MelodyArc: How we implemented AI as the Primary Worker in Customer Support
The AI Catalyst
On the Lex Fridman podcast in March 2023 (#367), Sam Altman pondered the potential of AI in job markets, particularly Ai in customer service. He speculated,
“I would say customer service is a category where I could see there being way fewer jobs relatively soon. I'm not entirely certain about that, but I could believe it ... like whatever call center employees are doing now.”
Since then, there's been a surge of startups harnessing ChatGPT and similar AI tools with the goal of revolutionizing customer service. Interestingly, no company has truly managed to deliver that quintessential human touch, the expertise we expect from a great human agent, using AI alone. Instead, the trend leans towards AI as a sidekick—assisting human agents. Why? Because customer service isn't just about answering questions. It involves understanding nuanced client specifics like policies, brand inclinations, and past interactions, then crafting the best solution to minimize customer effort and maximize satisfaction.
Table of Contents:
We founded MelodyArc in 2021, before the wave of generative AI took off. Looking around, we noticed the tech landscape was increasingly enamored with the idea of replacing humans with AI. But having been in the customer service field for years, we knew it was not an easy task. At MelodyArc, AI agents serve as the primary workers, offering efficiency and consistency in customer support. But there's a human element. Expert human agents stand by to augment the AI, validating, guiding, teaching, and adding that essential human touch to the AI's decisions.
AI as the Primary Worker
Our platform seamlessly integrates AI and human agents, drawing on the strengths of each. In the sections that follow, I'll delve into how our platform functions at a high level and share some key insights we've gathered along the way.
1. Internalizing external data
To have AI serve as the primary worker in customer support, it needs access to a client’s systems.
Like other platforms, we connect to client systems using APIs. Data is either pushed to us via webhooks or we retrieve it through triggered API calls. The primary entry point for this data is what we term the “ingress system”. This system channels tasks, such as customer messages, usually from popular ticketing platforms like Zendesk, Gorgias, or Kustomer directly into our platform.
Additionally, to enable AI and agents to actually take actions, not just generate responses, we also integrate with other platforms. This includes ecommerce solutions like Shopify, logistics providers like ShipBob, and tech tools such as Jira. For situations where an interface lacks API access, we've found workarounds, like using manual methods including webform submissions and custom crawler bots. That, however, is a topic for another time 😉
It's important to note that these integrations aren’t merely about data acquisition; they equip our platform with the context and capabilities to function. The combination of this data, and how MelodyArc utilizes it for both AI and human agents, makes our approach functional and efficient. We'll delve into the specifics of this utilization in the next section.
2. Servicing outside of straight lines
Delivering exceptional, on-brand customer support is complex. It's marked by multiple attributes: knowledgeable support from adept agents, precise adherence to brand voice, swift and accurate application of policies, and tailored experiences based on individual customer history.
Achieving this kind of support, especially on a large scale, is directly linked to a company's operational procedures and the proficiency of agents in understanding and executing those procedures. Traditionally, organizations capture these procedures in documentation akin to wikis. These pages often contain decision trees, rule charts, and flow diagrams outlining how to navigate various scenarios.
However, this model has its limitations. A primary challenge is the difficulty of keeping these resources updated. Changes in one area can cause a domino effect, necessitating updates in multiple interconnected areas. Additionally, transforming these complex updates into a format that's both human-readable and presentable is cumbersome.
Because of this, we took a different approach. Instead of formulating fixed procedure flows, we introduced “Points”—containers that encapsulate fragments of information, whether simple or complex. These Points can contain varied data: a list of available shipping statuses, rules determining if a shipment is on schedule, standard return policies, protocols for manual password resets, or even actionable commands like processing refunds or placing orders.
A Point is triggered when its predefined conditions or inputs are met. For instance, to utilize a Point focused on tracking a shipment, certain criteria like customer verification and the presence of a tracking number are required. Once triggered, a Point generally produces one or multiple outputs—like package tracking details in this instance.
These outputs can then become inputs for subsequent Points. This cyclical interaction continues until all triggerable Points have been activated.
Some Points are designed to decipher factors demanding qualitative judgment—assessing, for instance, the emotional tone of a customer's message or discerning their preference between a refund or a replacement. Addressing such qualitative queries demands the involvement of an agent, be it AI or human.Our AI-powered customer support platform alerts an agent whenever such Points are activated, and the cycle of input-output continues. platform alerts an agent whenever such Points are activated, and the cycle of input-output continues.
The advantages of our Point-based system include:
Adaptability & Personalization: Instead of being tied down to a rigid procedure, Points allow the creation of dynamic, undirected graphs. This means each customer interaction can follow a unique, personalized path rather than a one-size-fits-all solution.
Simplified Updates: By revising or adding new Points, information can be easily updated in the system without disrupting the entire procedural architecture. This fluidity accommodates complexities that would otherwise overwhelm human agents.
The solutions above enable AI to assume the primary worker role. Here’s how our AI agents operate:
AI First: Every incoming task, like a new customer conversation, is assigned to an AI agent.
Navigating the Graph: The AI agent traverses the dynamic graph, created by the Points, to resolve the task. This includes making attempts to interpret and answer qualitative questions.
Asking for Help: If, at any point, the AI agent is unsure about a decision, it seeks help. Help might come from another, perhaps more specialized AI, or more often, from a human agent.
The Human Role: When called upon, a human agent provides specific inputs, refining the AI agent's understanding and bolstering its confidence in tackling Points.
Maintaining Ownership: Even with human input, the AI retains ownership of the task. It determines the best route to resolution, whether there’s one solution or several. If the AI ever struggles to meet a confidence threshold for a decision, it can always call upon its human counterparts for further guidance.
In practice, especially with new clients, many tasks initially require human agent reviews to ensure optimal outcomes. Over time, as the AI gathers more data and becomes familiar with client needs and nuances, its dependency on human oversight diminishes.
We've provided a high-level overview of MelodyArc platform’s implementation of AI for customer support. We'll delve deeper into individual parts in future entries. There is tons more to cover!
Reach out to learn how MelodyArc can be personalized to support your business.
Frequently Asked Questions
What are the benefits of using AI in customer support?
AI in customer support helps both businesses and customers. Businesses save time, deliver more consistent results, and save money. Customers get faster answers, 24/7 support, and personalized interactions. It is worth noting, that the benefits of AI are reversed if the response does not answer the question asked by the customer, or reduce their effort to solve the problem.