WAN-IFRA: USA Today on measuring AI before trusting it in the newsroom

What the article is about

This June 2026 piece from WAN-IFRA profiles USA Today’s approach to deploying AI for a specific, high-stakes journalism workflow: drafting public records requests. The account covers what was built, what went wrong, and what the team concluded about what it takes to use AI responsibly in a newsroom context.

Context

The workflow the team chose to automate is genuinely complex. Public records requests must cite the correct state statute to be accepted; the laws vary across all 50 states and rejection often follows from citing the wrong one. A journalist drafting such a request from scratch might spend an hour on a single letter. USA Today built an agent to handle this task.

Key method and takeaway

The development process encountered a recurring problem. The agent hallucinated — producing slightly wrong statute citations that would cause a request to fail. Jessica Davis, who led the work, is quoted noting that “slightly wrong, in a public records context, means the request fails.” The team concluded that confidence in the AI model was not sufficient grounds for deployment. What they needed was a formal evaluation framework: structured tests that could establish, with data, where the agent performed accurately and where it did not.

Once evaluations were in place, the pace of the work changed. Deployment time for new features dropped from months to within a week, and shipping multiple features within days became achievable. Davis’s core argument, as the article presents it, is that evaluations — not a journalist’s positive impression after testing — are the foundation a newsroom must build before it can trust AI at scale.

The case study is useful beyond journalism. Any content team deploying AI for a task where accuracy and compliance matter faces the same problem: general performance on standard tasks does not predict performance on the specific, constrained task the team actually needs done. The USA Today approach — build a formal evaluation, measure against it, use the results to define the scope of AI’s role — transfers to other professional writing contexts.

The article also distinguishes between what Davis calls assistive AI (text generation, information surfacing, tasks where a human reviews the output) and agentic AI (agents that take autonomous action toward a goal). This distinction matters for where evaluations are most critical: agentic tasks that produce outputs feeding directly into consequential processes need more rigorous measurement than assistive tools where a human reads and approves each output before it is used.

Who it is useful for

Newsroom editors and data journalists implementing AI workflows; content operations teams building AI-assisted processes for tasks where accuracy is a requirement; and anyone at organisations that want to move past pilots into reliable, scalable AI use.