00:00 Speaker A
I kind of just want to jump off where uh Amazon is seeing kind of the the biggest uses of AI. Is it in agentic right now? Is it in accessing different models, developing different models? Where where are you guys seeing that?
00:15 Speaker B
We are seeing access across the board. People uh people choose AI uh Amazon for AI because over the last 20 years they’ve come to trust Amazon for security, reliability and safety. But in this agentic AI world, uh with Amazon Bedrock, they have the choice of using many many models that are they can pick the model that they need for their task. But and then they can build agents on it using Amazon Bedrock Agent Core. And then with uh products like Amazon Quick and Qro, they can then take those models and those agents and start applying them. Uh with Qro, it’s for building software, with Amazon Quick, it’s for maybe helping you build an agent for your team or uh making it easy for you to get your handle your email. So across the board, we have customers who come and train their uh models on on AWS, they do inferencing on AWS. There are people who build agents on AWS and then we provide agentic agents ourselves, which then our customers use. So across the board, we have uh you know, especially over the last six to nine months, there’s been an explosion in the kinds of applications people are building with those tools.
01:21 Speaker A
Yeah we we talked I think uh we being generally the Royal we about how uh I think 2025 is supposed to be the year of the agent, right? But it feels less so that it was 2025 and more like it’s 2026. You know, maybe 2025 we saw the starts a little bit, but now we’re actually seeing them in the real real world. I uh we were just talking off air. I have a few set up for different alerts that you know, pull from my email, go to different websites, send send me the info. I guess internally at at Amazon, how are you guys using agents and do you find them to be uh a lot I mean, obviously you would think that they’re more helpful, but do you see them as allowing you to take time from away from work to focus on other things? Do you see them giving you a little more work just because you’re doing more things? How how are they kind of being used?
02:16 Speaker B
So, over the last year, the agents have definitely gone from improved in capability because of the the underlying technology, the models that are available to us, the capabilities of these agents have gone up. So at Amazon, you across engineering, across finance, other departments, you see things like a team uh saving 4,500 years of work on a project. You have another team that had an 18-month project that they finished in 76 days. Uh there was another, our legal team is able to do tax research in minutes instead of 15 hours. And, you know, most relevant to many of us, the Amazon Add to delivery button, I use that every week. Um that was shipped two months ahead of schedule. So to answer your specific question, it manifests itself in the scale of things that people can do because you can accomplish a lot more and also the time to delivery, which matters a lot. Uh so there’s multiple areas in which using agents effectively impacts Amazon both from how we do work ourselves internally and the products that we ship.
03:32 Speaker A
When when it comes to the different AI companies, you know, you work with with Anthropic, you know, AWS has a huge relationship with them, but now we saw today Microsoft changing its agreement with OpenAI. I guess does that now allow AWS to potentially go in and and work with OpenAI at a different level because you guys already have uh uh deals, but this seems to open it up even further.
03:59 Speaker B
So the deal we announced with OpenAI, the partnership, it’s you know, this the one that we have in place right now, we are co-creating these stateful runtime environments, which allow our customers to build agents and run them on Amazon Bedrock. We are also working with them on obviously they are uh using our capacity from AWS to serve these uh runtime environments. For example, where people are doing these agents. Uh Amazon’s also the uh third party provider for the exclusive third party provider for opening our frontier, which is their agentic environment. And so across the board, so they’re using our chips. They we have a agreement on on serving agents to customers as well as allowing our customers to build agents use uh using this runtime environment which we are co-building with Open AI. So we already have a fairly deep relationship with them and it’s only going to we’ve just announced it. So it’s only going to improve over time.
05:07 Speaker A
How do you see agents improving, I guess everybody’s everyday productivity, right? I mean, I’ve like I I’ve literally just started stepping my toes in. Uh I I wonder though how that that can improve my productivity over time. What are your customers saying as far as that goes?
05:27 Speaker B
So, I mean a great example and this is at a corporate organizational level is somebody like Bristol Myers Squibb where they had 10,000 plus compounds and doing the analysis used to take them forever, months, weeks, months. Now they can do it in less than a day, right? and it’s meaningful when you’re analyzing that many compounds. Similarly, uh Genentech built an agent on Amazon Bedrock with Amazon Bedrock agents, which allows them to do things like uh look at biomarkers again, which they had to do very manually. Now they can automate that entire process. So that’s on the impact it can have for a business. Uh Visa, for example, is using agents uh on AWS to build for to secure their entire environment. On a day-to-day basis, they can range from going through email faster, you and I were talking about that earlier. uh to simple things like you have the the tax one that I talked about, you have somebody who’s doing research. Uh you may be using Amazon Quick for that because you want to do research on an M&A thing or understand the tax policy. You can now use AI agents to do that for you very very quickly, while you can focus on other things uh or go on to other work. The what I like to tell people is anything that feels very important but can be a little tedious or repetitive, use an AI agent for that’s a great way to start. And that’s where we see a lot of our customers starting as well.
06:57 Speaker A
How how is the uh I guess hallucination aspect of all of this changed over time?
07:03 Speaker B
So the way to think about it, multiple dimensions. One is models have gotten better at it. But the better the quality of your data, the better the results you will get. The way the way I like to say it is AI agents are much more than chatbots. People still think of agents and AI as just a chatbot, but that’s not the case. With an AI agent, you can give it specific outcomes that you want to generate and give it the inputs that you control. We call it steering because you are steering the agent towards the world that you want to see results in. So if you’re a domain expert, it actually helps you uh because you can give it the right set of information, you know where your data is, with high quality data and well defined outcomes, agents are able to accomplish a lot. But at Amazon, we think we’ve also thought about things like automated reasoning, which is a mathematical way of proving that something is correct. And so in Amazon Bedrock, we provide uh Amazon Bedrock, uh I think it’s called verification checks with automated reasoning. And those, let’s say you’re a company with a policy, some kind of compliance, you can create those checks and make sure that the results are always in compliance with those checks. So we’re providing it both from the tooling perspective and just the fact that the way people can use these agents and the capabilities are getting much much better. So you see much less.