As a major retailer selling everything from party food platters to duvet covers to kids' clothes, UK-based retailer Marks & Spencer has a customer base as diverse as its product offerings.
And yet, according to enterprise architect Akash Parmar, the company had, for a long time, tried to pigeon-hole customers into three or four interactive voice response (IVR) options any time a call came in.
Unsurprisingly, it didn’t go well.
“The key challenge we were facing was the fact we just didn’t understand. We were getting ten million calls a year and didn’t clearly understand what these calls were about,” he said.
It goes to show that the challenges facing a multi-billion dollar company can be the same ones plaguing a one-person start-up: namely, understanding and then communicating effectively with your target audience.
To solve this challenge, the company created a Natural Language Platform (NLP) l that captured more than 60,000 utterances from customers and then categorized them according to intent. Customers are now routed according to their message’s intent, not just the keywords they speak, and according to Akash, accuracy is over 90 percent.
Getting to good, though, wasn’t easy, and required not only some powerful problem solving, but some internal soul searching too. We sat down with Parmar, where he shared his tips about how company leaders can apply his company’s learnings to the understanding of their own struggles and communicating with customers.
1. Keep it simple and agile to start.
“In the early days of experimentation, it’s very important to build a platform that is flexible and can be changed quickly to try new things out.
Use text to speech for prompts instead of recorded messages so that different versions of welcome or disambiguation messages can be tried quickly. We started by saying, ‘How can we help you?’ The customers were leaving long utterances that were difficult to classify. When we changed it to, ‘In a few words, please tell us why you are calling?’ we started to get clear utterances we could use to build an effective intent model.”
2. Get everything you can out of every interaction.
“Listen to the customer and extract all the value from the customer’s utterance, and only ask relevant questions. Make sure that the customers get something valuable back for every effort they put in. In our case, if the customer says they have lost a card, we ask if its a bank card or a loyalty card but if the customer says they have lost a bank card, then we don’t ask a follow-up question.”
3. Dig into the ambiguous.
“The team was focused on making sure we went through every ambiguous utterance and made sense of it. We use to play a game where at the end of the day the person in charge of tuning the model would send a list of utterances which he/she couldn’t make any sense to the entire team. The person who could decipher an utterance got a pound in prize money. This game taught us that Jamaican Order means to make an order and Brad Pitt means Bra Fit.”
Also, ambiguous customer utterances pointed out where we asked ambiguous or confusing questions. For example, we asked “ Do you want to book a new bra fit appointment, or amend an existing one” and customers responded with “Yes” rather than telling us which of the two they were calling about. It is very important to make the customer journey on a natural language platform truly conversational and we learned that the hard way.
4. Employ human insight at key stages of the conversation.
“On top of the training the AI model, it was important that we train our advisors and educate our customers to make sure we get the most out of this platform. Unclassified calls ended up with a pool of advisors. If the intent was amending an appointment but it wasn’t routed properly due to an unclear customer utterance, the advisor would ask the customer to call back and say they want to amend an appointment, and they will be routed to the right team. The customers were happy to do that and we saw a number of these kinds of calls coming down over time.”
5. Involve your entire team, and iterate quickly with small but meaningful changes.
“We instilled a mindset shift to help the team realize that, with the right team and hypercare, there are ways to try things out in a production environment without causing a massive outage and scaring the customers away.
Instead of spending months building a big solution, spending a lot of money and then trying to justify that it was a success, we could make small changes in days, learn from the feedback and build a product that gives the right output before spending money and time to scale it out.”
6. Show your team your solution, and trust them even in their mistakes, to achieve internal buy-in.
“Show people a working prototype rather than just talking about it on paper. Every presentation I made had a slide with a phone number, which I asked everyone to call. I opened up the dashboard that showed, in real-time, that whatever they said on the phone got transcribed, classified, and routed accurately.
This was very powerful, as any doubt they had about the platform disappeared instantaneously, and for the rest of the presentation, we could talk about where could we go with this in the future.
Internally, it is very important to trust and empower people who are trying to go down this path and give them room to make mistakes and learn from them. Every success and failure was a collective effort of the entire team.”
Ultimately, Parmar said, it’s about solutions that serve not only your customers but your team as well.
“You’ve invested time and effort, and you want to be rewarded for it. And if you are rewarded for it, you will spend more time and effort. And more and more we do it, customers will be trustful, and the experience and the bond between the customer and the organization grow stronger and stronger.”