DigitalGenius is announcing its Human+AI customer service platform today, along with a $4.1 million seed investment. The platform integrates with existing customer service software suites — like Salesforce, Zendesk, and Oracle — to automate the most repetitive parts of customer service through AI and machine learning-powered chatbots. The work is to augment the process, while still keeping the human element decidedly at the center of things. It’s interesting to note that Salesforce was part of the deal, as that could conceivably help the startup scale quickly in this space, thanks to Salesforce’s massive distribution network and its suite of automation products ripe for AI.
AI is the biggest buzzword of 2016
Haven’t you heard? #Botopia is officially upon us. Facebook made it so! I talked to DigitalGenius chief strategy officer Mikhail Naumov to clarify what AI is and isn’t. “It’s important to decipher between Hollywood AI and practical AI you can use today,” he said. Right now, the media is throwing around technical terms like “AI,” “machine learning” or “deep learning” interchangeably, buzzword-izing the whole landscape. As Benedict Evans from Andreessen Horowitz aptly put it in an epic Tweet rant, “People try to bolt AI onto every new user interface model. But we don’t actually have HAL 9000, and may be 50 Nobel Prizes away from that.”
Mic sufficiently dropped.
“AI is a concept, not an actual technology,” Naumov told me. “If you talk to the people at DeepMind and Facebook’s AI Labs, AI gets all the attention in the press, but the industry itself is far away from Artificial General Intelligence,” he said, “It’s not Ex Machina.”
If you think of it as a hierarchy or taxonomy, AI is at the top, and machine learning is an offshoot, Naumov explained:
Machine learning is an approach or set of techniques where you use massive data sets to train machines in semi-supervised ways…Deep learning is a step below machine learning in the tree. Deep learning is a way to combine neural networks (machine learning frameworks) to produce outcomes of unsupervised machine training based on very large datasets. Those datasets can be images, like pictures of cats in Google. They can be historical moves of the world champion of the game Go.
Or, in DigitalGenius’ model, they can be historical transcripts of customer service logs. The deep learning algorithm that powers the DigitalGenius Human+AI platform is trained on conversation data from these logs. It unlocks new value by teaching the AI how to answer customer service questions across text-based communication channels like email, social media, and mobile messaging platforms. This frees up human agents to focus on solving more complex problems. Again, this isn’t unassisted AI. Google didn’t hand DeepMind the instruction manual to Go and tell it to play. That doesn’t mean it wasn’t an incredible leap forward, though. Technology’s limiting factor for AI is being challenged more rapidly than was once believed possible, which could change, well, everything in how we interact with devices.
AI comes to life when you set parameters
DigitalGenius focuses on customer care because, despite its inherent complexity, for most businesses, there are a relatively finite number of customer problems and solutions for a set of products. This is in contrast to sales or marketing, where there are relatively infinite states of content and services and modes of measuring intent indicators on how people find and buy things.
“Customer service is so repetitive,” Naumov said. “There are gold mines of historical transcripts that aren’t used for anything.”
That’s where DigitalGenius comes in with what’s really deep learning techniques, creating what are called mathematical vectors out of words. In this case, those words come from a huge dump of chat transcripts or call logs from a company like BMW (a DigitalGenius customer). From there, it’s like moves on a Go board, but simpler — there’s a defined set of parameters and trainable techniques for what makes a successful customer service agent, and DigitalGenius claims to mimic that behavior. The company includes what they call a “confidence threshold” in their AI/customer interactions, and if it drops below a certain point, a human customer service agent steps in or approves the messaging.
DigitalGenius has no console, so the company trains expensive systems already entrenched in a service organization on their deep learning models.
I asked Naumov why there’s so much buzz around this area, seemingly all of a sudden. “One of the reasons we’ve seen such a jump we haven’t seen before is because of the improvement of hardware. We’re training our model using GPUs. [Five years ago], you’d use GPUs to play video games.”
“Additionally, now AWS is on-demand, as needed, and startups can begin to do this in a significant way,” he said. This follows a growing trend in the industry, as rapid evolution in NLP (natural language processing) and machine learning techniques like deep learning are pushing the technology forward. A couple of short years ago, some of the dominant tools using NLP –like scraping the Web for search terms (customers complaining on Twitter or talking about your brand in some positive or negative way) — were lucky to be 50 percent accurate out of the box.
Naumov wisely warned about the importance of managing expectations. “If every brand starts building a chatbot — and they haven’t learned the mistakes — they’re going to have bad experiences, like Microsoft Tay.” So the question becomes, who will make the chatbots we actually want to use? I personally don’t think the technology is about ordering flowers on Facebook. With the limiting factor of technology out of the picture, what will determine bots that are useful (and invisible to us) will be beautiful and elegant design. Who can build interaction models that humans will like? That remains to be seen, but watch out, because many thousands of developers now have the tools to will it into existence.