How Talkmap Was Built: Cracking the Enterprise Call Center Nut with Generative AI
Radical new approaches to long-standing problems appear thanks to generative AI
Big idea #1: Start with a model of how customer conversations work
The bun-to-beef ratio of generative AI is out of whack—for every 1,000 clickbait articles about AI, you’ll find one with meat. That’s because beyond image generation, improving your emails, grammar, and coding, very few enterprise-grade generative AI solutions have proven themselves ready for prime time. For example, while most analysts agree customer support is an essential use of generative AI, try to find someone who's done it, and you’ll be sorely disappointed.
One AI innovator has cracked that nut. They’re called Talkmap.
Talkmap has proven that popular AI thinking—all you need is ChatGPT–is both inaccurate and incomplete. But popular thinking that all you need is ChatGPT is wrong. Here’s how they created an innovative view of language and built a company worth millions that uses AI to help call center agents turn customers into fans — years before ChatGPT became a household topic of conversation.
The mission: scale the human touch for billion-dollar brands
Big consumer product brands are worth billions, and each conversation agents hold with customers is a thread that forms the fabric of that brand. Every interaction with a customer builds or reduces the balance in the satisfaction, loyalty, and retention bank account.
The quest to improve customer relationships isn’t new. Toll-free telephone numbers sparked innovation beginning in the 1980s. Today, the market for customer engagement tools is over $40 billion. Each solution promises to increase empathy, optimize operations, automate chat, or analyze customer sentiment at scale. (1)
Unfortunately, human engagement is complicated, and most tools fail to improve it: Deloitte estimates that more than 50% of call center conversations fail to meet customer expectations. Those disappointed customers led to $1.6 trillion in lost revenue. (2)
Excellence in customer experience remains elusive for one reason: although the gold standard of customer success is the human touch, the human touch doesn’t scale.
Yet while the nail is still the same — happy customers — there’s a new hammer for the job: generative AI.
A vision of conversational intelligence at scale
Founded by Jonathan Eisenzopf, a former call center leader, Talkmap started with a simple vision.
Imagine 10,000 customer support agents holding millions of conversations every day. Now imagine you could partner each agent with a friendly robot. That robot knows everything knowable about your customer — why they love your product, the problems they’ve encountered, and their most recent interactions with your brand.
That robot uses her knowledge to whisper in your agent’s ear recommendations about where to lead conversations — from technical support to account questions. In this example, you’d have thousands of robots that effectively guess at the intent of each conversation.
And imagine that robot has a 360-degree view inside your business: special offers, knowledge of outages, and late-breaking news.
This robot is unbiased, well-trained, and governed.
And it’s designed for scale: you can deploy millions of variations of these robots that suit each agent at a marginal incremental cost because they run in software.
This is what Talkmap set out to build.
How Talkmap cracked the nut
Talkmap started to crack the conversational intelligence nut in 2017. Seven patents (3), hundreds of features, and dozens of blue-chip customers later, they’ve designed a system, TalkDiscovery, that uses generative AI to scale the human touch.
Eventually, they developed four big ideas about conversational intelligence that help generate predictions that help humans improve empathy and service at scale. Let’s explore their history based on these four ideas.
Big idea #1: Start with a model of how customer conversations work
The first problem conversations pose is that computers don't know how they work. By 2017, speech analytics were in the market—tools that turn voice into data. But that data is only valuable if humans know what they’re looking for: the exact word sequence or perfect keyword. Speech analytics doesn’t provide context, and for conversations, context is king. You must understand intent.
As a result, using speech analytics is like finding five needles in a haystack, then knitting a sweater with them.
Eisenzopf defined a mathematical model of human language with conversations in mind. He designed his model around goals, or what he called intent. Intents are goals, like “I want to fix my router” or “I need to understand a billing question.”
Fortunately, although linguistic researchers at Stanford, MIT, and Carnegie Mellon have created theoretical ideas about language structure, Eisenzopf defined a conversation as a collaboration among two or more people to reach a common goal. Those goals reveal themselves as a pattern as a conversation unfolds. Each utterance can be used to predict the goal of each participant. Large enterprises have millions of conversations with nearly infinite combinations of patterns, intent, and paths to solve them.
This goal-seeking idea set the stage for AI-driven transformers that direct conversations toward the desired outcome of both parties.
But mathematical models are just the beginning. Although they form the foundation of a solution, Talkmap needed to build a house.
Stay tuned for the next in our Big Ideas series…
Authors
Sources & Notes
1: Customer Contact Center Sizing Analysis and the Impact of COVID, Fortune Business Insights, 2023.
2: Top Growth Strategies for Tech CEOs of AI Technology Providers, March 10, 2023, Arup Roy
3: Talkmap’s approach is instructive for any enterprise to apply mission-critical AI. If you’re a patent-reading nerd like us, you’ll want to read these: Conversation Classification - US10,929,611 and US11514250B2, Training Chatbot from a Corpus - US11,004,013, Conversation Flow Exporter - US10,896,670