Abstract
- Change management is important. Effective AI integration relies on strategic change management and identifying clear problems.
- Organizing your data is important. The success of generative AI, especially in regulated industries, relies heavily on well-structured and organized data.
- Emerging AI applications. Innovative AI tools are transforming traditional tasks, promising significant gains in productivity and ROI.
Nearly two years after ChatGPT's debut, AI hype is giving way to reality. Enterprises are eager to build generative AI, but it's proving to be a challenge to pull off: AI models are expensive, data is riddled with challenges, and change management is proving not to be so straightforward. That's why only 21% of enterprises surveyed by Gartner earlier this year had deployed generative AI in production, while the rest were in the “piloting” or “exploring” stages of the technology, according to Big Technology data.
AI optimism fuels billion-dollar race
And yet the optimism around this technology is unprecedented. Every thriving company is considering how to integrate generative AI into their internal operations and external products. They're spending billions of dollars with big tech companies and consulting firms looking for solutions. And they believe all these experiments will eventually pay off. And they hope so, because the economic future of the current AI boom depends on it.
Related article: VC predicts 'disruption' as generative AI eats the web
Inside Amazon: The incremental advances of AI
I spent a fair bit of time last week talking to Amazon’s AI team and its partners about these on-the-ground realities, and I got my best understanding yet of what’s going on. I was surprised by the cautious tone of almost everyone I spoke to. “It’s going to feel a lot more incremental than we’re probably used to,” Matt Wood, vice president of AI products at Amazon Web Services, told me, arguing that it will add up over time. I also learned about some surprising products that broadened my view of the state of the art. Below is a breakdown of the major roadblocks and what surprised me on the product side:
Related article: 7 Technology Predictions for 2024
Change management
It's estimated that over 7 million people pay OpenAI $20 per month for premium ChatGPT, but at least in the short term, generative AI will be most valuable to enterprises. For enterprises, the use cases are clearer, and so are the returns. But for enterprises to successfully adopt AI, employees need to embrace new internal tools, change workflows, and increase automation. And executives need to apply technology to solve problems, not adopt technology for implementation's sake.
“A year ago, many customers were asking us, 'Tell me more about generative AI, how can we use it?' and setting aside budgets without any idea of what they would spend it on,” Luba Borno, vice president of Worldwide Channels and Alliances at Amazon Web Services, told me. “Without a clear 'why,' it becomes very difficult to move forward with the 'what' and the 'how.'”
Valerie Henderson, president of AWS consulting partner Caylent, also spoke to the severity of the change management challenge. “You can't underestimate this,” she told me. “I was having dinner with a client last night, and we were talking about this, and he said his biggest fear is that they'll build this and have one illusion – that they'll never get the right output, and people will 'quietly' stop using it.”
Last year, when countless press releases claimed that AI would replace departments or revolutionize companies, I was skeptical of the importance of the change management issue. But now, some people are starting to understand it. For example, Klarna CEO Sebastian Siemiatkowski recently convinced me that a significant part of customer service could be entrusted to large language models (LLMs). More on this later. Successful adoption is rare today, but it could become more common as the technology and organizational readiness improves.
Related article: Apple + OpenAI Math: A week in Silicon Valley
Model Cost and ROI
While AI applications look great in the prototype stage, they can become expensive to run when pushed to 100% of users. This is another issue that holds back broader deployment of generative AI, but it doesn't seem like it will be a long-term problem. Already, some companies are finding use cases with clear benefits that justify the expenditure. Borno says one energy company built a generative AI tool to assess tax needs in different countries and has already saved hundreds of millions of dollars from the output.
If AI model providers can lower costs and bring them closer to zero, wider use will become possible. “Everyone should be prepared for $0 intelligence,” said Logan Kilpatrick, a current Googler and former OpenAI employee. “It's going to happen sooner than you think.”
Data issues
Generative AI works best with well-organized data, which is why some of the companies benefiting most today are those in highly regulated industries with large amounts of well-structured data, such as financial services, healthcare, and life sciences.
Amazon's Matt Wood gave the example of a life insurance company with a pile of policies for 90-year-olds who were due to pay out soon. “They were scanned at some point, but nobody had read them,” Wood said. “So they were able to use generative AI to piece together that risk and get a more complete understanding of it.”
This application of generative AI is admittedly mundane compared to the revolutionary, world-changing capabilities promised by some in the tech industry. But the productivity gains from offloading this work to machines could be enormous. “We really want to automate a lot of the mundane tasks,” Wood says. “We want to be able to take the mundane tasks that are maybe considered cost centers in some organizations and turn them into things that drive invention and growth.”
New Use Cases
The most amazing product I heard about was a generative AI tool they built to help automotive companies diagnose car issues using words, images, and sounds. The AI solution they built ingested documents, including images, from the car's user manual and even trained it on the sounds the car makes when it needs service.
The product “listens” to the car and can then recommend repair options to the service center. “We have this system running in three dealerships,” said Alan Chhabra, executive vice president of worldwide partners and international sales at MongoDB, which developed the product, but declined to name the company.
As modern AI models emerge with greater intelligence, multimodality, and better economics, we will likely see more applications like this. And that will likely help the percentage of companies with generative AI solutions in production rise to 21%, even if that percentage goes on for a bit longer than the story suggests. “The more customer cases we have of it working in production, the more likely customers are going to take the risk, follow through, and get an ROI,” Chhabra says.
