In the months before AI startup OpenAI debuted its artificial intelligence chatbot ChatGPT, technologists at Amazon’s cloud unit, Amazon Web Services (AWS), were starting to field more inquiries from customers who were curious about exploring traditional AI and machine learning.
But after ChatGPT’s launch in November 2022, everything changed. The exploding customer interest in generative AI led to a $100 million commitment by Amazon to fund the AWS Generative AI Innovation Center, a global team made up of scientists, engineers, and business strategists that launched in June 2023 with a mandate to help enterprises deploy generative AI tools that would drive greater productivity for workers, improve customer experiences, and change their software product development processes. AWS doubled its investment in the Innovation Center two years later.
Over the course of its two and a half year existence, the Innovation Center has worked with well over 1,000 customers, including Formula 1, Nasdaq, Ryanair, and S&P Global. AWS says that more than 65% of the projects the Innovation Center has worked on this year have gone into production. It’s a success rate far above studies that have shown that a vast majority of generative AI pilots fail.
“When I say ‘production,’ this means the solution is live,” says Sri Elaprolu, a 13-year Amazon veteran who now serves as director of the Innovation Center. “It’s driving real business outcomes, and it’s driving value to the customer.”
Elaprolu says each project kicks off with an hours-long “discovery workshop” where AWS brings together a client’s data stewards, business leaders, and technologists. He says there have been instances where these three groups hadn’t yet all met to debate the new use cases they’d like to explore.
“We need to get that cross-org leadership alignment on the customer side to understand, are they all aligned on the same problem that we want to solve?” Says Elaprolu. “Because often what you run into is your business wants something done, but the technology is not ready, or the technology wants to experiment, but the business doesn’t care.”
Once everyone is aligned, the data gets a fresh look to make sure that the quality, volume, and access are all in order. Then, Elaprolu says, critical discussions are needed to outline expectations for the return on investment and the time horizon for when results can be achieved. “Yes, we want to be aggressive, and yes, we want to push it a little further,” says Elaprolu. “But what is realistic?”
Bringing employees along for the ride
The “discipline phase” comes next. That means working through any necessary change management processes to ensure that the company’s employees buy into using the new tools, or if applied externally, to gauge customer usage.
“You could take something into production, but the adoption of customers or users may not be as high as you wanted, which means you just lost all of the ROI that you were expecting to gain,” says Elaprolu.
GoDaddy, which helps businesses and individual consumers set up their domain registration, has worked with the Innovation Center for two years. One project the pair have put into production involves testing various products like Anthropic’s Claude and Meta’s Llama to determine which LLM would be best at predicting sales for GoDaddy’s customers, who tend to operate small businesses with lower sales volumes. AI can help them better predict demand, says Jing Xi, GoDaddy’s vice president of applied AI and ML.
Another project, still in pilot mode, involves adding AI capabilities to the domain-name search feature, which enables GoDaddy to serve up a potential unique web address with an image icon that may be relevant to the customer’s request. Xi says GoDaddy wants to be more certain about the user interface design and also more measured about a full deployment because of the potential impact on revenue. But she says the concept is a bolder bet that GoDaddy feels comfortable embracing with a second set of eyes from the AWS team.
“For the Innovation Center, usually the project that we wanted to try is a little bit more risky,” says Xi.
When the Innovation Center launched, it could take six to eight weeks to get a generative AI project into production, though time horizons can vary beyond that range. As technology has advanced and the team has gotten more experience working with enterprises, projects can now be deployed in as little as 45 days. The group has also expanded its focus as technology has advanced, now including more work on agentic AI and physical AI.
Off-the-shelf versus customized models
In 2024, Elaprolu set up a team within the Innovation Center to exclusively focus on model customization. Off-the-shelf large language models weren’t meeting the needs of certain customers in sectors like health care or financial services that were deeper in their AI journey and needed models that were customized to their specific industry requirements and unique data.
“Broadly speaking, model customization is an area where there’s a ton of activity going on,” says Elaprolu. “We expect this to only get much more intense as more enterprises go towards the core of their business.”
Another client is Cox Automotive, an auto-focused software provider whose brands include Autotrader and Kelley Blue Book. Cox Automotive kicked off its work with AWS in 2018 for a migration to the cloud, and with the exception of a few acquisitions, nearly all of the company’s tech stack is in the AWS cloud.
Part of the more recent work between the pair has focused on agentic AI. Marianne McPeak-Johnson, chief product officer at Cox Automotive, says she has more than 500 data scientists on her team but carved out a smaller group that was mandated to focus on accelerating agentic AI use cases. There were more than 57 ideas considered, and 20 have moved into full production.
This summer, AWS went to Cox Automotive’s office, where McPeak-Johnson said a group of about 100 employees across both businesses divided into six teams to explore new agentic AI tools. They worked through questions about model performance; orchestration that can potentially occur across multiple agents; as well as how the technology should be monitored and at what level of reliability.
“That partnership has been incredible because we’ve been able to take those concepts and start filling out our methodology to go to market,” says McPeak-Johnson. “All six of those concepts are now in pilot with customers.”
