How M&S cut costs building a natural language processing platform

M&S receives more than 10 million calls per year. But until recently, the chain had little information about the reason for those customer queries.  

As part of a organisation-wide digital transformation programme - which has included migrating off mainframe infrastructure to cloud-based systems and implementing new warehouse management systems - Bogdan Grigorescu, AI platform manager at M&S, helped deliver a natural language processing (NLP) platform for customer calls.

(NLP is a branch of artificial intelligence that deals with the interaction between computers and humans).

“We didn't know very well the reason for contact. What was the intent behind the call? The data was insufficient… siloed and the process was very manual - leaving lots of space for error. It was very costly too,” he told delegates at the Virtual AI Summit, during London Tech Week. 

Grigorescu’s team adopted a custom-build, multi-cloud platform, nicknamed ‘Ava’, which delivers a conversational experience for the caller. 

The company previously had switchboards in 13 different stores across the UK and Ireland (UKI), with call routed through a legacy phone system.

“We devised and implemented a natural language processing platform that accurately detects the reason for contact and uses that in managing the call. While also delivering conversational experiences... It is also very cost efficient, saving over £700,000 per year in platform costs alone.”

No lock-in

The automated system is able to recognise the intention of each call, so the customer service team is aware of the nature of the query and can deal with it straight away before speaking to the customer, with 35% of the call volumes fully ‘self serve’ - meaning customers don't need to speak to a human at all.

Another key benefit is the platform is it's vendor and technology agnostic, so any one of the products it is based on can be replaced within one week. In addition it is channel agnostic, so is capable of processing SMS, email, chat and social media queries. 

The team worked with a variety of partners to build the platform, says Grigorescu. “We used a lot of tools, varied as well, but we used them in harmonious ways.” 

For storage it used Azure and Heroku on the Salesforce platform as an active backup, while the natural language process engine was built with Google DialogFlow, Cloudspeech, Wavenet and Twilio. Collaboration tools included Skype for Business, Slack and Office365. That is to name just a few of nearly 30 companies Grigorescu lists as being involved in the project. 

It's an approach which has paid off. Following the roll-out of the platform, customer engagement rose 40% to 95%, call routing accuracy was up by 70%, and average handle time fell by 10% - the equivalent of just under 20 seconds per call. 

“Behind every successful solution there is a reliable platform… We delivered on the objectives and vision of building a platform that just works.”