How a Gateway City Reduced Non-English Call Volume by 52% with Multilingual AI Chat

This case study describes the multilingual AI search and chat deployment at a mid-sized city government in the Mountain West, serving approximately 340,000 residents across a metropolitan area recognised as a resettlement destination for refugee and immigrant populations. The city was chosen for the case study because its language access challenge is more diverse than most: American Community Survey data documents 38 distinct languages spoken by populations of 1,000 or more, and no single non-English language accounts for more than 22% of the LEP population.

The city has requested anonymity for the case study to allow more candid discussion of the challenges and mid-deployment adjustments. Metrics have been verified by the city's IT director and the department of communications. Financial figures are presented in ranges to preserve confidentiality.

The Challenge Before Deployment

The city's language access programme had three components before deployment: static Spanish translations for approximately 40 high-priority pages, a Google Translate button on all other pages, and a contracted phone interpreter service available during business hours in the top 12 languages. The programme cost approximately $340,000 annually and was not achieving its stated goals.

Two specific pain points drove the search for a new approach. First, call centre data showed that 41% of calls to city services were being routed to interpretation services, at a per-call cost between $8 and $14. Second, resident feedback consistently identified that the website was not usable for anything beyond the 40 statically translated pages, and the Google Translate button produced translations for government-specific terminology that residents described as 'sometimes wrong, always confusing'.

"We were spending a lot of money to provide something that residents were telling us didn't actually work. And the population we most needed to reach was the one most affected by the failure. It wasn't sustainable and it wasn't equitable."

, City IT Director, referenced case study organisation

The Pilot

The city ran a 4-week pilot beginning in October 2025 with AI Search and AI Assistant deployed on a subset of high-traffic service pages: the residential permits page, the utility billing page, and the parks and recreation programme registration page. The pilot was announced publicly through the city's communications channels, including specific outreach to community organisations serving the largest non-English-speaking populations.

Pilot metrics were tracked daily. The city set three success thresholds: at least 500 distinct multilingual queries during the pilot period, task completion rates for non-English users equal to or exceeding English-language users on the pilot pages, and no significant quality or accuracy issues reported through the AI chat feedback mechanism.

The pilot exceeded all three thresholds. Multilingual query volume was 2,340 over the 4 weeks, more than 4x the threshold. Task completion rates for non-English users on the pilot pages were 71%, compared to 68% for English users. And no accuracy issues were reported that required content or configuration changes; two minor terminology adjustments were made to the AI chat greeting messages based on community feedback.

Full Deployment

Full deployment across the city services website and the parks and recreation portal completed in November 2025, three weeks after the pilot concluded. The deployment approach preserved the city's existing translated content for the 40 high-priority pages, added AI search and chat as a supplementary layer for the remaining content, and included staff training for the communications and IT teams responsible for ongoing management.

Two specific configuration decisions shaped the deployment. The AI chat was configured to escalate any query flagged as high-emotional-content (queries mentioning threats, safety concerns, or legal disputes) directly to human agents. And the AI answers included an explicit callout when the resident's question could not be fully answered from the city's content, routing them to the appropriate department contact rather than providing a partial or speculative response.

The Six-Month Results

Call Volume Reduction

The 52% reduction in non-English call volume was the headline result and the one the city communications team highlighted in reporting to city council. The reduction was uneven across languages. Spanish-language call volume dropped 61% (the highest single-language reduction), followed by Vietnamese (54%), Amharic (48%), and Arabic (43%). Smaller-population languages saw more variable reductions, in most cases in the 30–50% range.

The city verified that the call volume reduction was not simply displacement to another channel. Email inquiries did not increase measurably, walk-in traffic at city offices was stable, and community organisations serving immigrant populations reported that residents were increasingly using the website for questions they previously would have called about.

Language Distribution Discovery

One of the most operationally valuable outcomes was language usage data the city did not previously have. The AI platform's analytics showed exactly which languages residents were using, in what volumes, and for what types of questions. This data revealed two languages, Nepali and Karen, that had significant resident usage but were not on the city's prior list of top-supported languages. Both were subsequently added to the city's phone interpreter contract and to the static translation programme for future high-priority page translations.

Financial Impact

The annual financial impact modelled from the 6-month data is significant. Interpretation service costs dropped by approximately $180,000–220,000 annually. Contact centre load reduction (partly attributable to AI chat, partly to improved multilingual search) translated to approximately $340,000 in avoided contact centre expansion that had been planned for FY2026. Against an annual AI platform investment in the mid five figures, the ROI is straightforward.

What Would They Do Differently

Asked what they would do differently in retrospect, the city IT director identified three things.

1. 1Start the community outreach earlier. Awareness in the immigrant communities that AI chat existed and worked in their language was a significant driver of adoption. The organisations that heard about it early adopted it fastest. Building relationships with community-based organisations two months before launch, rather than at launch, would have accelerated the first-90-day results.
2. 2Add the parks and recreation portal to the initial deployment scope rather than as a subsequent expansion. Parks and recreation was where families first engaged with city services in language, and getting the AI available there from launch would have driven earlier trust and engagement.
3. 3Build the staff feedback loop into the deployment from week one. The city's content team started reviewing zero-results queries at week eight. Reviewing from week one would have surfaced content gaps sooner and improved answer quality faster.