Months before AI startup OpenAI debuted its artificial intelligence chatbot ChatGPT, engineers at Amazon Web Services (AWS), Amazon's cloud division, began fielding inquiries from customers interested in exploring traditional AI and machine learning.
But everything changed with the release of ChatGPT in November 2022. Following an explosion in customer interest in generative AI, Amazon has committed $100 million in funding to the AWS Generative AI Innovation Center. The AWS Generative AI Innovation Center is a global team of scientists, engineers, and business strategists that launched in June 2023 with a mission to help companies deploy generative AI tools that increase employee productivity, improve customer experiences, and change software product development processes. Two years later, AWS doubled its investment in innovation centers.
Over its two and a half year existence, the Innovation Center has worked with well over 1,000 customers including F1, Nasdaq, Ryanair and S&P Global. AWS says more than 65% of the projects its Innovation Center worked on this year have gone into production. This is a much higher success rate than studies showing that the vast majority of generative AI pilots fail.
“When I say ‘production,’ I mean the solution is in production,” says Sri Elaprol, who worked at Amazon for 13 years and is now director of the Innovation Center. “It delivers real business outcomes and delivers value to customers.”
Elaprolu said each project begins with a multi-hour “discovery workshop” where AWS brings together the customer's data managers, business leaders and technical experts. He said there have been instances where all three groups were not yet together and discussed new use cases they wanted to explore.
“We need to understand the alignment of leadership across the organization on the customer side. Are we all aligned around the same problem that we want to solve?” says Elaprole. “Because a lot of times you run into situations where companies want to do something, but the technology isn't ready, or the technology wants to experiment, but the companies are apathetic.”
Once everyone is aligned, your data will have a new look, ensuring that the quality, quantity, and access are all in place. Next, says Elaprole, a significant discussion is needed to outline the expected return on investment and the time frame by which results will be achieved. “Yes, we want to be proactive, and yes, we want to go a little further forward,” Elaprole says. “But what is realistic?”
bring employees along for the ride
Next, the “training phase” begins. This means you need to run all the necessary change management processes to ensure your company's employees buy-in to the new tool, or measure customer usage if it's being applied externally.
“You can put something into production, but the customer or user adoption rate may not be as high as you hoped, meaning you lose out on all the ROI you were hoping to get,” Elaprolu says.
GoDaddy, which helps businesses and individual consumers set up domain registrations, has been working with the Innovation Center for two years. One of the projects the pair has put into production is testing different products, such as Anthropic's Claude and Meta's Llama, to determine which LLM is best at predicting sales for GoDaddy's customers, who tend to run small and medium-sized businesses with low sales volume. Jing Xi, vice president of applied AI and ML at GoDaddy, said AI can help predict demand more accurately.
Another project still in pilot mode involves adding AI capabilities to the domain name search feature, which will allow GoDaddy to provide potential unique web addresses with image icons that may be relevant to a customer's request. Xi said GoDaddy wants to be more certain about the design of its user interface and more cautious about its full implementation because of the potential impact on its bottom line. But she says the concept is a bolder bet that GoDaddy is comfortable accepting with a second set of eyes from the AWS team.
“For the Innovation Center, the projects we wanted to try are usually a little more risky,” Xi said.
When the Innovation Center launched, it could take six to eight weeks to move a generative AI project into production, but the time frame can extend beyond that. As technology advances and the team gains experience working with companies, projects can now be deployed in as little as 45 days. The group has expanded its focus as technology advances, and now includes increased work on agent AI and physical AI.
Ready-made models and customized models
In 2024, Elaprolu established a team within its innovation center dedicated to model customization. Off-the-shelf, large-scale language models did not meet the needs of specific customers in areas such as healthcare and financial services, which are deep in AI efforts and require models customized to specific industry requirements and unique data.
“Broadly speaking, model customization is an area where there is a lot of activity,” says Elaprolu. “We expect this situation to further intensify as more companies move towards the core of their business.”
Another customer is Cox Automotive, an automation-focused software provider with brands such as Autotrader and Kelley Blue Book. Cox Automotive began working with AWS to move to the cloud in 2018, and with the exception of a few acquisitions, nearly all of its technology stack is in the AWS Cloud.
Some of their recent research has focused on agent AI. Marianne McPeek-Johnson, chief product officer at Cox Automotive, says the team has more than 500 data scientists, but the company has carved out a small group tasked with focusing on accelerating agent-based AI use cases. More than 57 ideas were considered and 20 ideas moved into full-scale production.
This summer, AWS visited Cox Automotive's offices, where about 100 employees from both companies divided into six teams to explore new agent AI tools, McPeak-Johnson said. They addressed questions about model performance. Orchestration that can potentially occur across multiple agents. The same goes for how the technology is monitored and what level of reliability is ensured.
“This partnership has been incredible because we were able to take these concepts and perfect the methodology to bring them to market,” says McPeek-Johnson. “All six of these concepts are currently being piloted with customers.”
