Tennessee: AI-powered customer service across chat, email, and voice

The problem

Tennessee's Department of Safety and Homeland Security (TDOS) covers driver services, highway patrol, and state-level homeland security — but around 95% of the contact it receives from the public is about driver services: driving licences, appointments, and related enquiries. The volume is significant: on any given day, the department receives between 3,000 and 5,000 calls, but has the capacity to answer only about 1,000.

The gap is not in the complexity of what is being asked but in the sheer number of people asking it. It is a constraint most government agencies will recognise — demand grows with the population, but staffing does not keep pace. In Tennessee, the problem is especially difficult to address because expanding staffing levels is a lengthy and challenging process.

With hiring effectively constrained, the TDOS turned to technology to close the gap between the number of people who needed help and the number it could realistically reach. The department had explored this before. A previous chatbot had been in place, but it was a rule-based system that could only respond to questions that matched its preprogrammed options. Anything outside of those went unanswered. What the department needed was something that could handle the full range of questions the public actually asked, not just the ones someone had thought to anticipate in advance.

The solution

Starting in December 2025, the TDOS began rolling out a series of AI-powered tools across its main communication channels: chat, email, and voice. All three use the same data source — the agency's public website — and are built on Zendesk's agentic AI platform. They are separate tools, each adapted to how people communicate on that channel, but they draw on the same underlying information.

The website acts as the single source of truth.

A central design decision was to use the agency's own website as the only data source the AI draws from. The reasoning was practical: the website is the agency's official record of how its services work. If the AI gives a wrong answer, it means the website is wrong — and the website needs fixing regardless. This avoided the problem of maintaining a separate knowledge base that could fall out of date or contradict what the public sees online.

Channel 1: chat
The chatbot launched in December 2025 and handles roughly 1,000 interactions daily. It answers about 70% of the questions it receives, drawing on the website content. A further 10% are deflected — messages that are not genuine questions, such as people venting frustration about a news story or expressing general complaints without asking anything specific. About 20% of interactions are passed to a human agent because the bot cannot resolve them.

Channel 2: email
After the chatbot launched in December, the team turned to email — the department's second communication channel. The logic was the same: if the AI could answer common questions using the website as its data source, it should be able to do so regardless of whether the question arrived by chat or email. The email bot launched in February 2026 using the same underlying system and data.

In practice, the results were different. The email bot currently answers about 30% of enquiries, compared to 70% on chat. The gap reflects how differently people communicate on each channel. Chat users tend to type a single, direct question. Email users write at length — explaining their situation, providing background, recounting what has happened — before arriving at what they actually need. The AI struggles to extract the question from a longer narrative.

Channel 3: voice
The voice bot is in final testing and expected to launch shortly. Voice is the agency's largest channel — where the gap between demand and capacity is greatest. The challenge is different again: on chat, the bot can provide a link to an online service. On a phone call, reading out a web address is not practical. The team is working on text-back options so the voice bot can send a follow-up text message with the relevant link after the call.

The impact

1. Around 700 citizen questions are answered every day
The chat bot answers approximately 700 questions daily — roughly 25,000 per month. Before the AI was deployed, these questions either went unanswered (because staff could not get to them) or were handled by a rules-based chatbot that could not cope with anything outside its pre-programmed responses. The AI provides answers 24 hours a day, seven days a week, including nights and weekends when no staff are available.

2. Only 20% of chat interactions now need a human agent
Of the roughly 1,000 daily chat interactions, 80% are either answered by the bot (70%) or deflected because they are not genuine questions (10%). Only about 20% are escalated to a human. This means staff time is concentrated on the enquiries that genuinely need a person, rather than being consumed by questions the AI can handle.

3. The AI revealed problems with the agency's own website
Using the website as the AI's data source created a feedback loop that nobody had anticipated. When the bot gave a wrong answer, it meant the website was wrong. With a thousand chat interactions a day, the AI quickly surfaced problems with the website that had previously gone unnoticed. Wrong information, missing pages, and questions that citizens repeatedly asked but the website did not address. Within about a three-week period, there was a very rapid iterative process of cleaning up the data. The AI did not just answer questions — it showed the agency where its own information was incomplete or incorrect.

4. The agency can respond to emerging situations within minutes
One of the advantages of the approach is the speed of adaptation. When Tennessee was hit by a major ice storm in January 2026 that shut down government services across multiple counties, a new use case was immediately created within the AI system. When citizens asked whether their local driver services center was open, the bot could tell them which counties were closed. A human agent would have needed to look that up. Similarly, when citizens started receiving scam text messages claiming they owed tolls or fines, the agency was flooded with enquiries. A use case was created to identify these questions and tell people to delete the messages. Both responses were live within minutes of the need being identified.

Lessons learned

• Using the website as the data source turned wrong answers into a quality improvement tool. When the AI gave incorrect information, it pointed directly to content on the agency's website that was wrong or missing. The AI did not just answer citizen questions — it audited the agency's own public information at a speed and scale that manual review could not match. For any government agency considering AI-powered citizen services, the choice of data source has consequences beyond the AI itself.
• The speed of refinement is impacted by how the tool is managed. In this instance, one person was responsible for both the service and the technology, which made it possible to identify a problem, decide what to change, and make the fix quickly. This is not always how government teams are set up. If multiple parties are involved, the process flow may be exponentially longer. For agencies considering similar deployments, how the day-to-day refinement is managed — and how many handoffs are involved — will shape how quickly the tool improves.
• AI in government needs ongoing management. Managing an AI tool is similar to managing an employee. It is like a human being that you have to keep training day in and day out. It never stops. The private sector is creating dedicated “bot manager” roles for this work. In government, where creating new positions requires a lengthy approval process and existing staff have full workloads, the question of who monitors, refines, and maintains the AI tool after launch is often unresolved. The tool does not manage itself, and neither the business team nor the technology vendor typically has both the knowledge and the capacity to do it.
• The AI only provides information — it does not execute anything. The system does not modify accounts, process payments, update records, or take any action. It answers questions. This boundary is what allowed the process to move quickly and made it possible to iterate in a live environment: if the AI gives a wrong answer, that is a problem to fix, but it has not changed anything in the system. For agencies considering how much risk to accept when deploying AI, limiting the tool to information rather than transactions significantly reduces the consequences of errors.
• Different communication channels need different approaches. Chat works well because people ask short, direct questions. Email works less well because people write long narratives that the AI struggles to interpret. Voice introduces a different challenge: how to provide links and online resources to someone on a phone call. Using the same data source across all three channels provides consistency, but each channel requires its own design work to account for how people actually communicate on it.