As business uses for artificial intelligence move from theory to practice, enterprises grapple with an age-old dilemma — build or buy?
Nutanix does both. That helped them decide to build what became known as SupportGPT, a chatbot for answering tough, complex questions about using Nutanix’s hybrid multicloud software for IT operations. After analyzing GenAI offerings on the market, a team inside Nutanix determined it was best to develop an app for their own use.
“I had all these vendors saying, ‘Look, we've had plenty of companies try it themselves,” said Chad Singleton, vice president of Support Readiness at Nutanix. “‘Go for it. We know you'll fail. When you do, give us a call and we'll be here waiting.’”
That call was never made because the homegrown SupportGPT hit the mark. And it happened quickly. The idea came from a Nutanix hackathon in April 2023, when employees shared ideas for new business applications. A proof of concept was approved and work began in May. It was production-ready in April 2024 and has been in use ever since.
“What once took hours to research now takes less than a minute,” said Singleton, recalling feedback from one of the team’s System Reliability Engineers (SREs) after using SupportGPT to help a customer.
SRE's used to search a Nutanix database using keywords from questions or customer requests. Once they found a relevant article, they would read and find the information the customer needed. Now, they just search in SupportGPT using natural language processing (NLP), generating the answer in a few seconds.
Strategic Place to Start
Getting from the idea to the implementation of SupportGPT took tightly coordinated teamwork. Kathy Chou, senior vice president of SaaS Engineering at Nutanix, directed the project. Her team uses data-driven innovation to deliver excellent experiences for Nutanix customers, partners, and employees. Chou’s team joined forces with business sponsor Singleton and Debojyoti “Debo” Dutta, vice president of Engineering and AI for Nutanix, who provided guidance and data scientists.
The Support team wanted something they could roll out soon, but there was trepidation.
“I wasn’t going to pilot or test tools in a sandbox until I was certain it would get us what we needed,” Singleton told The Forecast.
Using chatbots for customer support has become common practice, but Singleton’s team wasn’t satisfied with the chatbots they saw being used by many financial or consumer services. They help customers address complex IT system issues that must be correctly diagnosed and fixed with clear and concise technical details.
“We needed a world-class tool that gave our support engineers accurate information, so customer requests can be handled quickly without escalation,” Singleton said.
Singleton and others feverishly studied the benefits of GPT and the risks of giving customers inaccurate GPT-created advice. This led them to a significant next step: requiring human support engineers to validate and present accurate information to customers.
“A human-validated support system could help ensure the accuracy of GPT responses,” Singleton said, recalling what motivated them to move ahead. Accuracy is essential for Nutanix customer support team, which has averaged above a 90 Net Promoter Score for over eight consecutive years.
“Our service is a business differentiator,” Singleton said. But as Nutanix grows, he said the team must innovate without skipping a beat.
Creating Business Value From GenAI
Business interest in GenAI applications skyrocketed after OpenAI, Meta and other digital technology companies began releasing their large language model capabilities to the world. Exuberance has pushed GenAI to the top of the hype curve in 2023, but a year later many enterprises have moved beyond the hype. A recent report by Google called The ROI of Gen AI shows 74% of enterprises using GenAI report getting ROI within the first year. More than 8 out of 10 of those using GenAI reported a 6% or more increase in revenue.
The report also showed that 84% of organizations move a GenAI use case from idea to production in less than six months. The Google report states more business use cases and ways to get started.
Dutta said his engineers learned a lot developing Nutanix’s GPT-in-a-Box, a turnkey solution that makes enterprise IT infrastructure AI-ready. That team was working on the second version of GPT-in-a-Box while they collaborated with Chou’s team to build the new SupportGPT application. He said AI application development has changed dramatically in the past year.
“This year, everybody's going gangbusters because it’s easy now to set up RAG (retrieval-augmented generation) and implement many use cases, to drive business needs,” Dutta said.
“Today, enterprises are hosting hackathons within their company to find new use cases that they can quickly put into production.”
He’s a firm believer in the benefits of AI.
“Once this technology trickles down, it's gonna have a dramatic effect on our enterprises and society,” Dutta told The Forecast. “There'll be at least 15 to 20% efficiency in every walk of life using these capabilities.”
Nutanix’s experience building its own GenAI app to improve its own business underscores what many companies are working through as they search for practical and beneficial ways to leverage AI. Short and long-term economics play a critical role in the decision-making process.
“Software can be a multiplier, boosting productivity without building up operational expenses,” Dutta said.
But when it comes to hiring a third-party vendor to run sophisticated AI applications, it can incur ongoing cost increases for running more GPUs and periodic application updates.
“Instead of getting multipliers of efficiency, you get multipliers of op-ex,” he said.
Build vs. Buy Bake-Off
Chou said her team knew they could build SupportGPT themselves but wanted to evaluate the best of what was available on the market.
“Our open source model approach ensures we have best-of-breed tools and flexibility to build a GenAI platform that can be leveraged across Nutanix,” said Chou. She saw SupportGPT as a momentous step toward creating GenAI apps for other aspects of Nutanix’s business.
Chou said SaaS Engineering Vice President Manoj Thirutheri used Retrieval Augmented Generation or RAG, which includes Large Language Models (LLM), with specific datasets like a company’s knowledge bases (KBs). RAG allowed them to fine-tune the final results to focus only on accurate, useful data and filter out data that might fuel inaccuracies and hallucinations.
“The team's relentless pursuit of excellence paid off,” she said. “The current production run of SupportGPT achieved an impressive 80% accuracy rating.”
She said this allowed them to improve accuracy in the next iterations which included AI experimentation and continuous SRE feedback. A well-curated gold standard dataset enabled her team to achieve high accuracy and reliability.
“We worked together hand in hand, collaborating across the AI team (Dutta's team), SaaS Engineering (Chou's team) and the customer service team (Singleton's team),” Chou said. “This ensured the specific needs and goals were clearly communicated along the way.”
Dutta said all the knowledge gained from building SupportGPT would benefit Nutanix.
“Whatever we built would give us a big headstart on any future AI development project,” he said.
Ongoing SupportGPT application efforts feed into SaaS Engineering’s learning management system (LMS), which integrates all aspects of the company’s data-driven operations.
SupportGPT Improves Customer Service
Nutanix’s SREs working customer support cases often sift through massive amounts of technical content, KBs packed with solutions, technical documentation, and summaries of previous support engagements. An SRE on a support case might burn hours a day manually reading and mentally synthesizing page after page of technical content.
Conventional automation can deflect some customer support tickets to a self-help portal where customers can figure out things independently. GenAI lets SREs ask highly specific questions and receive easy-to-understand summaries of information buried in KBs and other data sources. It also can empower bots that automatically deflect cases, lightening the load on SREs.
“When I started to see the first results, I knew it was a game changer for us,” Singleton said. “We just needed to do it right.”
Doing it “right” meant infusing the Three H’s: be honest, harmless and helpful. These, along with human validation of data, are some of the key aspects that went into creating SupportGPT.
Singleton anticipates using GenAI across a customer’s full support journey.
“Case summarization can help us draft a KB before closing a case,” Singleton said. “It can provide the first iteration or first draft of the content, so support engineers don't need to start from scratch.”
GenAI frees SREs to focus more on complex and challenging issues.
“We're removing the mundane so SREs can focus on the hard problems to solve,” Singleton added. “That's what challenges them. That's what they like doing,”
The SupportGPT tool gets 500 to 700 queries a day, helping SREs take work that used to take hours and finish it in minutes.
“The feedback is great,” Singleton said. “The accuracy remains high.”
“Our goal is to help our SREs run faster with the same effort,” Singleton said. “We don't want to remove the human aspect that has seen us achieve our 90 NPS score to a bot or any other AI system.”
Dutta said the effects of AI are rippling through Nutanix and that taking on projects like SupportGPT contributes to the overall success of enterprise AI.
“We demonstrated that even with limited resources, it is possible to build powerful AI solutions that deliver real business value,” Dutta said. “It has a multiplier effect.”
Ken Kaplan is Editor in Chief for The Forecast by Nutanix. Find him on X @kenekaplan.
Tom Mangan contributed to this story. He is a veteran B2B technology writer and editor, specializing in cloud computing and digital transformation. Contact him on his website or LinkedIn.
© 2024 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
