The US 2020 election was unique in many aspects; fought in the middle of a global pandemic, and in a very polarised environment. But Joe Biden’s victory was fueled by an extensive digital campaign, which relied on new methods to reach out to voters. Among these was conversational AI deployed via Facebook Messenger and designed by Amplify.ai, based in Palo Alto, California.
“The Biden campaign was actually our fifth US Democratic presidential campaign client. We entered into that partnership with a lot of learnings under our belt,” John McCrea, CMO, Amplify.ai explained to indianexpress.com over a call.
Amplify.ai entered the US political market early 2019. The company has also helped work with MyGov in India with AI-powered chatbots to help with the COVID-19 pandemic.
In the Biden campaign, the conversational chatbot was deployed on Facebook Messenger in the final three weeks of the campaign. The aim was to reach out to potential voters, especially those who were undecided.
People would comment on a post from say Joe Biden, and then Amplify.ai’s software would reply via Messenger. To the voter, it could appear as though Joe Biden was himself replying, though McCrea says their intention was never to fool anyone. If a user asked if this was a bot, the AI would confirm the same.
“Engagement in the form of comments acts like fuel for our conversational AI platform. We were able to take voters who were supportive of the candidate and get them to do a number of things that are positive for the campaign, like signing up for the email list or opting in to receive text, or even directly, asking them to donate,” Mcrea explained further.
With the conversational chatbot, the idea was to connect the campaign with voters in battleground states. These were voters the campaign was not able to reach by phone. Sending volunteers to knock on their doors was also not possible given the pandemic.
The company’s chatbots helped Biden’s campaign reach close to a quarter-million voters in 14 battleground states, a fact also acknowledged by Vishal Disawar, the Digital Organizing Platforms Director for Biden’s campaign in a Medium post.
According to McCrea, they were able to help Biden supporters execute a voting plan such as figuring out their closest polling booth, timings. Further, if a user was undecided still after talking to the AI, the system would do a live handover to volunteers to help persuade them. Only 3,000 handovers took place, which McCrea said shows how effective the conversation AI was in the first place.
Learning from the other side
Biden wasn’t the only big campaign that Amplify.ai has handled. The company’s first client was Kamala Harris in May 2019, when she was running as a Democratic Presidential candidate. Amplify.AI has also handled chatbot deployments for seven Democratic Senate campaigns.
McCrea admitted that when the company started with Harris’ campaign in 2019, they knew they were entering uncharted waters with respect to conversational AI. “Our political landscape is so polarised that it’s actually difficult for humans to talk to each other about political matters, especially if they disagree. We knew we were going into the deepest end of the pool with respect to natural language processing and conversational AI,” he explained.
The problem meant that he had to take on the role of lead conversational UX designer for these conversations. “The key challenge was how do we phrase things in such a way that the voter will tell us upfront, are they positive neutral or negative, and then have very different flows for those different pathways,” he explained.
An interesting learning for Amplify.ai from these early campaigns was that their AI could engage with voters across the political spectrum. For example, initially if the chatbot encountered a Donald Trump supporter, the team knew there would be no donation coming. In such scenarios, the company didn’t even bother designing a conversational flow. They just had the conversation end with a nice, polite, thank you.
But that abrupt end in the conversation made Trump supporters really mad, according to him. “They were already a little bit angry and it just made them more so. So we developed a deeper flow that was essentially like we want to hear from all voters, and we’d love to understand your views better. That worked like a charm,” he said.
This resulted in Amplify.ai getting a completely new data set with over a half-million Trump supporters who would answer questions for them, like how they identified politically, did they vote for Obama, etc. Not only did the conversational AI overcome barriers to communication with voters across the spectrum, it also helped add to the company’s learnings on building future conversational workflows.
Keeping it natural
For Amplify.ai’s chatbots, unnatural conversations were never really a challenge. “I’ve seen more conversations between American voters and AI than any other human has. This feeling of stiffness, inauthenticity really isn’t the core problem here. Honestly, it’s almost the opposite end problem, even though we don’t attempt to present the AI as being the candidate, or as being a human,” he said.
According to him, a pretty high percentage of people thought that they were talking to a human in these conversations. And if at any point in the conversation someone asked whether it was a bot, they would stop the interaction to let them know that was the case. But that very rarely happened.
Creating the chatbot for Biden also meant that the chatbot had to embody at least some of the qualities of the Presidential candidate. “We knew we were working with the likely next president of the United States and I kept humans involved and looked every day for opportunities to improve the quality of the conversations,” Mcrea explained.
“Joe Biden, as a figure, is a nice, empathetic, warm person. And so I felt like the bar that we needed to hold ourselves to was a lot higher,” he added.
One example was this around donation conversations within the chatbot. The pandemic also meant that many supporters, who might have donated, were perhaps unable to do so, because maybe some of them lost their job.
“Before the pandemic, if someone said no to a donation, we might reply, ‘No worries’. But I realised that’s a terrible way to handle things. For example, if you say might you be willing to make a donation and the person says, ‘I like to, but I’ve lost my job’. You don’t want to reply, ‘No worries’ to that,” McCrea elaborated.
Conversational AI as the future, but…
Despite the success, he admits that conversational AI as technology is still mainstreaming. “One of the challenges is that kind of not only being able to be multilingual, but also to be able to handle things that are part one language, part another,” McCrea said.
In fact, one of the biggest learnings for Amplify.ai has been that existing natural language processing engines were really not up to the task of what they wished to achieve. They had to build their stack from scratch.
In his view, most natural language processing engines have been trained on book level pros sentences and paragraphs, with perfect spelling and punctuation, which is not what happens when humans interact with a chatbot. For example, someone might spell you as a simple ‘u’ in a conversation with a chatbot.
Still, consumers have changed the way they want to interact, and are living in a messaging-first world. “They want instant self serve, 24X7 through chat. That’s just the reality consumer brands and governments and political campaigns need to adapt to,” he added.