The call center, which was staffed by public health nurses and volunteer medical professionals, was primarily focused on supporting patients who needed to be isolated or placed into quarantine units. To keep the center from being flooded with calls, Kirk worked with IT staff and a vendor to quickly stand up a COVID-focused chatbot.
The software platforms, which use artificial intelligence to respond to human questions, have become a key part of state and local governments’ rapid pivot during the pandemic, particularly as call centers were inundated and many agencies sent their workers home. Three-quarters of states are now deploying chatbots to respond to COVID-related challenges, including assistance with applications for unemployment insurance and fielding questions about symptoms and testing, according to the National Association of State Chief Information Officers (NASCIO).
For example, when volume at its call centers exceeded a million calls in March, the Texas Workforce Commission quickly stood up its own chatbot. Named Larry to honor the commission’s former executive director, the bot has responded to 6 million questions about unemployment benefits since its launch on April 1. In Missouri, a proposed $16 million plan to support call centers with chatbots over several years moved more quickly to standing up the technology in four departments to meet immediate needs.
“It helps you be resilient,” says Eric Roche, budget officer for the city of Pearland, Texas. “If your building loses power, the bot’s in the cloud and available 24/7. If you need to launch a similar one, you can copy and paste it and have it up and running quickly.”
Before the pandemic, chatbot technology was part of a larger push towards tools powered by the emerging field of artificial intelligence, joining autonomous vehicles and facial recognition among much-touted use cases for governments.
Chatbots are the top use of AI in the enterprise, according to Gartner, and the research and advisory company placed the technology squarely at the peak of inflated expectations in its annual “hype cycle” review of government technology adoption in 2019. The pandemic may have since established proven use cases that could move the technology further along the adoption curve more quickly, but key limitations remain to be addressed — both on the technology and organizational fronts.
Overcoming Chatbot Negativity
Even with rapid advancements in AI, too many government solutions are “still for the most part answering known questions with known answers for anonymous users,” says Forrester Principal Analyst Ian Jacobs. “That’s fine if you need the address of a planning office, but not fine if you’re trying to check on approval of the planning variance you applied for.”More than half (54 percent) of U.S. users have negative perceptions of interacting with chatbots, according to Forrester research. Jacobs doesn’t mince words when asked why.
“Chatbots sucked and people hated them,” he says. “The public sector was no different than any other domain. In fact, they may have been worse because resource-strapped agencies at the local level didn’t have the resources to invest.”
Even so, government solutions are rapidly evolving. Roche jokes that one of Kansas City’s first bots didn’t exactly endear itself with the city’s PR staff. The reason? Its name — Trashbot. But the phone-based bot and its Snowbot sibling allowed people to automatically report missed garbage pickups and icy roads over the phone, saving $9,000 in the first four months and reducing wait times, says Roche, then the Missouri city’s chief data officer.
Roche developed Kansas City’s first chatbot in 2017, a Facebook Messenger assistant focused on providing easier access to the city’s open data project. Despite the ease of launching the chatbot — creating it took two weeks and didn’t require writing a single line of code — the challenge was that “it became really hard to keep up to date and add new information to,” says Roche. “When we created the other bots, we had a plan to keep technical debt at bay.”
But as technology improves, government officials have to be prepared to anticipate new challenges. When Kansas City’s voice bots expected “yes” or “no” responses to questions, for example, they often stumbled on a common regional colloquialism: “yup.”
“There has to be someone vigilantly monitoring it,” says Roche.
Many recently launched government chatbots have been purpose-built for specific departments and topics — Knoxville’s chatbot, for example, launched in March focusing on just two areas: COVID-19 and the Census. However, Forrester’s Jacobs believes that ultimately government chatbots should aim to provide 311-style information across all agencies or departments. Doing so would provide a consistent experience — and the same answers to the same questions across the entire government.
To that end, the city of Williamsburg, Va., upgraded its text message-based chatbot to a Web version in July. The transition to the Web expanded use by 2,000 percent in the first few weeks, in large part because the chatbot was available on every page of the city's website, not via a text number that citizens had to track down in order to use the service.
As a smaller city, Williamsburg doesn’t have a 311 service or call center, so the chatbot has played a key role in helping citizens quickly find answers to common questions and to place field service calls. IT Director Mark Barham gives the upgraded technology high marks for handling the wider range of questions fielded online to date, but notes that more advanced AI is a necessity for chatbots to understand and respond to all requests. “You have to have some type of machine learning associated with these, because they’re not going to give you the right answer all the time,” he says.
Advances in AI bring challenges
AI is making great strides in improving chatbot technology, in part by training and learning from past responses and navigating longstanding data headaches like the inconsistent formatting of addresses. However, some of the technology’s improvements open the door for new challenges.With the adoption of technologies like OpenAI’s GPT-3, a text-generating algorithm that can generate uncannily human writing, erroneous answers will be tougher to flag. Roche likens the AI-generated text to “high school papers with a mix of facts and misstatements and outright lies that are still passable reading.”
As with printed and online materials, governments will have to contend with developing chatbots capable of conversing in multiple languages. Chip, the city of Los Angeles’ chatbot, now supports more than 50 languages, but for many governments ensuring nuanced real-time translation of complex subject areas such as COVID-19 symptoms will require extra care. “We’re very specific about the agencies we work with on translation to make sure those are truly accurate,” King County's Kirk says.
The next generation of chatbots will identify specific users and provide personalized answers to complex questions, such as the status of a zoning application or a specific fine or fee, according to Jacobs. But getting to that point will require governments to address two key challenges that are slowing development of other services. First, they will have to tackle the identity management issues that also are a necessity for one-stop Web portals capable of handling functions across multiple departments and agencies. Second, they will have to digitize and map out complex processes and workflows in ways that AI-powered chatbots can navigate.
“Vendors are doing pieces of that in the commercial market, but the public sector is lagging behind,” Jacobs says. “That’s only partially limited by the state of AI and more by all of the things before it — being able to map out complex governmental processes, understanding rights and responsibilities, and being able to absorb entire regulatory frameworks. That’s the bigger problem going forward.”
As AI becomes more autonomous in training itself, governments also will have to think about implicit bias that may be unintentionally reflected in the historical data or feedback that the chatbots draw from, says Roche. “Is it giving everyone across the city the same answers, or is it giving people at different income levels different answers?” he asks. “People hear about implicit bias and think there’s nothing they can do about that. There is — but it takes the right staff.”
A largely untapped opportunity will be using chatbots to empower governments’ own employees, allowing human call center employees, for example, to tap into the bots to quickly get answers that help them serve citizens better, Jacobs argues. “We’re starting to hear more about that in the private sector,” he says.
And while much attention is being placed on AI’s ability to help chatbots train themselves, government leaders can see how these lessons learned can be brought to bear on improving other services. In Williamsburg, Barham learned the top questions posed to the chatbot involved business licenses, which sent him to examine the city’s website.
“It means they can’t find the information on their own,” he says. “We have to look at how we’re aligning information.”
And listening to chatbots can help governments prepare for what’s next. In King County, public health officials monitor the questions posed to its COVID-19 chatbot daily. Doing so, says Kirk, has helped “get a sense of new and emerging questions to get a sense of what the public’s thinking about on any given day.”
