It’s the third message from Rachel you’ve gotten but you haven’t had the chance to get to her email. She’s not pestering you, but she has consistently followed up with you once a week for half a month already and hasn’t lost track. You decide you should clear your inbox. You apologize for not replying earlier, but you were busy.
She totally understands and says ‘no problem,’ talks about that software you had looked at about a month ago and asks if you want more details. You ask one last question to hear if that program would cover this one nagging problem you’ve been working on at your desk for the last couple of months and she responds in the affirmative.
After sharing some pleasantries on a couple more emails in the thread, she sends you more info and offers to connect you with customer support. She sounds kind of nice and you’re pretty impressed with her sales abilities. You send her another message, hinting your company might be looking for someone to do sales in the near future. But there’s one problem.
She’s a robot.
That’s a real scenario that has played out over at Conversica, a Forest City, California company that raised $34 million in a Series B funding round last month for its AI-driven sales assistant that takes generated leads and does the follow-up work with natural-sounding messages.
“It interacts with end customers only via email,” Head of Product Sales Gary Gerber told Geektime. Conversica generates profiles with names and email addresses, then gets assigned a list of contacts. “It implements a sales assistant and the assistant literally becomes a member of the sales team.”
The company started off as a sales lead company in the automotive industry. They were having trouble keeping up with their contacts and even initiating in-person conversations between dealers and salesmen. That drove them to hire friends working in AI development to create the first iteration of the system, which became the priority when dealers found out what was driving increased efficiency in lead follow-through.
“The dealers said they didn’t care about the leads, they cared about the medium.”
She is a she for a reason
According to all the tests they have done, people feel most comfortable talking to a sales rep when they think she’s a she and she happens to be in her mid-20s. That results in usage of certain first names which might have been popular among new moms in the early ’90s. Altogether, Gerber brags they have a ridiculous 38 percent response rate.
“People could do this in theory but they spend most of their day prioritizing which leads to hit,” says Gerber, meaning some leads are never contacted based on guess work and some abandoned after just a couple of tries. But this system doesn’t ignore any of them, designing custom emails for each customer and follow-up messages and only stopping contact attempts after hearing a response. “A certain percentage of those abandoned leads would have been a good sale. It usually takes 8-12 times usually to reach someone!”
As mentioned before, Conversica assistants don’t launch a barrage of emails, waiting a few days or possibly a couple weeks before trying again.
Gerber notes, “She’s real simple. She’s not trying to sell anything. People feel very comfortable with her. She’s personal, friendly, never in a bad mood. She’s persistent.”
“Her goal is, like a dating service, her goal is to facilitate a conversation between the two of you.”
Gerber is sure that simplicity is also playing a huge role in the response rates, but certainly persistence and automation allows for more opportunities to break through with the messages eventually, hitting that 8-12 sweet spot and finally eliciting a reply.
“Rachel, like a person, will read that ‘No way was I ignoring you’ and say ‘Great,’ then formulate a response.”
The system overall is governed by what Gerber dubs the “AI Council,” which includes mechanisms that general NLP (natural language processing), measure sentiment, analyze intent, and a couple more focused on other angles relevant to crafting a response to a customer.
They have some big name clients, including Epson, Oracle, HP Enterprise, Chrysler, and even IBM. The system is still guided by human hands as a customer service team still takes some referrals and helps build Conversica’s repertoire by adding new data.
“Ninety-five percent of the emails are responded to completely anonymously. Five percent of the time the assistant isn’t sure. That boots it to one of our people. The person formulates the right response.”
One of the drawbacks that created that 1-in-20 scenario might not surprise people working on NLP or translation.
“They’re typically . . . where our system has issues, is sarcasm,” Gerber says.
That doesn’t set off many red flags though.
“Virtually never do people come back and say ‘Youre a bot! Quit spamming me!’ But what we do get are some pretty amusing responses. We’ve had companies try to recruit our customer assistants, saying ‘Oh wow, you’re so persistent.'”
Then of course some people get kind of attached.
“I will only work with Stephanie!”
The ATM effect?
When automatic teller machines (ATM) were first introduced in the ’70s, people fretted many bank tellers would lose their jobs. Banks actually ended up expanding their services by freeing up tellers to deal with more advanced things like investment banking. As debate surges over the need to compensate a perceivably antiquated human workforce with the onset of automation, Gerber sees the buddings of more labor.
“Of course many people’s first reaction is, ‘Oh my gosh! I’m gonna lose my job!’ Some customers say, ‘Oh great, I can save money by firing some people!’ But a huge percentage of customers have told us ‘because you’re increasing my sales funnel and generating better leads, I had to hire more people to deal with the new customers.'”
They have started offering service assistants and begun to beta test other programs like assistants for accounts receivable. They also plan to cover more industries.
“The key difference with what we do and general purpose is a chatbot is trying to cover everything. He/she is trying to be all things to you, but we are very focused on a specific conversation focused on driving behavior.”
So how close is Conversica to creating an audible version of the technology?
“Can we do it with voice? I suspect the technology isn’t quite there,” posits Gerber, who reminds Geektime that he isn’t on the development or design side of the business. “The catch is that it works so well with email right now there is no pressure to do that.”
The technology is also useful for SMS. In that scenario, like a nervous high school boy, Rachel or Sheila might wait a couple minutes between texts to respond just to avoid the perception she is coming on too strong.
“Technologically you can expect a response within a minute. Sometimes she decides not to respond so quickly so she doesn’t freak you out. What cadence makes sense? Should I wait?”