Real-time quality coaching in customer service

Customer service is in the midst of an AI-driven reboot. 

Chatbots and other advanced automation tools are shouldering ticket workloads with increasing scale and efficiency. Human agents are using the tools to more capably and quickly resolve higher-priority caseloads. 

New AI-powered assistants are emerging to play a more subtle but critical role in service operations: serving as real-time quality coaches that can analyze language, tone, grammar, and relevance and provide agents with actionable advice and suggestions on the fly.

At a recent Data Science Salon—a conference for AI and machine learning practitioners—in San Francisco, Anshuman Guha, staff machine learning engineer at Freshworks, shared how AI is powering new forms of real-time agent assistance, improving scoring of agents, and delivering customers and agents a better experience. (The new capabilities are part of an update to Freddy Copilot, part of the Freshworks Freddy AI platform.)

“As the agent is typing,” Guha explained, “windows will pop up suggesting a change in tone—for a relevance or sentiment issue, or grammar problem, or something else that needs to be corrected.”

Read also: How machines learned to chat

In customer support environments, immediate responses are essential. “The popups are there to grab an agent’s attention in less than a second,” said Guha. Minimal lag improves customer experience while helping reduce cloud compute costs. Those can add up quickly, Guha explained, when a company has 100,000 agents handling millions of customer interactions annually.

Here are some of the new AI-powered features and the benefits they offer:

Tone detection and response suggestion

Generative AI-powered assistants can detect tonal cues in agents’ responses and suggest improvements. When they identify a negative tone, Copilot recommends a response that is more empathetic, professional, or solution-oriented.

To detect negative tones, assistants are trained on a blend of real-world examples and examples derived from gen AI prompts; combined, they form baseline standards for tone. New customer conversations are then added to the baseline to help refine and clarify standards. 

With a supervised machine-learning model that blends artificial and real data, “we have a lot more control,” said Guha. “We are better able to define what is a bad tone from a customer service perspective.”

Improved relevance of agent responses

There is no margin for drift or delay when it comes to agent responses in customer conversations. Their relevance, Guha explained, is often what stands between a successful resolution and an unhappy escalation or abandonment.

If a customer asks a particular question but receives an off-topic response (such as: “Are you still connected?”), the conversation is often over before it begins. With technical information especially, real-time guidance is essential for agents to keep up with customers who are more versed in the product or problem than they are.

Customers receive irrelevant responses up to 20% of the time, said Guha. Features developed by Guha’s team at Freshworks can reduce that number by relying on large language models (LLMs) to make continual suggestions, helping agents respond with the most relevant, helpful information.

By using AI to score actual customer responses on a scale of 1 to 10 and training the AI assistant against these inputs, Guha’s team improved the relevance of agent responses. “The intention of this feature is to have it be based solely from a language perspective,” he said. “We are able to check or help flag to agents that it’s okay, they are going in the right direction.” 

Multi-language grammar correction 

Grammar-check software has been available since the 1970s; using it within the context of the agent-customer conversation should remind business leaders of its everlasting value. (Women are 81% less likely to buy a product advertised with spelling or grammar errors, according to SurveyMonkey. Similar results apply to the language used in customer service.)

At Freshworks, these new AI features tap into Grammarly’s open-source knowledge base and can perform agent corrections simultaneously in seven major languages. That presents a significant advantage for agents for whom English may not be a first language.

Since grammar is a rules-based discipline, real-time correction is the simplest of the new AI features to implement. “Relevance and tone are always debatable,” said Guha, “but grammar is pretty straightforward.” 

One last key feature of the new tools is the option for agents to pick and choose what to do with the suggestions. They can click “fix it all” and have their response rewritten entirely—or decide on which elements they agree with and which they don’t, giving the rep ownership of their work and communication.  

We want to hear from you! Please send us your feedback, and get informed about exciting updates from The Works. Drop us a line: theworks@freshworks.com.