detection and prevention across the public sector. Here is a summary of her key points.

🎵 Prefer to listen? Hear Heather's full presentation here >>  🎵

The core opportunity

Heather opened with a direct statement: AI will transform government operations. Not in the distant future; now. She identified four practical benefits already being realised: stronger fraud defences at scale, reduced manual workload, improved efficiency, and better experiences for both staff and citizens.

From machine learning to agentic AI

She walked through the evolution of AI tools in fraud work. Machine learning has been in use for some years - it is good at scanning large structured datasets and identifying anomalies in claims, contracts and transaction records. For example, the US Treasury used it in 2024 to prevent and recover over $4 billion in improper payments.

Natural language processing and generative AI extend detection into unstructured data, such as text, documents and images. And now, agentic AI brings it together: automated agents that coordinate multiple steps, call on both machine learning and generative AI, manage processes, and work alongside humans to reach an outcome.

Eight opportunities in tax fraud

 Heather outlined eight specific areas where AI can be applied in taxation fraud work:

• Early warning systems - spotting known patterns and flagging high-risk cases before fraud occurs. HMRC already uses predictive modelling to risk-score taxpayers and apply differentiated treatment.
• Precision detection - learning from existing cases to target interventions more accurately. Poland's STIR model detects likely VAT fraud and automatically moves to block accounts.
• Case prioritisation - ordering casework by risk, not just volume, and building in fairness checks to avoid bias. The IRS uses case selection tools to balance risk and customer impact.
• Automation of investigation activity - ingesting documents, verifying data, and running full-population reviews rather than samples, surfacing only the highest-risk cases for human review.
• Incorporating new data sources - France uses AI to scan aerial photographs and compare them with declared property information, flagging discrepancies for investigation.
• Network analytics - mapping connections between individuals and entities to build a richer picture of potential fraud.
• Proactive fraud hunting - running what-if scenarios across datasets to identify emerging patterns before they become established.
• Improved customer service- chatbots and virtual assistants handling routine queries, guiding applicants through processes, and giving contact centre agents faster access to information.

Benefits in action

Across benefits,  Heather highlighted automated identity and eligibility checks, synthetic identity detection, biometric verification, real-time payment monitoring, and AI embedded into application processes to detect automated or fraudulent submissions. Medicaid in the US, she noted, has used payment card transaction monitoring to identify account takeover with greater accuracy; an approach she suggested could translate to Healthy Start cards or transport passes in the UK.

AI is also the threat

Heather was direct about the other side. AI is making fraud threats worse. She identified three specific risks:

• Hyper-realistic synthetic content - fake documents, bank statements, and identity evidence are increasingly difficult to distinguish from genuine ones.
• AI-enabled probing at scale - fraudsters using automated agents to submit large volumes of false applications, learning which ones succeed and adapting accordingly.
• Self-inflicted fraud - AI systems given poorly defined goals that lead to unintended behaviour. She cited a trial where an AI tasked with helping customers access benefits began falsifying information to improve application success rates.

Measuring ROI

On the question of return on investment, often cited as a barrier,  Heather acknowledged the difficulty but pushed back against inaction. She recommended focusing on effectiveness measures alongside financial ones: true positive rates, not just false positive rates; engagement and trust metrics; and honest assessment of whether fraud detection is actually improving. A survey by SAS found that 100% of civil servants using AI reported productivity improvements, with 57% citing efficiency as the primary benefit.

Her conclusion

Heather closed with a clear message: organisations do not have a choice about whether to engage with AI. The threats are already here; the tools are already available. The question is how to respond. Her prescription: build fraud prevention into AI-enabled processes now, with robust guardrails, real-time monitoring, kill switches, and a serious approach to identity authentication. Test, measure, and log everything. And accept that working alongside AI, not instead of it, is the new baseline.