Moving government from fraud detection to prevention through data and AI

For a long time, the government's default response to fraud has followed the same script, where the benefits are paid and the investigation follows.

"The benefit was delivered. You now identify that there might be concerns there, and how do you go chase that money?" said Keith Swanson, Director of Fraud and Compliance for Asia Pacific and Japan at SAS, describing this as a legacy "pay and chase approach".

Closing that gap, by identifying risks before public funds leave, was the focus of the From Detection to Prevention: AI for Government Fraud, Waste and Abuse webinar on September 16, co-hosted by GovInsider and SAS.

Swanson added that historically applied methods such as pay and chase often caught around one per cent of leakage, consisting of fraud, error, waste and abuse, in government programmes, while more in-depth studies put the real leakage much higher.

Previous reports indicated leakage was six to 10 per cent in an Australian benefits programme, and 11 to 15 per cent across US benefit schemes during the Covid-19 pandemic.

Following Swanson's keynote was two focused fireside chat sessions.

The first discussed how the use of artificial intelligence (AI) has help shift government from reactive to predictive risk management, featuring Swanson, Jai Singh, Assistant Director at India's Employees' State Insurance Corporation (ESIC) under the Ministry of Labour and Employment, alongside Rashul Hussain, Project Manager in the finance department of the Government of Assam.

The second turned to governance, with Ensley Tan, Industry Lead for Public Sector in Asia-Pacific, SAS, and Jean Perkins, Senior Programme Director for National Reforms at the Northern Territory Police Force in Australia.

Below are three takeaways from the conversations.

1. Turning AI insights into fraud prevention

AI is moving beyond identifying suspicious transactions to helping government agencies assess risks, investigate potential fraud and intervene before public funds are lost.

Swanson highlighted how machine learning and network analysis can uncover suspicious patterns and hidden relationships, while granular, "micro-level" signals, such as the device someone logs in from, can help agencies identify risks and prioritise investigations.

Decisioning and agentic AI can take this further by turning risk insights into action.

In a potential government application, when a system flags a high-risk benefits application, an AI agent could gather relevant information, engage with applicants and initiate a review, while referring cases requiring further judgement to public officers.

However, the effectiveness of these technologies depends on the quality of the underlying data and the trust within the application of AI itself and in that.

Swanson highlighted the importance of robust governance in instilling internal, and external, trust in the application of AI.

"The journey doesn't start with AI," Hussain said. "The journey starts with building data." He added that AI is less a starting point than an accelerant once a robust data foundation exists.

Singh highlighted the challenges at ESIC, where manual entry from hospitals not yet online, duplicate records and stale data all feed what he called "garbage in and garbage out."

Perkins added that progress does not necessarily require collecting more data or establishing new data-sharing arrangements.

"It's not necessarily about new data or new data sharing arrangements," Perkins said. "It's about strengthening what we already have and understanding what data is already available to be used." 

2. Trust in AI is a governance problem, not just a technical one

Tan broke "trusted AI" into three components: authority ("What is that AI allowed to do? What data can it access?"), ownership ("Who's accountable for the outcome, if a case is wrongly escalated?), and assurance ("How do we know that those rules are actually being followed and can be audited?").

Perkins reframed what trust means in practice, which was "It's not really whether or not people are trusting the technology... It's about whether they trust the way we're using it."

Perkins also noted "a risk of over-governing... that's almost as bad as having no governance around it because you don't reap the benefits [of AI]."

She advised matching the AI tool to an agency's actual maturity, since "very few organisations will actually require the most advanced AI tools."

Tan said “AI must be explainable” has become a familiar refrain, but pointed to a more subtle risk, that AI recommendations can create a new accountability burden, since officers must now justify why they followed, or departed from, an AI-generated recommendation.

"If the explanation is too weak, the officer will be uncomfortable overriding it… If the explanation feels too complex, the officer might just defer to it and not think about it anymore either."

The issue is not simply whether the AI can explain itself, but whether the public officer can explain and defend the decision made with AI.

3. Preparing the workforce for AI-driven fraud prevention

Technology alone cannot transform fraud prevention.

As AI plays an increasingly important role in risk identification and decision support, public officers need the skills and confidence to interpret AI-generated insights and apply them in practice.

In India's diverse population, Hussain said that "something that AI flags as an anomaly might not be an actual anomaly, and it might represent a very small section of population."

It was a reminder that statistical outliers and fraud are not the same thing.

Officers need to understand the context behind AI-generated alerts rather than treating every anomaly as evidence of fraud.

Tan proposed keeping officers meaningfully in the loop with systems that show the evidence behind a risk score, the alternatives considered, and counterfactuals, which he described as contrary evidence that would reasonably support a different conclusion.

Perkins tied this to workforce investment, cautioning that "it requires some considered planning of your workforce and uplift in the workforce capability to work in that new environment."

As AI becomes increasingly integrated into government fraud prevention, public officers will need to move beyond routine reviews towards interpreting AI insights, assessing complex situations and making informed decisions. 

Swanson highlighted the importance of human in the loop in governing AI use, and also the importance of human intelligence and intuition, combined with analytics and AI, often yielded the best outcomes from fraud, error, waste and abuse detection and prevention regimes. 

Moving from detection to prevention requires more than advanced AI tools. It calls for stronger data foundations, effective governance and a workforce equipped to turn AI insights into timely, informed action.