North Carolina’s Raleigh-based State Auditor says AI helps investigators find suspicious transactions faster, according to WRAL’s new September 30 reporting. The strongest specific claim is substantial: Dave Boliek told WHQR in March that analyzing a municipality’s purchasing-card records had gone from two months to two days.
What backs that up? State law authorizes AI-assisted work, the office has recruited technical staff, and Boliek has described how software flags transactions for people to investigate. But the public sources checked October 1 did not provide the time logs, error measurements or full costs needed to verify that improvement. The evidence supports an account of how the work is done; the size of the benefit remains unverified.
What the state funded
Session Law 2025-89 authorized the Division of Accountability, Value, and Efficiency, known as DAVE, to employ people using AI and other tools to examine government spending, recipients, duplication and fiscal soundness. The law also appropriated $6 million in recurring funds for each of fiscal years 2025-26 and 2026-27 for up to 45 division positions.
The appropriation covers division staffing. It does not tell us how much the office actually spent on AI or how many authorized positions it filled. The agency’s 2025 Data Scientist job posting sought experience building fraud, anomaly and inefficiency models, while its data-science internship page describes machine-learning work and says experience integrating large language models is preferred. These are records of the skills the office sought. They do not establish a system’s performance.
What investigators say they do
In a May interview with PBS North Carolina, Boliek described an internal sandbox, a team of 18 data scientists and work on enclosed large language models intended to protect personally identifying information. Those details are his account of the operation as of that interview; we did not verify the staffing figure against payroll.
The March WHQR interview adds an important qualification to the two-month-to-two-day claim: Boliek said his team refined the tool to reduce errors and had people check its results. The reporting supplies no measured error rates or comparable time records, including the effort spent checking the software’s work. A quicker initial scan could be useful, but evaluating its benefit requires counting that follow-up.
In WRAL’s September 30 report, updated October 1, Boliek again described AI flagging transactions followed by human forensic work. An office spokesperson said AI policies are continually updated through a subcommittee using other state and agency frameworks. The office has described controls; their actual text and validation results were not among the records obtained for this article.
One investigation provides a useful check on what the published methods reveal. The office’s November 2025 Elizabeth City report describes reviewing purchasing-card receipts, approvals and compliance with spending rules. It does not identify which steps used AI. That limits what the report can tell us about the tool’s contribution; it does not establish that AI was absent.
More reports do not measure AI’s contribution
WRAL counted more than 80 audits or reports this year. Our check of the official report index also found more than 80 entries dated 2026, spanning different kinds of reports. Neither count identifies the portion of the work performed with AI. Measuring a productivity gain would require comparable work and staff effort across periods.
The October 1 review covered the enacted law, the agency’s public recruitment pages, the first five pages of its public report index, the public budget-dashboard landing page and the cited PBS, WHQR and WRAL accounts. The report index does not identify which entries were AI-assisted, and the budget dashboard landing page did not expose AI-specific spending in the material inspected. That limited search cannot determine whether additional internal or unreviewed records exist.
Records that could change the answer
Four kinds of records would make the claim testable:
- AI policy text, subcommittee records and a model inventory could show the governance rules and systems actually covered.
- Validation plans, human-review dispositions and false-positive or false-negative results could test accuracy rather than simply show that screening occurred.
- Contracts, invoices, personnel allocations and actual division expenditures could identify the cost of implementation and human review.
- Report-level method disclosures and comparable workload logs could test whether AI changed production time or contributed to specific published work.