Solving the live agent vs AI conundrum: The human-in-the-loop approach
Michael Housman is Chief Science Officer and co-founder of live chat platform RapportBoost.AI. Prior to founding the company he spent nearly a decade architecting data science platforms for early-stage technology companies seeking to leverage advances in machine learning in order to build more intelligent software applications, most recently as the Chief Analytics Officer at Evolv.
Housman has been published in a variety of peer-reviewed journals, presented his work at dozens of academic and practitioner-oriented conferences and has had his research profiled by The New York Times, Wall Street Journal, The Economist and The Atlantic.
In this interview with CX Network, he discusses his human-in-the-loop approach to breaking down the walls between live agents and artificial intelligence, how organisations can ensure they don’t lose the human touch when automating processes and he provides his insights on the biggest artificial intelligence and intelligent enterprise developments between now and 2020.
ZDR: Zarina de Ruiter, Editor, CX Network
MH: Michael Housman, Chief Science Officer and co-founder, RapportBoost.AI
ZDR: Hi Michael, many thanks for joining us on CX Network today. I’ve been reading about your unique approach to eliminating friction between the live agent and AI when it comes to chatbots, which you call the human-in-the-loop system. Could you explain this for our readers?
MH: Absolutely. We think that simply replacing humans outright with bots is a recipe for disaster. In fact, I read recently that Facebook pulled back the reins on their bot platform because 70 per cent were failing out of the gate. That’s bad customer experience and that’s going to leave a bad taste in people’s mouths.
Instead, we try to build a mutually beneficial relationship between the humans and the artificial intelligence. The AI we’ve built is capable of analysing millions of conversations and identifying actions that can produce a measurable improvement in the conversation. It does that for each and every interaction and offers up that guidance to the chat agent but it’s the human chat agent – our “human in the loop” – who ultimately makes the decision about what guidance to follow and what to discard.
The system observes when the chat agent follows our recommendations and then observes whether there’s a change in the expected outcome. When it sees that the human took our advice and the conversation improved, the artificial intelligence learns about what works in the wild and it can render that advice more frequently.
So in that way the AI helps the human and the human trains the AI. We think that’s the right approach.
The AI we’ve built is capable of analysing millions of conversations and identifying actions that can produce a measurable improvement in the conversation.
ZDR: Can you share an example of how you’ve approached this with one of your clients and the success they’ve seen on the back of this?
MH: We’ve found a big opportunity for this technology in retail. For example, with one e-commerce retailer that does about $25 million in annual revenues, we delivered three key insights to them about best practices for their chats. They didn’t trust the system entirely at the time so they only shared those three insights with half their agents and then used our technology to follow those agents over a three-month test period.
What they found is that those agents exhibited between 10 and 70 per cent improvement in each of those three behaviours. That that drove a 7.6 per cent increase in total order revenue and a 7.8 per cent increase in conversion rate. It was a small test that yielded a big result and we’ve identified an opportunity for a $2.3 million revenue enhancement if they do those behaviours plus a few more.
ZDR: That is really great! From a customer perspective there is also a big debate between chatbots and the risk of losing the valuable human touch along the way. As an expert on emotional intelligence in AI, how do you solve this conundrum and ensure the human connection doesn’t suffer?
MH: This is the big thing that no one seems to be talking about. Everyone’s focused on just automating interactions to reduce headcount and not realising that we’re all naturally social beings that crave interaction and a personal touch.
In our own research, we’ve found that the vast majority of the success of an interaction (whether measured by customer satisfaction, buying behaviour, or repeat visits) is driven by how you say something as opposed to whatyou say. Researchers don’t seem to get it, but what’s funny is that when I share those results with lay-people, their reaction is usually something along the lines of: no duh.
I think the way you sold that is by first reverse-engineering what makes an interaction engaging and authentic. That’s why our first step with any client is to have the machine learning engine analyse all of their existing chats. Then, once we feel like we have a good sense of how best to engage their customers, we offer up feedback and recommendations to their human agents.
As the artificial intelligence learns about what the human touch is and why it matters, we’ll eventually point the technology at bots. It’s easy to programme your bot to add a smiley face at the end of a specific message but it’s really hard to teach a bot to pick up on the nuances of conversation and empathise with customers when needed or mirror the customer’s level of excitement.
That’s our ultimate goal: to build emotionally intelligent bots.
ZDR: As artificial intelligence and chatbots are rapidly becoming a huge part of the wider CX market, what do you predict will be the biggest AI trends, challenges and/or disruptors affecting the industry by 2020?
MH: To the point earlier, I think that once the basics are fixed and the bots don’t fail so quickly, then the next thing people are going to work on is making the bots more engaged and personalise. So that not only means enhancing their EQ but also making them capable of remembering previous interactions, knowing your preferences and adapting to ensure that they provide a better and better experience as they learn more and more about the customer.
I also think that we’re going to see a massive uptick in the extent to which bots are used for sales-facing endeavours. Right now, most bots have been programmed to engage in customer service. That’s going to change quickly and we’re going to see a lot more instances of Alexa-like platforms where you can tell a bot what you’re looking for and it’ll offer up useful recommendations and, ultimately, complete the transaction.
I read somewhere that conversational commerce is expected to grow to a $600 billion market, which sounds insane but when you consider the growth of virtual assistants, bots and voice-activated bots like Alexa, I actually think it’s pretty reasonable.
ZDR: Our recent research shows that while 36 per cent of businesses are currently just in the beginning or planning phase of their AI journey (with an additional 6 per cent seeking vendor information and 7 per cent being in pre-deployment stage), by 2020 a majority of 53 per cent of organisations want to be established. What key steps should they take now to help them get there?
MH: It’s a long road, but my advice to all organisations is to start small and identify areas of the business that could benefit from: (a) increased automation; or (b) enhanced decision making through the use of AI. Those are really the two things that AI does and I’d be looking for specific areas of the organisation that could benefit from either or, ideally, both of them.
The other advice I’d give is to not just hand over the reins to a machine right off the bat. I know this sounds crazy but I spoke a few months ago to a large insurer that was exploring the use of bots without having utilised any live chat agents. I thought it would likely be a disaster because they hadn’t learned yet about how their customers would utilise chat.
We’re big believers in “human in the loop” and we think you should start with humans, help them do their job better with AI and then slowly shift the highly repetitive tasks towards machines. Let the humans focus on what they do best.
ZDR: Respondent in the research also say that the biggest barriers to AI investment to reach that next stage include (1) Integration with existing infrastructure/legacy systems; (2) How quickly it can demonstrate ROI; and (3) Lack of understanding on what solution area to focus on. How can they break through these barriers?
MH: Completely agree. The third point – identifying the right area to focus on – is really about identifying the right business priorities and areas where AI can make an impact. It’s tricky but I’d start with a matrix, list all the departments / initiatives of the organisation and then rate them based on: (1) how important they are to the business; and (2) how much of an impact AI can make.
From there, you need to identify a solution – likely through a vendor – and I’d be very wary of anyone who says that they require massive integrations or a long time to realise ROI.
Let me offer up an example: the first stage of our engagement with a customer is what we call RapportRoadmap where we tell customers to give us their chat data and KPIs (in whatever format it’s in; no integrations), we feed it to the machine and we come back to them within 30 days with actionable insights and a dollar figure representing the ROI they can expect to achieve. Our RapportCoach tool then requires some lightweight integrations to help them start to achieve that lift.
If step one of an engagement was scoping out a massive integration with multiple systems, then I would tell the customer to be legitimately sceptical. I’ve rarely seen solutions like that work well.
ZDR: When it comes to the wider intelligent enterprise landscape (the suite of novel technologies that are changing how companies operate internally and externally, such as AI, IoT, RPA, machine learning, and data analysis procedures), which do you see to have the most impact between now and 2020 within the CX space – and how?
MH: Honestly, I have the biggest hope for technologies that are aiming to democratise data science and put it in the hands of lay people. Organisations now amass so much data and the real rate limiting factor is the availability and bandwidth of people who can analyse it and generate insight.
There are a ton of BI tools and dashboards – like Tableau, Birst, MuleSoft and countless others – that enable basic analytics, but I’m more excited about tools like DataRobot that are trying to put data science in the hands of people who haven’t trained in it for years.
I think that’s especially true of customer experience where people have been seeing dashboards for decades. But if you give virtually anyone the ability to pose a question, test hypotheses using machine learning models, arrive at a conclusion and then implement that change, you’re going to see massive improvements in customer experience. That’s what’s going to drive the biggest impact.
ZDR: Finally, what top tip can you share with organisations to allow them to derive the most value from developing their AI capabilities between now and 2020?
MH: Don’t try to eat the elephant and don’t turn to massive AI solutions that purport to solve all your problems. Start small, be focused and look for problems where a very specific application of AI can potentially achieve a lot of good. Once you’ve gotten a few successful projects under your belt, you’ll realise that that momentum can continue to build to larger and larger projects.
To access the full-length article, click here.