Ashish Nagar is the CEO and founding father of Level AI, taking his experience at Amazon on the Alexa team to make use of artificial intelligence to rework contact center operations. With a powerful background in technology and entrepreneurship, Ashish has been instrumental in driving the corporate’s mission to boost the efficiency and effectiveness of customer support interactions through advanced AI solutions. Under his leadership, Level AI has turn out to be a key player within the AI-driven contact center space, known for its cutting-edge products and superior implementation of artificial intelligence.
What inspired you to go away Amazon and begin Level AI? Are you able to share the particular pain points in customer support that you just aimed to deal with together with your technology?
My background is constructing products on the intersection of technology and business. Although I even have an undergrad degree in Applied Physics, my work has consistently focused on product roles and organising, launching, and constructing latest businesses. My passion for technology and business led me to AI.
I began working in AI in 2014, after we were constructing a next-generation mobile search company called Rel C, which was much like what Perplexity AI is today. That have sparked my journey into AI software, and eventually, that company was acquired by Amazon. At Amazon, I used to be a product leader on the Alexa team, repeatedly in search of opportunities to tackle more complex AI problems.
In my last yr at Amazon, in 2018,I worked on a project we known as the “Star Trek computer,” inspired by the famous sci-fi franchise. The goal was to develop a pc that might understand and reply to any query you asked it. This project became generally known as the Alexa Prize, aiming to enable anyone to carry a 20-minute conversation with Alexa on any social topic. I led a team of about 10 scientists, and we launched this as a worldwide AI challenge. I worked closely with leading minds from institutions like MIT, CMU, Stanford, and Oxford. One thing became clear: at the moment, nobody could fully solve the issue.
Even then, I could sense a wave of innovation coming that might make this possible. Fast forward to 2024, and technologies like ChatGPT are actually doing much of what we envisioned. There have been rapid advancements in natural language processing with firms like Amazon, Google, OpenAI, and Microsoft constructing large models and the underlying infrastructure. But they weren’t necessarily tackling end-to-end workflows. We recognized this gap and wanted to deal with it.
Our first product wasn’t a customer support solution; it was a voice assistant for frontline employees, corresponding to technicians and retail store employees. We raised $2 million in seed funding and showed the product to potential customers. They overwhelmingly requested that we adapt the technology for contact centers, where they already had voice and data streams but lacked the fashionable generative AI architecture. This led us to comprehend that existing firms on this space were stuck prior to now, grappling with the classic innovator’s dilemma of whether to overhaul their legacy systems or construct something latest. We began from a blank slate and built the primary native large language model (LLM) customer experience intelligence and repair automation platform.
My deep interest within the complexities of human language and the way difficult it’s to unravel these problems from a pc engineering perspective, played a big role in our approach. AI’s ability to know human speech is crucial, particularly for the contact center industry. For instance, using Siri often reveals how difficult it’s for AI to know intent and context in human language. Even easy queries can trip up AI, which struggles to interpret what you’re asking.
AI struggles with understanding intent, maintaining context over long conversations, and possessing relevant knowledge of the world. Even ChatGPT has limitations in these areas. As an example, it may not know the most recent news or understand shifting topics inside a conversation. These challenges are directly relevant to customer support, where conversations often involve multiple topics and require the AI to know specific, domain-related knowledge. We’re addressing these challenges in our platform, which is designed to handle the complexities of human language in a customer support environment.
Level AI’s NLU technology goes beyond basic keyword matching. Are you able to explain how your AI understands deeper customer intent and the advantages this brings to customer support? How does Level AI make sure the accuracy and reliability of its AI systems, especially in understanding nuanced customer interactions?
We’ve six or seven different AI pipelines tailored to specific tasks, depending on the job at hand. For instance, one workflow might involve identifying call drivers and understanding the problems customers have with a services or products, which we call the “voice of the shopper.” One other might be the automated scoring of quality scorecards to guage agent performance. Each workflow or service has its own AI pipeline, however the underlying technology stays the identical.
To attract an analogy, the technology we use is predicated on LLMs much like the technology behind ChatGPT and other generative AI tools. Nonetheless, we use customer service-specific LLMs that we now have trained in-house for these specialized workflows. This permits us to realize over 85% accuracy inside just just a few days of onboarding latest customers, leading to faster time to value, minimal skilled services, and unmatched accuracy, security, and trust.
Our models have deep, specific expertise in customer support. The old paradigm involved analyzing conversations by picking out keywords or phrases like “cancel my account” or “I’m not completely happy.” But our solution doesn’t depend on capturing all possible variations of phrases. As a substitute, it applies AI to know the intent behind the query, making it much quicker and more efficient.
For instance, if someone says, “I would like to cancel my account,” there are countless ways they could express that, like “I’m done with you guys” or “I’m moving on to another person.” Our AI understands the query’s intent and ties it back to the context, which is why our software is quicker and more accurate.
A helpful analogy is that old AI was like a rule book—you’d construct these rigid rule books, with if-then-else statements, which were inflexible and always needed maintenance. The brand new AI, however, is sort of a dynamic brain or a learning system. With just just a few pointers, it dynamically learns context and intent, continually improving on the fly. A rule book has a limited scope and breaks easily when something doesn’t fit the predefined rules, while a dynamic learning system keeps expanding, growing, and has a wider impact.
An amazing example from a customer perspective is a big ecommerce brand. They’ve hundreds of products, and it’s unimaginable to maintain up with constant updates. Our AI, nevertheless, can understand the context, like whether you’re talking about a selected couch, without having to always update a scorecard or rubric with every latest product.
What are the important thing challenges in integrating Level AI’s technology with existing customer support systems, and the way do you address them?
Level AI is a customer experience intelligence and repair automation platform. As such, we integrate with most CX software within the industry, whether it’s a CRM, CCaaS, survey, or tooling solution. This makes us the central hub, collecting data from all these sources and serving because the intelligence layer on top.
Nonetheless, the challenge is that a few of these systems are based on non-cloud, on-premise technology, and even cloud technology that lacks APIs or clean data integrations. We work closely with our customers to deal with this, though 80% of our integrations are actually cloud-based or API-native, allowing us to integrate quickly.
How does Level AI provide real-time intelligence and actionable insights for customer support agents? Are you able to share some examples of how this has improved customer interactions?
There are three sorts of real-time intelligence and actionable insights we offer our customers:
- Automation of Manual Workflows: Service reps often have limited time (6 to 9 minutes) and multiple manual tasks. Level AI automates tedious tasks like note-taking during and after conversations, generating customized summaries for every customer. This has saved our customers 10 to 25% in call handling time, resulting in more efficiency.
- CX Copilot for Service Reps: Service reps face high churn and onboarding challenges. Imagine being dropped right into a contact center without knowing the corporate’s policies. Level AI acts as an authority AI sitting beside the rep, listening to conversations, and offering real-time guidance. This includes handling objections, providing knowledge, and offering smart transcription. This capability has helped our customers onboard and train service reps 30 to 50% faster.
- Manager Copilot: This unique feature gives managers real-time visibility into how their team is performing. Level AI provides second-by-second insights into conversations, allowing managers to intervene, detect sentiment and intent, and support reps in real-time. This has improved agent productivity by 10 to fifteen% and increased agent satisfaction, which is crucial for reducing costs. For instance, if a customer starts cursing at a rep, the system flags it, and the manager can either take over the decision or whisper guidance to the rep. This sort of real-time intervention could be unimaginable without this technology.
Are you able to elaborate on how Level AI’s sentiment evaluation works and the way it helps agents respond more effectively to customers?
Our sentiment evaluation detects seven different emotions, starting from extreme frustration to elation, allowing us to measure various degrees of emotions that contribute to our overall sentiment rating. This evaluation considers each the spoken words and the tonality of the conversation. Nonetheless, we have found through our experiments that the spoken word plays a rather more significant role than tone. You possibly can say the meanest things in a flat tone or very nice things in an odd tone.
We offer a sentiment rating on a scale from 1 to 10, with 1 indicating very negative sentiment and 10 indicating a highly positive sentiment. We analyze 100% of our customers’ conversations, offering a deep insight into customer interactions.
Contextual understanding can also be critical. For instance, if a call starts with very negative sentiment but ends positively, even when 80% of the decision was negative, the general interaction is taken into account positive. It’s because the shopper began upset, the agent resolved the difficulty, and the shopper left satisfied. Alternatively, if the decision begins positively but ends negatively, that is a distinct story, despite the proven fact that 80% of the decision may need been positive.
This evaluation helps each the rep and the manager discover areas for training, specializing in actions that correlate with positive sentiment, corresponding to greeting the shopper, acknowledging their concerns, and showing empathy—elements which can be crucial to successful interactions.
How does Level AI address data privacy and security concerns, especially given the sensitive nature of customer interactions?
From day one, we now have prioritized security and privacy. We have built our system with enterprise-level security and privacy as core principles. We do not outsource any of our generative AI capabilities to third-party vendors. All the pieces is developed in-house, allowing us to coach customer-specific AI models without sharing data outside our surroundings. We also offer extensive customization, enabling customers to have their very own AI models with none data sharing across different parts of our data pipeline.
To handle a current industry concern, our data shouldn’t be utilized by external models for training. We do not allow our models to be influenced by AI-generated data from other sources. This approach prevents the problems some AI models are facing, where being trained on AI-generated data causes them to lose accuracy. At Level AI, the whole lot is first-party, and we do not share or pull data externally.
With the recent $39.4 million Series C funding, what are your plans for expanding Level AI’s platform and reaching latest customer segments?
The Series C investment will fuel our strategic growth and innovation initiatives in critical areas, including advancing product development, engineering enhancements, and rigorous research and development efforts. We aim to recruit top-tier talent across all levels of the organization, enabling us to proceed pioneering industry-leading technologies that surpass client expectations and meet dynamic market demands.
How do you see the role of AI in transforming customer support over the subsequent decade?
While the overall focus is usually on the automation aspect—predicting a future where bots handle all customer support—our view is more nuanced. The extent of automation varies by vertical. For instance, in banking or finance, automation could be lower, while in other sectors, it might be higher. On average, we consider that achieving greater than 40% automation across all verticals is difficult. It’s because service reps do greater than just answer questions—they act as troubleshooters, sales advisors, and more, roles that cannot be fully replicated by AI.
There may be also significant potential in workflow automation, which Level AI focuses on. This includes back-office tasks like quality assurance, ticket triaging, and screen monitoring. Here, automation can exceed 80% using generative AI. Intelligence and data insights are crucial. We’re unique in using generative AI to realize insights from unstructured data. This approach can vastly improve the standard of insights, reducing the necessity for skilled services by 90% and accelerating time to value by 90%.
One other vital consideration is whether or not the face of your organization ought to be a bot or an individual. Beyond the essential functions they perform, a human connection together with your customers is crucial. Our approach is to remove the surplus tasks from an individual’s workload, allowing them to deal with meaningful interactions.
We consider that humans are best fitted to direct communication and will proceed to be in that role. Nonetheless, they’re not ideal for tasks like note-taking, transcribing interactions, or screen recording. By handling these tasks for them, we liberate their time to have interaction with customers more effectively.
