AWS re:Invent 2025 - Agentic AI's new generation of Industry and Line of Business solutions (PEX202)

🦄 Making great presentations more accessible.
This project enhances multilingual accessibility and discoverability while preserving the original content. Detailed transcriptions and keyframes capture the nuances and technical insights that convey the full value of each session.

Note: A comprehensive list of re:Invent 2025 transcribed articles is available in this Spreadsheet!

Overview

📖 AWS re:Invent 2025 - Agentic AI's new generation of Industry and Line of Business solutions (PEX202)

In this video, Michael Schmidt and Ryan Thomas from AWS discuss agentic AI's transformation across industries. They explain how AI has evolved from reactive generative responses to autonomous multi-agent systems capable of complex decision-making. Using a circus industry analogy, they emphasize focusing on customer value over technology hype. Key statistics reveal that while 80% of customers have deployed AI, only 25% see benefits, but those successful implementations achieve 45% cost savings and 60% revenue growth. The session covers industry-specific use cases including construction scheduling with weather-dependent concrete delivery, media entertainment production planning, predictive maintenance in automotive manufacturing, distributed energy resource management, and real-time financial fraud detection. Healthcare examples include Philips' five-second medical image analysis and Cure.AI's lung cancer detection using 25+ million scans. They stress starting with clear business problems, experimenting early, and leveraging AWS's agent marketplace and partner ecosystem for success.

; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.

Main Part

Welcome to Las Vegas: Setting the Stage for Agentic AI in Industry

Good morning everybody. Welcome to Las Vegas. It's wonderful to see you all, and thank you for making time to spend with us today. Welcome to Las Vegas. I hope you've had a good morning so far. It's a little chilly outside, and I'm sure you're wondering why I'm dressed like this. More on that to come. My name is Michael Schmidt. I look after industry partnerships, and together with Ryan Thomas, we're looking forward to spending some time talking about all things agentic AI. We're going to specifically focus on industry and what people are really doing with it. I think you're going to find some really interesting stories that we're going to talk through today.

We represent the partner organization at AWS, and specifically our team focuses on the line of business and what customers are really doing in industry. Our team looks ahead six, twelve, eighteen months to understand what transitions are happening, what's going on in the industry, and what customers are doing. Right now, we're talking a lot about AI and what we're seeing in the market, as well as what customers are asking for in the future. Today we're going to talk through some really specific use cases and stories that I'm hoping you'll be able to learn from.

You don't have to look far for talk about AI. It's in the popular media. Our families are talking about it. The newspapers are talking about it. For those of you who live in the United States, you've probably discussed it over Thanksgiving in the last couple of weeks. It's making its way into popular media. There are things like Velvet Sundown, AI-generated music on Spotify that's become very popular. People have heard about Tilly Norwood, which was launched at the Zurich Film Festival this year—AI-generated talent. It is a controversial topic, an interesting topic, and a fascinating topic. But never before in my lifetime has a technology transition been so public and so prominent in the popular narrative. It's pretty much everywhere, whether it's a doomsday prediction, a model update, or a meme. It is everywhere we look right now.

It has been fascinating over the last couple of years, really in the last twelve months since re:Invent last year, to see the rate of change and the pace of innovation. Initially with generative AI, it was more reactive. You give it a prompt, you ask it a question, and it gives you a response, whether that's a line of code, a piece of text, or a picture. It's reactive, it's stateless, and yet powerful. Then we saw the emergence of early AI agents, which leveraged reasoning frameworks to make more sophisticated outcomes and produce more sophisticated responses.

The early agents were still primarily used for simple, well-defined tasks. But as they've evolved over the last year, we've seen a transition into fully autonomous multi-agent systems that can seamlessly execute multi-step processes and solve multi-step problems. They're capable of tackling more complex problems using real-world use cases, and we're going to talk about a number of those as we go through today. This has primarily been driven by a couple of fundamental shifts in our industry. The content of this presentation has evolved just over the last six weeks. I can't remember a time in my career in this industry where the topic and content of a presentation can change so much in the last six weeks before re:Invent, just because of the rapid rate of change that we're seeing in the industry. It's remarkable.

The advancements in the models are astounding. In the last twelve months and even in the last six months, we've gone from foundational models that provide reactive responses, quality responses, and insightful responses, to proactive models that can reason, reflect, observe, and act. The difference is this: when you ask a simple foundational model from twelve months ago to find a date and time to go hiking on Mount Fuji, it could tell you that the weather's going to be good on Wednesday. But more proactive agentic models now will tell you that the weather's going to be good on Wednesday, they've noticed that your fitness has been strong the last couple of weeks, so this route might be appropriate for you, you may want to consider these suppliers, the weather window is short, so you may want to consider turning around by twelve o'clock. These are far more sophisticated, compound analyses and proactive responses that are unprecedented.

Today, we're going to talk through a number of things and share a number of different stories. I think what you're going to like most about this is that it's going to be very practical. We're going to touch on so many customer stories that you're going to find insightful. To set the scene for us, a couple of people have asked me to make them a café latte already this morning. So Ryan, can you please tell me why am I dressed like this? It's the question we all want to know.

The Circus Story: A Blue Ocean Strategy for the AI Era

As my boss Michael set the stage with all of the technical jargon and the evolution of FM's, I get to be the one to tell you a story. We're in Las Vegas. Welcome to Ryan and Michael's circus. I want to tell you a story that probably many of you know, and it's going to set the stage for what we're seeing, some of the trends we're seeing in AI for what differentiates our partners and customers from being successful and less successful, from growing their top line and reducing costs to just incurring cumbersome, more and more technology, more and more complexity that's not driving value.

In the 1970s, the circus was a booming industry. In fact, it was one of the top family live entertainment places that you would take your family. Over time, as we got into the 1980s, we started to see sports teams and concerts create a more family-oriented experience, and the circus business started to decline. Enter a whole new business strategy for the circuses. Ringling Brothers and all of the others went out and hired the top talent. They said that what's going to differentiate us is Bozo and this clown and that clown. We started to enter the golden age of clowns.

Clowns were demanding top salaries. They were the highest paid performers, outpacing acrobats, jugglers, and animal acts, and getting to a point where they were becoming the front of the circus. In fact, it was how the circus was going to market. It was how they were trying to create the customer experience that they thought their customers wanted. But circus revenues continued to decline. The costs went up, and the clowns were not only taking money that could have been spent for other performers, but taking the money that those businesses could have been using to invest in a better product and a better customer experience.

It wasn't until two Canadian brothers, French Canadian brothers, came along and created a blue ocean strategy. They said we're going to reinvent the circus industry. We're going to go out untainted by competition, and we're going to reinvent what we think our customers want to see. They went out and said we're going to hire the top acrobats that are suddenly available out there. We're going to create a whole new performance. We're going to create a whole new scene, and we're going to bring families back to the circus and create a better customer experience.

They've evolved that to today. If you've walked down the strip at all, anywhere, you've seen the water show. A whole new experience. Multi-billion dollar industry went from one billion when they started to twelve billion in a gigantic market share. They've reinvented themselves over and over and over, right here in Las Vegas and around the world. They've created the concert. They've taken over the live music events, and they've created one with the Beatles and other concerts and so on and so forth.

Then they reinvented themselves just a few years ago. Thirty-five percent of their market is now a brand new adult-only circus, transforming from the family-oriented to a whole new business model where they went from a twelve billion market to about a sixteen to seventeen billion market.

The AI Value Gap: Why Only 25% See Real Benefits

So what does this have to do with AI? Pick your favorite number. This week, you're going to see all kinds of numbers. You're going to see trillions and quadrillions and everything else. One that Michael and I have focused on is in North America alone, somewhere around two point eight to four point five trillion by 2030 from AI and specific to a lot of the line of business, industry-specific use cases that our customers are evolving themselves and really using to transform their business.

What's stunning is that two years ago, this was zero. Last year, we were all talking about chatbots, talking about back and forth, as Michael said, some of the generative AI solutions. Now we're having a completely different conversation. We're talking about how do I achieve line of business outcomes using agentic AI and thinking completely different about those problems. We've moved from a world where I want to store data and I want to get my data ready for generative AI to a world where I'm using these agentic systems to think, decide, and act.

Eighty percent of customers, and this is a survey from McKinsey, have said that they've deployed some form of AI. Something has gone into production, something that they hope to materialize and move their business. But here's the catch.

Only about 25% of those have seen some benefit. Which means, much like the circus, we have very few winners and a lot of people who haven't seen the value, haven't seen the growth, and haven't seen the targeted desired outcomes that they had hoped to achieve. But here's the catch to that. Those that have seen the value, that quarter that have seen value, are realizing 45% average cost savings and 60% higher revenue growth, outpacing even the best case scenario for a lot of these customers and outpacing our partners who thought that they were going to grow in agentic and generative AI, driving outcomes that are well beyond what they had even hoped.

To pull the cliche back to the circus, you don't need more clowns. We need more engaged customers, but that creates a whole host of problems for our partners and our customers. I'll turn it back over to Michael who's going to give you a little bit more insight into that.

Emerging Trends in Agentic AI: From Multi-Agent Systems to Hybrid Human Interactions

There's a lot of excitement right now and a lot of anticipation, a sense of urgency. Enterprises, startups, and governments are all thinking they need to be doing something with AI. They should be doing something with AI. But there is a high rate of failure, with a high percentage of projects being pulled back and projects that don't end up going into production. In one way, that's a good thing because it means that people are experimenting, trying things, evaluating, modifying their thinking, and learning as they go.

But what are the common traits of the workloads that are being successful? What are the traits of the projects that are being successful and going into production? We're going to dig into some of that right now. For many of us, there's uncertainty: how far do I go, when will I see value, will people use it, will it be secure, will it be reliable? But I have to do something. That tension and anxiety is very real for a lot of people and organizations when they're thinking about their journey to AI.

So what are the trends that we're seeing? With agentic AI, the emerging trends can be categorized in a few areas. First, powerful multi-agent systems. We're going to walk you through some really cool examples of powerful multi-agent systems today and get your mind thinking about what that really means. Hyper personalization is definitely central to lots of great use cases that you'll hear and see about this week, creating personalization at a level that we've never seen before, whether that be through customer service or the media that people consume. Anywhere there's a customer involved, hyper personalization is central to the conversation right now.

Advanced memory is another key trend, with models developing extremely sophisticated knowledge bases to commit things to memory, learn from mistakes and successes, and make better predictions, better decisions, and produce better outcomes in the long term. And of course, hybrid human interactions. People have heard the term human in the loop, where the integration of decision-making and workflows between humans and AI is extremely topical right now. We're encouraging AI to make more and more decisions, and when AI makes better decisions and you reach a point where perhaps a certain decision vector gets above a 90% or 95% success rate, then perhaps AI could be making that decision while we focus on the 5% of exception cases. Being able to evolve a workflow and improve it as you integrate AI into it is the name of the game. Start small and conservative, and evolve the thinking and decision-making over time.

Value Creation Through AI: Workplace Productivity and Industry-Specific Applications

So what does that mean in terms of value creation that people are seeing from AI? Certainly three main areas, and you'll see these reflected in a number of the examples we're going to walk through today. Firstly, workplace productivity, which has certainly been around for a while. Anybody who's used AI to generate documents, summarize documents, pull things together, or generate code has experimented with these well-known examples. But workplace productivity goes a lot further, improving workflows and reducing the number of steps and stages in a workflow by using agentic AI, which is central to the foundational value proposition of what agentic AI is doing for us today.

Incident management is another area where there are a number of great stories. We quote internally about incident management and use it internally in terms of managing our own infrastructure, being able to use AI to intelligently make decisions and remediate problems.

This is very central to business workflows. Smart supply chain, demand forecasting, and smart pricing are all workloads that are deeply entrenched in business workflows, unlike any technology transition we've seen before. Innovation and research has been massively transformed by what AI has been able to do. Whether it be complex federated models to do deep data analysis, advanced simulations, or predictive medicine, the fields of innovation and research have been dramatically transformed by AI over the last couple of years and especially the last 12 months.

We're going to talk about the value proposition that is unique to industry. What I can tell you is that industry to industry, the value proposition and the value that people are realizing with AI is unique. The use cases are often unique industry to industry. In my own experience, never before has there been a technical transition that has been so deeply embedded in the line of business and in the business workflow, not just the technology. Business transformation and business workflow is where business impact is made.

What I can tell you, and we'll talk about in terms of what's been successful and where we've learned the most, is that those use cases deeply embedded in line of business are statistically the ones that have been the most successful, rather than the ones where we should tool around with AI and see what we can do with it. While that's a fun experiment, those ones are not as quickly transitioned into ROI as we've seen in the ones closest to line of business.

Small Organizations Moving Fast: Construction and Media Entertainment Transformations

We're going to talk about industry specifically and what the data is showing us, what we're seeing in some of our industries. So my first hypothesis is that big organizations are moving faster than small ones. I don't think so. I think that all organizations have an opportunity and an appetite to move quickly, but particularly small organizations have often seen the opportunities to look at transformational uses of AI that perhaps have been missed by many others.

I want to tell you a story about how AI is transforming the construction industry. The construction industry is big on process and projects are incredibly complex with many moving parts. Scheduling elements of a project, whether it be scheduling of cranes, trucks, formwork, and all of the things that go on in a construction project are deeply intricate. How that overlays on labor forecasting, cost modeling, and overtime management is significant. Using agents to manage these elements of the workflow has been transformational in the construction industry.

Not to mention P&L management. If you can overlay those things against the management of your P&L for your project and then do deep milestone analysis to be able to make trade-off decisions about elements of the construction, there's an enormous amount of ability to model and produce good outcomes by looking deep into your P&L and your project and making those key decisions to be able to meet that milestone, bring forward a milestone, or assess risk against a milestone, both financially and otherwise.

I was talking to a partner recently at an event, and they were talking about construction specifically. One of the things he said to me which really stuck with me was that concrete is pretty fundamental in construction. Concrete is a pretty key element. The thing about concrete though is that concrete can only be made on certain days because it's very weather dependent. So one of the things they did was they went out and looked at an agent to help manage the weather forecast so that they could more accurately schedule concrete deliveries on certain days.

They have concrete jobs that are three days, five days, and one day. They have labor requirements for all of those things and prerequisites that they need certain things to be available. They need to make sure they have a three-day weather window for a three-day concreting job. So the weather agent can manage the short, medium, and long-term weather forecast and be able to make dynamic interactions with the scheduling agents to be able to schedule the right labor, make sure the formwork is in place, and make sure the production managers are on site and ready for concrete deliveries.

For concrete deliveries, the beautiful thing about that is, and I think one of the most undervalued things about Agentic AI is the composable nature of the architecture. That is bringing in an extra factor, an extra autonomous decision-making process. It's just bringing in another agent or a different agent that has a specific autonomous, contextually relevant specialized use case. I love this one because it's a nice simple example that we can all gravitate to.

Another one that I really like is that Agentic AI is all over the media entertainment industry. There's a lot happening in media entertainment with Agentic AI, and specifically, I was talking to a partner that was doing things with storyboarding and looking at production planning around storyboards either for motion pictures or even in terms of advertising. Looking at being able to manage the production budget on a shot-to-shot basis. Using agents to baseline the P&L of a production, look at trade-offs in using agents to manage casting trade-offs, what certain casting decisions are going to cost us at different times, whether I use an actor that's going to cost this much or that much, have relevance in this market versus that market, overlay against my sales strategy in this market versus that market, the appeal of that particular casting decision in that market in alignment with my sales strategy.

Agents are ideal for this because they can take on these autonomous tasks and bring them together in terms of this deep analysis. They can make decisions about placement of talent, but also can make decisions about placement of products, perhaps in a motion picture. They can also manage things in media entertainment to do with transposing audio, inserting local content, but also they can do things like profanity detection, which is key to compliance for media entertainment in specific markets. Profanity is not acceptable in certain markets and it is in others. You can use agents to do profanity detection and remediation in real time.

What the Data Shows: Top Use Cases in Automotive, Energy, and Financial Services

Not to mention product placement or even celebrity detection. So many interesting uses on an industry-to-industry basis of what people are using agents for. The other thing I've heard is that AI is targeting jobs. Is AI targeting my job, Ryan? I hope not, but let's talk through it. We named this section "What does the data show us?" Michael just showed us that a lot of small companies, like local construction companies, are taking advantage of AI and able to reinvent their businesses much more quickly.

The other hypothesis that we hear all the time talking to customers and partners is how many people am I going to replace and how many jobs am I going to replace? It's a little bit misleading to think about AI like that. Michael talked a lot about human in the loop. He talked a lot about not just supplementing humans, but going the other way, and we're going to talk a little bit about that. What I want to show you, I'm going to show you three of our largest industries: automotive, energy, and financial services.

I'm going to talk through the number one use case where AWS is seeing quite frankly the highest revenue, where our partners are seeing the most change, and where our customers are actually innovating the fastest right now. Not to say that that'll be where it is in six months, but what you're going to see is each one of these is not necessarily a human job. It's about creating a better customer experience. It's about creating a better set of desired outcomes for the line of business and actually transforming where they're going.

I'm going to start with automotive. The number one use case we see across the board and quite frankly in all industrials, whether it's manufacturing a car, manufacturing a Nike shoe, or pharmaceutical manufacturing, is predictive maintenance. Companies are leaning more and more into the smart factory concepts. They're addressing fragmented data across multiple systems. They're dealing with the challenge of complex data integration between domains and people, and how do I get that into a natural language processing capability where I can actually make sense of this data, have AI supplement this, and get ahead of where something's going to break down.

Today, AWS looks across our industrial customers. It's somewhere about thirty percent of their engineering time is spent just searching for data. How do I make a decision? Well, I need the data. How do I get to the data? I've got to spend time curating and pulling together structured and unstructured data, pulling that into one place. Thirty percent of engineering time can now be used for other bigger, better, faster decisions and supplemented with AI.

Fifty percent of product development delays are caused by inaccessible knowledge scattered across the enterprise, meaning that product developers are trying to learn, they're trying to improve their smart factory floors, they're trying to evolve how they have humans on the floor and how they use robotics across the automation to produce products, now fifty percent faster.

The outcome is more persona-based access to institutional knowledge, more intelligent decision-making, and context-aware interactions with multimodal data. It adds advanced search and retrieval capabilities that weren't there before, and all of that adds up to predictive maintenance. We're not replacing humans. We're doing it much faster, much more accurately, and reducing operational costs significantly because I'm not waiting until something breaks down and having to stop the factory from producing. I can get ahead of it and start thinking about how to prevent manufacturing from ever stopping.

Apply that same logic to energy. Two years ago, we crossed the threshold where $1 spent on non-renewable energy sources and renewable energy sources crossed that point. We're now almost at $2 spent on every renewable resource versus $1 in the world, and that number continues to accelerate. This is driving a whole new set of use cases around distributed energy resource management. How do I predict what utility providers can provide on the power grids they serve and manage for their users? How can I give 100% renewable energy to those willing to pay for it and a whole new generation of consumers that want that in their homes?

How do I think about dynamic pricing for peak and off-peak periods, pricing for when I have wind, solar, or hydro? How do I think about that in a whole different way? Agentic AI is a perfect set of use cases. We're seeing rapid adoption in Europe, starting to see catch-up in Asia, and even a little bit here in the US. This has given utility and power companies the ability to address customer concerns they were never able to address before.

Finally, financial services, the one that probably most of you work in. By far, the most traction we've seen is from retail banks, capital markets, payments companies, credit card companies, and credit unions. KYC, AML, financial fraud, and financial crime detection now happens in seconds, replacing a process that used to take days or months. Large institutions like Bridgewater, Morgan Stanley, and JPMC have really set expectations with their customers that they're going to detect fraud now, stop it now, and that identification is happening in near real-time to identify and remediate potential fraud, whether it's cross-border payments or someone who's stolen your credit card and identity trying to make unauthorized purchases.

My favorite use case is JFK Terminal 4. Just yesterday, I was there, and a year ago, anybody going through Terminal 4 experienced remarkably slow security lines that made travel in and out of New York really painful. Over the last year, I was just there yesterday, which is supposed to be one of the busiest days of the year, and I went right through security. What's transformed it is a massive net of agents controlling everything from security lines to the app you probably have on your phone, whether it's Delta or United, now showing you the best security line, whether TSA or Clear is faster.

One of my favorite use cases is how to get through airports when you need a luggage cart, wheelchair, or other mobile accessory for someone needing additional assistance. The app can now show you using our partners' technology exactly where those carts and wheelchairs are and predict how long you have to wait. It knows exactly where you are and where you need to go, and it can schedule all of that for you without talking to a rep, calling anybody, or even clicking anything on an app. It knows that you need that and schedules it all ahead of time.

So with that, I'm curious what else you're seeing, Michael? Thanks, Ryan. Funnily enough, true story, I think my son is actually in Terminal 4 at JFK at this very moment, and he just texted me this morning to say that he went through Amsterdam Airport last night, and it was an absolute train wreck because they have not yet embraced agents for making that airport run so smoothly. I'll be interested in his feedback today. It's funny, he's literally there right now, so funny coincidence.

AI in Highly Regulated Industries: Real-Time Compliance and Retail Innovation

Let's talk a little bit about another hypothesis that AI is not working. For highly regulated industries, I'm going to tell you that it is working really well for highly regulated industries. We're going to talk about some interesting use cases and let's start with compliance. Compliance is one of those use cases that cuts across every industry.

Whenever there's an agreement, whenever there's a contract, there's a use case for compliance. But compliance has been radically transformed across lots of different industries. I hinted at one example before, but I'm going to talk about a couple more now. It's been radically transformed across all industries, whether they be regulated or non-regulated industries.

Agentic AI means that compliance management is becoming an active part of the business process, not just something that you look at as an afterthought. Compliance is moving from a model where we think about something happened and ask whether it was compliant, to compliance being able to be managed in real time. That manifests itself in different ways in different industries, so I'm going to give you a couple of quick examples here.

In media, we can look at compliance in real time in terms of video streams and audio streams to make sure that, for instance, profanity is monitored, but also that the correct language is used or that the right information is being presented on the screen for that market in compliance with local laws, requirements, and regulations. It can analyze product placements and all those sorts of things in line with a contract that you might have with an organization to deal with product placement.

Retail certainly is an industry where we've seen an enormous amount of interesting use cases and enormous amount of growth. We've seen up to 350% growth in AI solution adoption in retail, whether it be dynamic pricing, which is really compelling. You have agents that go out and monitor competitive pricing and make judgments and recommendations based on what the right price would be for your product, looking at the P&L and your revenue targets for this month, using a series of agents to manage your pricing decisions.

Smart supply chain is another major use case, being able to manage where you have stock and how you should manage your discounts in what markets to address what stock you have in different locations. Those two things work so nicely together. Multiple agents look at managing your P&L versus your discount versus your competitors while looking at your supply chain. There are some very interesting multi-agentic workflows in place in retail, which are super compelling and the ROI just jumps off the table. It is so easy to see because you can offset that straight onto your P&L.

One of the things about retail is that it's not regulated, so there are a lot of use cases that are emerging and a lot of new use cases. Agentic AI is moving super fast in these industries. Twelve months ago, if you came to re:Invent, we would have been talking about just walk out and virtual try-on in retail. Those are two great use cases, but they're pretty old hat now. Dynamic pricing and smart supply chain have really transformed and accelerated hugely this year.

We're also seeing interesting agentic use cases in-store, and that's one that really jumped out at me in the last couple of months. For example, we've seen compliance as a use case in-store to manage cameras in-store to monitor the shelves and make sure that the presentation of a product is in line with the contract that they have in place with the retail outlet. Or even to ensure that the shelf is stocked, so that if they're running out of a certain product, the warehouse is alerted to say the shelf is looking a bit thin on this item.

The compliance side of that's very compelling. As a supplier, you can make sure that your retail outlets are correctly merchandising your product on the shelf, and it's done all in real time. In-store also comes down to things like safety and fire codes. Monitoring your in-store foot traffic using agents enables you to say, okay, we have the maximum number of people in the store, or are we approaching that point, or in a certain part of the store, are we overcrowded? Do we need to manage the amount of people coming in at the door? Being able to manage against fire codes is another great example of real-time compliance management that's completely dynamic.

One of the other interesting use cases when it comes to in-store monitoring is being able to look at customers' movements inside the store. Agents can track people's movements inside the store, look at where they're dwelling in the store, what stands they're standing in front of. So that if Ryan walks up and looks at perhaps the same set of sunglasses three times in ten minutes,

they could say, you know what, Ryan is pretty interested in that stand of wayfarer sunglasses, and deploy a human agent, a sales agent, to go and say, "Hey Ryan, are you aware that we've actually got a discount today on the wayfarers? You might want to consider tucking a pair in your pocket and walking out with them today." This is a whole new use case, right? What I think is the most compelling about it is really the sky's the limit. This is not a use case that even existed twelve months ago. Deploying a human out to talk to someone who's looked at a pair of sunglasses three times is not something we even dreamed about twelve months ago. Customers are actually deploying this technology in retail outlets today, which is super exciting.

Healthcare and Life Sciences Revolution: From Philips to Moderna and Cure.AI

So Ryan, perhaps you want to tell us a couple of examples in more of a regulated context. Yeah, thank you. I mean, Michael said 350% growth right now in the retail space using agentic AI. I would love to give you a number for healthcare and life sciences, but it's growing at such a rate we don't really have a good metric. It's literally changing month over month, but it is growing at an outpaced pace. It's things like Philips, where Philips is literally transforming the way they go to market by creating agents that supplement everything from radiologists to medical clinics. They've transitioned from what might take days or even weeks to get an analysis on things like medical images to five seconds, meaning they're supplementing the radiologist, supplementing the pulmonologist, the cardiologist with real-time information.

We're going to talk a little bit more about what that looks like with one of their customers in just a second. But that results in a thirty times cost reduction. It results in more accurate diagnosis and then obviously in recommended therapeutics. Philips has taken that concept and created their entire AI tool suite for all the developers, all of their engineers, and that allows them to get things out faster and more compliant so that their entire SDLC is built on compliant AI now. The agents are constantly monitoring as you're building code to ensure you're not in any violation of HIPAA requirements or European requirements, and to determine where we're deploying this technology and where we think it's going to go to production, getting ahead of that regulatory burden or compliance burden.

Things like surgical procedures are now more precise, and it's allowing them to do it in a more secure, more auditable, and compliant way. One of my favorite examples is Moderna. Moderna not only did something similar to what Philips was doing with all of their electronic health records and making sure that they were compliant, they went a step further. They said, I'm going to take my HR organization and my IT organization, I'm going to put them together. They created a new role, CHRO. What the CHRO does is look across both organizations and say, how am I using AI to make my people more productive? And how am I rolling out AI so that my people are making AI more productive? They've literally combined the two organizations.

Michael started this presentation by going through three themes, and we've talked a lot about the business workflow. This is about the human-AI interaction and how we can make our work, our workforces not just in the loop, but actually make humans use humans to make AI more productive. Fast forward to today, nineteen out of the top twenty pharmaceutical companies globally are already deploying some sort of AWS generative or agentic AI into production. Four of the five largest genomic sequencing companies in the world are using AWS or some of our partner technologies. Ten of the ten largest medical device companies in the world are already using AI on AWS, and eighty percent of healthcare and life sciences unicorns are already deploying some sort of AI to help their customers and leverage this.

The punch line is it's you, it's our partners and our customers that are building this technology, taking advantage of Agent Core, of Bedrock, of a lot of the transformed technologies to enable you, but it's really those products that our customers are then consuming to make their AI processes better. One example I'll give you right now: According to the European Lung Cancer Association, in a recent study, fifty-four percent reduced fatalities from lung cancer if detected early in the process. Cure.AI using this technology rolled out a set of agents with AstraZeneca for any clinic in the United States. If your doctor says, "Hey, you might be at risk for lung cancer, go get a chest scan," when you get that pulmonology scan today, the first person to look at it is not a human being at all. It's not a radiologist, it's not a pulmonologist. It's Cure.AI with AstraZeneca running a set of agents to detect early signs of lung cancer.

What that does is a couple of things. One is, if you think about the life cycle of a radiologist through the course of their career, they might see hundreds or thousands of chest scans.

AstraZeneca and Cure are training every one of their models on a minimum of 25 million and sometimes many hundreds of millions of scans. This not only gives them a higher level of accuracy and predictive diagnosis, but it also allows them to go back in time. I could take a chest scan of Michael's chest today and compare it to Ryan's chest scan from five years ago, then ask what therapeutic or treatment did Ryan take and what was the resulting outcome?

We can take that comparison and apply it not just to a couple of people throughout the career of a radiologist, but through thousands or even millions of people that have similar medical imaging. With that, I turn back to Michael. What have we learned today? We talked about a lot of things, and I tell you what, we won't be beaten on the number of customer stories. I hope those have been extremely valuable to you. Let's pull together some of the common threads that we've discussed through these great customer stories.

Lessons Learned: Start Small, Think Composable, and Pick the Right Path Forward

Here's a startling statistic: 40% of Agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. That's a significant number. When I looked at that number, I had two reactions. On the positive side, I think that means people are experimenting, and that is a healthy thing. Experimenting and not waiting for perfection before starting is central to success with Agentic AI.

Useful agents need to understand natural language, reason effectively, and maintain context over long interactions. Most importantly, they need to do that in a trustworthy and secure way. They need to integrate seamlessly with existing systems, maintaining proper authentication and authorization. Of course, they need to interact with other agents. One of the absolute foundational premises of Agentic AI is that agents work well with other agents as well as humans. We need to ensure consistent performance and scale to reduce the levels of complexity.

Enterprises are doubling down on agents. Here are two more interesting statistics from Gartner. 33% of enterprise software apps will include Agentic AI by the end of 2028, up from less than 1% in 2024, which is probably not a surprise. This one is interesting though: 15% of day-to-day work decisions will be made autonomously through Agentic AI by the end of 2028. I think that's a very conservative number actually. My personal view is it will be much higher than that. When you think about the volume of decisions that Agentic AI can make on a day-to-day basis, I think we're going to see that number greatly exceeded.

Scale requires operational excellence. The path from proof of concept to AI production agents and realizing value is a journey, but it's not a journey that you're on by yourself. It's a journey where we have many programs and tools to make you successful. Keep it reasonably simple, but performance, scalability, security, context, and governance are all central to your decision-making as you go on a path of AI. Handling user identity correctly, detecting degrading accuracy, and troubleshooting errors are all things to be taken into account.

I said before that one of the most fundamental aspects I've seen when talking to customers and partners in terms of success factors is to be fundamentally grounded in a business problem, not just technology for technology's sake. We talked very early in the presentation about how everybody wants to be doing something with AI and feels they need to be doing something with AI. But I strongly encourage you to work from the customer problem back, from the use case back, from the business pain point back. Like no other technology transition that I've experienced in my life, this one is deeply ingrained in business process, not just technology on its own.

What are the outcomes we're looking for? What are the employee tasks that we could consider automating? Once you have that clear business problem in mind, you're on your way. The Agentic mindset and we have a customer-tested framework to think about working backwards from the business problem, using data as your differentiator, using trust to drive adoption, and demanding the best price and performance to enhance your environment with multi-agent architectures. Most importantly, this isn't happening in the future—it's happening today. When you think Agentic,

think composable. An agent using an agentic architecture enables you to use more agents and integrate more agents. But most importantly, you don't have to come up with them all yourself. Other people are building agents all the time. There is a wealth of agents available for you to deploy, purchase, and procure through your existing purchasing contracts with AWS in the agent marketplace. They're specifically listed against use cases and customer problems. I thoroughly encourage you to check it out and dig deep into it. There's a new solutions page with lots of great content there.

The other point I want to leave you with today is to start small but start now. Don't wait for a perfect moment to have it all figured out. Experiment and try things and pull them back. It's a good opportunity to work with the tools, see what's possible, and enhance your thinking through experimentation rather than trying to come up with the perfect architecture or perspective to start with. Start small, start now.

Lastly, pick the right path for you. You can buy agentic solutions and use agentic tools that are available right now through AWS, such as Amazon Quick Suite and Kiro, to name a couple. They're available for you to deploy from your AWS account. When you're ready, go for it. You can build your own custom agents, and we've talked a lot about specific use cases in industry where building custom agents can address line of business and industry problems, create new use cases, and provide new abilities to differentiate yourselves and accelerate your customers. Those agents can be reused, and you can even sell those agents to help other organizations and monetize the value that you drive from building your own agents.

Lastly, utilize partners for specialized expertise. We have AI competency partners—140,000 partners in the network—and a good subset of them are pursuing AI competencies to more diligently serve our customer base and help customers on their journey. Whether it's a consulting partner skilled in taking a business case through to production outcomes or ISV partners who build software and agents to address specific industry problems and challenges that customers have on AWS, partners play a crucial role. Well, thank you so much for spending time with us today. It's been an absolute pleasure. I hope you have a wonderful time at re:Invent this week, enjoying getting to know each other and learning. Take the time out, put the email down, and we look forward to seeing you all again soon. Thanks very much everyone.

---

; This article is entirely auto-generated using Amazon Bedrock.