00:00 Speaker A
I want to jump right in from where Amazon is considering the biggest uses for AI. Is it in the agent now? Is it about accessing different models or developing different models? Where are you guys seeing that?
00:15 Speaker B
Access is increasing overall. People choose AI and Amazon for their AI because over the past 20 years they have come to trust Amazon for security, reliability, and safety. But in this world of agent AI, with Amazon Bedrock, you have the choice of many models to use, and you can choose the one you need for your task. But then you can build agents on top of it using Amazon Bedrock Agent Core. And with products like Amazon Quick and Qro, you can take those models and agents and start applying them. In the case of Qro, it’s to build software; in the case of Amazon Quick, it’s to help your team build agents, or to make it easier to process email. So across the board, we have customers who come to train models on AWS and do inference on AWS. There are people who build agents on AWS, but we provide our own agents and our customers use them. Overall, especially over the past six to nine months, we’ve seen an explosion in the types of applications that people are building using these tools.
01:21 Speaker A
Yeah, we talked, I think we’re royals in general, I talked about how I think 2025 is supposed to be the year of the agent, right? But it feels more like 2026 than 2025. You know, we may have seen a little bit of the beginning in 2025, but now we’re actually seeing them in the real world. Oh, we were just talking on the radio. As you know, I have several settings for different alerts. Get it from your email, visit various websites, and submit your information. How do you use agents internally at Amazon? Do you think you use agents a lot? Sure, you think they’re more helpful, but do you think they give you time away from work to focus on other things? Do you think they’re giving you a little more work just because you’re doing more? How are they used?
02:16 Speaker B
Therefore, the ability of agents has definitely improved in the last year. This is because the underlying technology, the models available to us, have increased the capabilities of our agents. That means Amazon is saving teams 4,500 years of work on projects across engineering, finance, and other departments. Another team completed an 18-month project in 76 days. Well, there’s one more thing. Our legal team can complete your tax audit in minutes instead of 15 hours. And the most relevant for many of us is Amazon’s “Add to Delivery” button, which I use every week. Well, it was shipped two months earlier than scheduled. Answering your specific questions will show you the scale of what people can do and all-important time-to-delivery, because you can accomplish more. Well, there are multiple areas where the use of agents can effectively impact Amazon, both in the way they work internally and in the products they ship.
03:32 Speaker A
When it comes to different AI companies, we’re working with Anthropic and AWS has a big relationship with them, but today we saw Microsoft change their contract with OpenAI. I think this opens up AWS to potentially work with OpenAI on another level. Because you’re already trading, but this seems to open up even more possibilities.
03:59 Speaker B
The deal, the partnership, that we announced with OpenAI, as you know, is what we’re doing right now, where we’re co-creating these stateful runtime environments so that our customers can build and run agents on Amazon Bedrock. And obviously they’re also helping us use AWS capacity to service these runtime environments. For example, where people are running these agents. Oh, and Amazon is also the exclusive third-party provider to explore our frontier: the agent environment. And overall as well, they use our chips. We have an agreement to service agents to our customers and to use this runtime environment that we’re building together with Open AI to enable customers to build their use of agents. So we already have a pretty deep relationship with them and we just announced it. So it will only get better over time.
05:07 Speaker A
How do you think agents are becoming more productive on a daily basis? I mean, I’m literally just starting to get my feet wet. Hmm, I wonder how that can improve my productivity over time. What are customers saying about that?
05:27 Speaker B
So, as a great example, and this is at an organizational level, a company like Bristol-Myers Squibb had over 10,000 compounds that would take forever, months, weeks, months to analyze. Now you can do it within a day. This is meaningful when analyzing a large number of compounds. Similarly, Genentech built an agent on top of Amazon Bedrock using Amazon Bedrock Agent. This allows tasks such as double-checking biomarkers. This had to be done very manually. Now we can automate that entire process. It’s about how it impacts business. For example, Visa is building to protect its entire environment using agents on AWS. On a day-to-day basis, this can range from processing emails faster. You and I talked about it earlier. Well, from simple things like there are people who have the tax thing that I talked about, who are doing research. Well, maybe you’re using Amazon Quick to do M&A research or understand tax policy. With an AI agent, you can do that very quickly, and in the meantime, you can focus on other things or work on other things. What I want to tell people is something that I think is very important, but can be a little boring or repetitive. Using AI agents for this is a great way to start. And many of our customers are starting there as well.
06:57 Speaker A
How have the hallucinatory aspects of all this changed over time?
07:03 Speaker B
So the way to think about it is multidimensional. One is that the power of the model has improved. However, the higher the quality of the data, the better the results. What I’m saying is that AI agents are much more than chatbots. Some people still think of agents and AI as just chatbots, but that’s not the case. With an AI agent, you can give it input to control and give it the specific results you want it to produce. We call this steering because it directs the agent toward the world in which we want to see the outcome. So if you’re a domain expert, you can provide the right set of information, you know where your data is, and your agents can accomplish a lot with high-quality data and well-defined outcomes, so it really helps. But I think Amazon has also been thinking about things like automatic inference, which is a mathematical way of proving something is true. So Amazon Bedrock offers, um, Amazon Bedrock. Well, I think it’s called validation checking using automatic inference. And let’s say it’s a company that has policies and some kind of compliance. You can create these checks and ensure that your results always comply with those checks. So we’re delivering this both from a tool perspective and the fact that the way people use these agents and the capabilities are much better. So you’ll see much less.
