Using AI to Bring Voice of Customer to the Forefront of Your Organization
Mining speech analytics data as a research tool
At face value, speech analytic tools offer an incredible value proposition: decode the mystery of why your customers are calling, provide real-time analytics and recommendations to agents, and even track conversion rates on sales calls. However, research by AI-powered conversation intelligence platform Tethr found that most investments in AI-powered speech analytics fails to deliver its intended return on investment. Why?
Most companies use the tool to automate quality assurance to score more calls than a human with a checklist ever could, and while that’s one of the best basic use cases for speech analytics, scaling quality control doesn’t actually improve quality unless the right insights are extracted and acted upon. It’s like using a powerful telescope to spy on your neighbor, noticing they’re about to commit a crime, and doing nothing about it.
In a recent webinar for CCW Digital, ‘Using AI to Bring the Voice of the Customer to the Center of the Enterprise,’ Matt Dixon, chief product and research officer at Tethr shared some buyer-beware tips on what to consider before shouldering the costs of a speech analytics system, as well as how to use it to fullest potential.
Problem 1: It’s too hard to get the insights out of the VoC data itself
Despite multimodal contact centers that enable customers to chat with a brand on Whatsapp or Facebook Messenger, the majority of Voice of Customer research is still captured in the voice channel. And while a speech analytics tool scores every call and provides an aggregate readout on key quality indicators, a human has to interpret those indicators and turn them into business objectives, such as spotting a new product opportunity and communicating it to the development team.
“The problem with that is that it’s really hard to do, and it’s much harder to do than you think,” Dixon warns. Vendors condition prospective buyers into thinking a speech analytics tool is a plug-and-play solution that uses call data to extrapolate how a company can improve its sales performance, boost NPS, or reduce customer effort, but the AI itself has to be pre-programmed to capture the specific data you want, and then told what to look for in that data. It needs a training set or a set of topics, which Dixon likens to a “roadmap for the AI.”
“You need to teach the AI how to detect customer sentiments like frustration, anger or uncertainty,” Dixon said. “You need to teach it how to detect agent behavior, competitive mentions, and different product, campaign or offer references.”
The use cases for speech analytics from one organization to the next are as different as their founding legacy, which is why speech analytics tools don’t come out of the box ready to go - rather, their machine learning capabilities must shaped by the user. Once you’ve established what you want the analytics tool to measure, you have to finetune those categories and training sets to eliminate false positives or negatives, which is time-consuming and expensive.
“Many companies are doing little more than what we call keyword spotting,” Dixon explains, referencing a method by which companies determine the main reasons customers are calling. “This is really a proxy insight at best; it’s not looking at the full context of the customer conversation.”
Without predictive or causal models, keyword spotting is but an anecdotal takeaway that likely reinforces what agents already know - after all, they’re the ones who listen to customers day in, day out. What it doesn’t doesn’t do is answer why the customer called with that particular complaint or asked that particular question. Is the product missing a key feature? Are you mismanaging inventory for a particular item? Should you switch to a more reliable shipping company?
Worse still, many organizations latch on to the wrong KPIs based on what vendors are touting. A much ballyhooed feature of speech analytics tools is their ability to “understand” customer sentiment by studying voice inflections during the call.
Perhaps the notion that AI can process human emotion is what makes this selling point so compelling, but as Dixon points out: “Very few companies are going to make a business decision based on a customer’s tone of voice, but they are going to base a business decision on the actual words that customers say.” Syntax analysis, he said, gives a better readout on what customers think, and not only because words have connotations.
Problem 2: The insights end up being trapped inside the tool rather than being used in the real world
At its most basic level, speech analytics provides you with an unstructured audio recording for every call, and transcribes it into unstructured text. Then what? A human has to structure the text data and mine it for insights. Imagine if you had transcripts of the last 90 calls to your call center in a folder on Google Drive. What would you do? Annotate comments in the margins? Even if your comments were actionable, they would still live in separate Google Docs.
Most companies merely teach the machine to find descriptive insights using current disposition codes - essentially automating the quality assurance checklist. That’s when the investment starts to diverge from business objectives, and executives begin demanding why the speech analytics tool guzzles so much time and money when it’s really just a fancy dashboard that helps scale quality control.
Companies buy these tools in the first place in order to improve outcomes like sales, NPS or CSAT, which directly touch the bottom line. But once again, without a knowledgeable analyst to interpret the data, the tool starts collecting dust. Some companies will hire data scientists to build causal models and manage the tool itself, but Dixon warns that data scientists are “in short supply, expensive and they turnover like crazy.”
To make the most of a speech analytics tool, you need to understand its limitations and costs in terms of time and manpower needed to implement it, finetune it, and maintain it. Your organization should be using the insights from call data to build an effective customer engagement model based on Voice of Customer demands. But that higher-level strategic thinking is the purview of the chief customer officer, not the machine.
Dixon compares AI to what a home inspector does for a prospective homeowner. “The AI will point out all your problems - there’s a puddle on the floor and your roof is leaking - but if you don’t know where it’s coming from and how to fix it, then your recourse is to put buckets on the floor to collect the water that’s leaking from your roof because you don’t know how to fix it.”
Agents need to be coached in how to use the tool and also trained to mine the data with an inquisitive mind.
Problem 3: Even if we get the insights and we can move the needle on things, it’s too expensive
Most companies are unaware that the sticker price on a speech analytics tool is the tip of the iceberg. Some vendors charge per-minute costs for ingesting, transcribing, processing and storing data. There may be hardware and software costs if it’s an on-premise solution.
Companies should also account for the maintenance and replacement cycle for the hardware and software, and any specialist training costs for the staff members who will be using the tool, or what it would cost to pay a salaried data scientist or team of data scientists. Some users ask the vendor to help with fine tuning and maintenance, which also drives up costs.
“Then you have problems of trying to ship the data to downstream platforms, such as CRM platforms or visualization tools,” Dixon said. An omnichannel contact center will seek to integrate its call data with an existing CRM or ERP, but some vendors charge steep fees to extract and migrate data, while others even require you to purchase an extract license.
Before investing in a speech analytics tool, it’s important to understand what it can and cannot do, consider what business objectives you want it to achieve, and then evaluate whether or not you have the manpower and financial resources to truly optimize this powerful tool beyond quality assurance automation.