The Future of FWA: Agents, GenAI, and Reasoning

The current state of the art: reasoning models and GenAI

As we observe a plateauing (and diminishing) of LLM capabilities MFCUs, SIUs, and program integrity offices are pondering and proving out the role of agents and GenAI in their workflows. Most PI offices continue getting lots of mileage out of rules-based and predictive models for most of their needs. The current administration has aggressively encouraged the use of AI across government agencies. The latest state-of-the-art AI technologies have compelling uses in the future of FWA and program integrity.

Agentic AI Monitoring

Program integrity professionals have long familiarity with machine learning algorithms through the use of predictive and later, prescriptive analytics. The last few years have seen the rise of LLMs and their capture of the mindshare of the public. Agentic AI technologies are born from combining RPA technologies with AI/ML and large language models. Agentic AIs are goal-oriented and execute workflows autonomously. Program integrity professionals can define goals originating from existing analytics and put agents to work. Some agents can accept tasks like RPA. Newer agent AIs can leverage reasoning models and carry out tasks using clever prompting.

Agents for Key Risk Indicators

Many PI professionals I speak with are constantly asking for smarter, more efficient, and more accurate pre-payment authorization systems. The constant deluge of data, and scaling issues deriving from rules-based algorithms turn pre-payment authorization into a challenging problem making pay-and-chase a valid solution. Agentic AIs leveraging machine learning can more easily scale in pre-pay systems. Since Agentic AIs are goal-oriented, PI agencies can establish goals based upon local, regional, or state-wide conditions, predictive analytics, or value-based risk indicators.

Agent Observers

I’m often asked, “if we notice a trend in whatever, can we do xyz?” In many cases systems can trigger actions on spikes and inflection points in behavior. They’re easier to detect and if-else logic is easier to program. Some predictive analytics, the right modeling, and clever automation can make the smartest spike detection systems, like RIViR®, even smarter. Agentic AIs are presumed to always be on, and a possible goal for the agent could simply be, notice a trend in either direction, on one or more data points that start automation.

The PI Professional’s Role

Once upon a time customers would tell me, “I want our team to make the decisions.” Today, “can we automate clear cut decisions?”

The industry has done a lot to reduce hallucinations and build in safeguards. The latest reasoning models and agentic AI have the same thing in common. These systems are trained on already existing data and events. For every scammer busted, there are new ones popping with more sophisticated scams. In order to combat people-driven fraud, PI people will always be needed to make the tough decisions and train the next generation of agents.