Miles Jamison 
- Tricentis’ Allan Troup said duplicate refunds and reused Social Security numbers are potential AI use cases
- Michigan and Vermont are exploring AI applications for tax fraud detection
- Testing should cover fraud scenarios, calculations and data flows during peak filing periods
Allan Troup, senior director of public sector SLED at Tricentis, said state revenue agencies exploring artificial intelligence should establish clear objectives and testing requirements before putting the technology into broader operation.
In a commentary published in Route Fifty, Troup said agencies are contending with tight budgets, staffing shortages and more complex fraud patterns as they manage revenue operations.
How Are States Exploring AI?
Troup cited tax-fraud detection as an emerging use for AI, noting that Michigan has proposed funding AI-driven analytics within its Department of the Treasury, while Vermont’s Agency of Administration has identified AI tools for detecting tax-fraud risks. He also identified duplicate refund claims, reused Social Security numbers and claims involving deceased taxpayers as potential areas for AI detection.
According to Troup, defining the operational objective first can give agencies measurable criteria for evaluating an AI pilot and provide them with information to use when responding to lawmakers and auditors.
How Should Agencies Validate AI Systems?
Troup said agencies need to account for changes that can affect AI performance after a system is introduced. Models, tax engines and related applications can be updated over time, requiring agencies to repeat testing rather than treating validation as a one-time exercise.
He recommended testing fraud scenarios, calculations and data flows, along with system performance during periods of heavy transaction volume. Troup noted that agencies can add AI-enabled functions around existing tax administration systems while evaluating new uses.
What Records Should Agencies Maintain?
The Tricentis executive said revenue agencies should be prepared to explain how AI systems use data and produce risk determinations. He said pilots should include audit trails of automated actions, documented decision rules and logs of what goes into and comes out of each model.
Troup also called for human review procedures for flagged cases, particularly when AI is involved in enforcement or financial determinations. He said agencies should monitor AI behavior over time and establish clear parameters for its use.
