Minnesota Slashed Translation Costs by $100K a Month Using ChatGPT

Minnesota’s government translation pipeline used to be a mess—decentralized, expensive, and so slow that a single request could take a month. The state’s Enterprise Translations Office (ETO), launched in 2023, just flipped that script by weaving ChatGPT into its core workflow, and the numbers they’re posting now deserve a hard look from every other state agency wrestling with language access.

The ETO serves a state where over 20 percent of residents speak a primary language other than English, with Spanish, Somali, and Hmong leading the pack. Before the AI overhaul, departments were each doing their own thing with external contractors, leading to inconsistent quality and punishing delays. Now, ChatGPT handles the first pass on translations, which are then meticulously reviewed by human experts. The result is a system where the machine learns from every correction, so the same mistake never happens twice. Adam Taha, the ETO’s director, nailed the core tension: “The balance between accuracy and speed is the crux of the problem.” This workflow appears to have cracked it.

Since the beta, the office has churned through over 3,000 translation requests totaling more than 2 million words. Urgent jobs that once took weeks are now turned around in as little as two hours. The financials are equally stark—the state reports saving over $100,000 per month by reallocating resources away from the old contractor-heavy model. The team didn’t just chase raw speed, though. They built custom GPTs loaded with culturally specific terminology for Hmong and Somali, tackling the nuance that generic machine translation usually butchers. Karine Lao Her, the Hmong language team lead, said the work “has already made a significant difference in how our community connects with state resources.”

The program went from beta to full agency rollout in just four months, and it’s already spawning a new pilot for real-time voice interpretation using ChatGPT in a live setting with another state agency. It’s a pragmatic blueprint that acknowledges AI’s flaws by keeping a human firmly in the loop while refusing to let perfect be the enemy of dramatically better. Whether that quality holds up when you move from documents to the messy reality of live conversation is the real test.