A bear starts chasing two hikers. One of them stops to wear running shoes. The other asks: “What are you doing! You can’t outrun a bear,” to which the hiker replies: “I don’t have to outrun the bear, I only have to outrun you.”
That’s a well-worn story Abhimanyu Singh sometimes uses to pitch to companies wanting to use artificial intelligence to handle the deluge of customer support queries.
In Singh’s world, the hiker with the running shoe is the human agent backed up by artificial intelligence. “For us to make a case, we only have to deliver better results than humans,” he says. His company, Agara Labs, makes the
running shoes artificial intelligence software.
Singh’s spiel mostly works because enterprises are hungry to automate processes to save costs and the solution he’s pitching is the next level of automation. It helps that the company has engineers with a proven track record and a storied Silicon Valley investor to back the pitch.
Most of us haven’t heard of Agara Labs, but perhaps Halli Labs rings a bell? Halli Labs was a stealth mode startup acquired by Google in July 2017, just a few months after it was founded. Turns out, it was more of an acquihire. It’s co-founder Pankaj Gupta is one of the engineering heads for Google’s ambitious India gambit now. After the acquisition, four others joined Google with Gupta. The rest of the team, comprising of top talent from Twitter, IIT-BHU and Stayzilla (where Gupta and Singh worked earlier), run Agara Labs named after a lake in Bengaluru.
The startup has been in stealth so far. But of late, it has started talking to venture capital firms to raise more capital. It is also ready with a product that promises companies that outsource customer support and other functions generous savings. The idea is to use deep learning techniques to automate tasks like writing emails to customers.
“Nearly 80% of the customer queries were for 10% of the questions. So it’s not really that crucial to apply AI to these problems,” says Shamik Sharma, former CTO of Myntra who’d automated a large chunk of customer queries at Myntra. So a rules-based system that dishes out canned responses might just do the trick. But is that all that Agara Labs is up to?
At the Agara Labs office in Bengaluru, Arjun Maheswaran and Akhilesh Sudhakar, experts in deep learning and natural language processing, show me an email generated by Blue Dot, the virtual agent they’ve built. Maheswaran was earlier a part of the deep learning team at Twitter, where he lead efforts to apply AI principles to detect spam and abuse on the platform. Sudhakar used to be a researcher at the Natural Language Processing Lab of IIT-BHU and then at Microsoft Labs.
The email in front of me is stripped of the details of the sender and recipient. To surmise: the sender is angry and is complaining about a product he bought. He’s written in to say that he won’t be using the product again.
And then I get to see a response to the mail. The reply begins with a well-worded apology to placate the customer, followed by a line thanking the customer for writing in and a few steps to make sure the fault doesn’t occur again. The virtual agent has thrown in a complimentary voucher.
As someone who has worked at a call centre nearly 10 years ago, I felt like I was qualified to judge the product which was supposed to help agents save time. While I took calls, I’d watched an older colleague, Jose, plug away into the night trying to get ahead of the emails that never stopped coming. I wouldn’t have known that the reply in front of me was generated by a robot. The Blue Dot-generated mail could have saved Jose some time, for sure.
Agents used to take about 10 minutes to resolve a query. They’re now able to do this in 5-7 minutes depending on the complexity of the case, says Singh who estimates that after full-blown deployment Blue Dot can save costs in the range of double-digit millions annually for its current client — a consumer products major from North America whose name Agara executives decline taking thanks to non-disclosure agreements.
What’s more striking is that the reply exhibits ‘empathy,’ a valuable trait taught to customer support agents during their training days and it also tries to get the customer to use the product again.
“We haven’t told it that if you find this, do this. All we’ve given it is a lot of training history saying this is how people handled it. Now you learn how to do this,” says Singh. At their best, we’ll see virtual agents capable of handling customer queries end to end. In this ideal world, companies can scale their systems to answer more queries by simply throwing more computing power at it.
But will humans be replaced? Not quite. Companies aren’t yet ready to do away with humans yet. As much as there is curiosity in the C-suite about using AI, there isn’t enough trust in such systems because they are early on the horizon. “Our systems are designed to augment humans,” says Singh.
Such systems not only rely on past data but also feedback from cases that it has handled. It’s a non-starter if there’s no data to train these models. This is where Agara Labs has had a bit of beginners luck. The consumer products client has shared nearly 1.5 million customer queries and human responses to those queries with Agara.
The deep learning model adjusts itself depending on the feedback it receives. So how does the feedback loop work? Responses generated by the machine are sent to human agents for feedback. “By the time our solution goes into production, it’s ideally better than your best guy,” says Singh.
This can get complicated. For instance, global companies have to handle customer queries in different languages, from different geographies and account for cultural nuances. For example, an early lesson you’re taught if you’re into troubleshooting computers is that you can ask Americans to open up a computer to replace parts but Europeans will likely baulk at the idea.
Agara Labs also stands out as the only Indian startup in recent times that has raised seed capital from Kleiner Perkins Caufield & Byers, a top US venture capital firm which said goodbye to India in 2014. That’s probably because the size of the opportunity ahead of companies that bring artificial intelligence is ripe for a venture capital play. The market is large: India’s IT- BPO industry is expected to make some $137 billion in 2018-19. The Indian BPO industry is about $28-$30 billion in size. And then there’s scope to find an exit on the way as large services companies with stockpiles of cash turn acquisitive.
Automation is the new fad in the customer support industry. Automation is the number one priority for four out of five enterprise C-Suite executives surveyed by HorsesForSources.
But much of it is currently focussed around what’s called Robotic Process Automation. The idea is to automate repetitive tasks and help save cost. It’s a glorified ‘if-else’ based system where software mimics a very well defined process previously carried out by humans. It works well when we know exactly what is supposed to happen next. Companies like UiPath, Thoughtonomy, Automation Anywhere, Pega, BluePrism and WorkFusion are leaders in the space.
Deep learning-based systems start where robotic process automation ends. It comes in when rules aren’t defined. It tries to mimic human judgment. It’s the next level of automation which some industry insiders call ‘cognitive automation’. “From plain and simple automation, the next level is cognitive automation where there’s a certain aspect of deep learning is involved,” says Akshat Vaid, engagement manager at management consulting firm Zinnov. “What you’re saying (about Agara Labs) sounds like cognitive automation.” The market for robotic process automation is likely to be about $6.5 billion in size by 2020, Zinnov estimates.
One of the leaders in applying cognitive technologies to solve enterprise problems is IPSoft. “RPA is one level. The next level is where cognitive technology comes in,” says Uday Chinta, Managing Director, IPSoft. The company counts banks such as Credit Suisse and insurance companies such as MetLife as its customers.
The holy grail, according to Chinta, is a cognitive agent who directly engages with the user and solves a problem without the user even having to know the inner workings of the processes. That’s a difficult proposition. People can ask the same question in many different ways. The rules don’t apply here because you can’t write rules. “You need to be able to naturally understand the same request that comes in different ways,” says Chinta. Then, there’s the need to understand a user’s context. For instance, if the user asks to reset his password, the cognitive agent needs to understand if it’s for his email account or social media and so on.
Agara Labs wants to develop and deploy similar solutions on voice and chat-based customer support channels. “We’re developing systems for chats and phone calls. They’re real-time systems and you don’t have the luxury of time. We expect that to take at least a couple of quarters,” says Singh.
Subscribe to FactorDaily
Our daily brief keeps thousands of readers ahead of the curve. More signals, less noise.