🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - Integrated AI Agents: Moving Beyond Summarization to Automation (AIM207)
In this video, the speaker traces computing evolution from SaaS 1.0 to SaaS 2.0, explaining how traditional cloud tools created rigid departmental schemas that AI can now transcend. They introduce Computer by DevRev, a unified platform that bidirectionally syncs data across systems in near real-time, enabling agentic AI to search, reason, and take actions across organizational boundaries. The solution combines SQL-based fact grounding with conversational interfaces, semantic search, and MCP integration, allowing citizen developers to upskill agents at personal, team, or enterprise levels while maintaining proper access controls and auditability.
; 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
From SaaS 1.0 to SaaS 2.0: The Evolution of Cloud Software and Its Limitations
Hi, how's everybody doing? It's exciting to be here. This is very ambitious. Over the next 20 minutes, we're going to cover roughly a half century of computing, but we'll try to go as fast as possible. We'll start off and try to end with AI, so we'll start and end with AI.
First of all, show of hands—anybody recognize any of these logos on the screen right now? I imagine so, since this is definitely a SaaS conference. These are products that have represented the best of what SaaS has offered us over the last 15 to 20 years. This is how we went from spreadsheets sitting on our local computers to software that could run in the cloud, cloud-native with the scale, safety, and speed of cloud behind us, giving us auditability, transparency, scalability, and security. That's been a great journey, and I call that SaaS 1.0.
But there's a flip side to it. We have now gotten to a place where all of these disparate tools have created their own rigid schemas that define at the department level what good looks like. They have these schemas—whether it's an opportunity, a ticket, an incident, an event, or a case. Each of these respective department tools has created its own terminology and glossary, and that glossary comes to really define how we use these products and how we do our daily jobs. There's something about that which actually gives us an opportunity to think further and ask how AI would revisit this if we were to look at where we go from here, from SaaS 1.0 to SaaS 2.0.
The AI Implementation Challenge: Why 95% of Pilots Fail and What It Takes to Succeed
I want to step back and ask how the greats over the generations of computing have tackled this problem. What have they done so incredibly well? First, they catered to the builders. They created the right separation of concerns so that a lot of problems could be solved at scale, allowing developers, great minds, and entrepreneurs to do great things on top of the platform that was built.
This is great because before the cloud really came to be, a lot of people were reading books about client-server architectures, JSON, and APIs, saying they could do that at home or in the workplace. You'd set up a Linux server, and before you knew it, you'd have an IT closet full of servers and racks serving real use cases. Maybe you're handling transactions online, but at some point you realize you have to do security patches and security updates, deployments, and ensure zero downtime. You have to make sure the hard drives are still working. This became a problem, and that's where a lot of these tools started to come together, creating the right abstractions for us to do a great job.
It's similar to what we're starting to see here as well. Everything starts to look really easy to begin with. I've got an LLM at my disposal, and this LLM is exceedingly easy to use conversationally. It can do some really great things—it can classify, prioritize, summarize, synthesize. You can do all of these things and put this to work to solve real problems toward specific conversations and specific tickets, and so on. But what happens is you can start to increasingly get to a place where you get back to that really messy IT closet.
We've started to see this in the market now. You're starting to see situations with headlines like this. I don't know, show of hands—people who actually saw this story or heard of it? Companies went out and said they're going to stop using humans to do problem A, problem B, problem C. They've got this covered. They're going to do it using agents. It's that simple, but it wasn't really that simple, was it? Six months later, there's a new headline from the same company. No, actually we're going to need those humans again.
This has been happening, and why is that? It's because we have to really start to understand what are the right abstractions, where are the areas of opportunity, and what are those separation of concerns. As we figure that out, we're starting to see headlines like this. MIT reports, and I know many of you have probably seen this on your LinkedIn feeds, where we talk about 95 percent of these pilots failing. What does that really mean? I think what it means, among other things, is the fact that you not only need a frontier model layer, but you need to really think about the data layer.
You need to think about how data becomes accessible and in whose hands. Access provisioning makes sense. You have to start to think about replication and data sanitization, and that's just the data layer. On top of that, you need to think about provisioning, auditability, agents, workflows, and repeatability. You have to start to think about end-to-end use cases. How do you serve real use cases in customer success, customer support, field sales, development productivity, and project management? These are real problems, and until you're willing to not just think about the data layer and build a prototype on top of it, when you get ready to actually look at the data layer, the operational layer, and the application layer, that's when you're ready for production.
It's a journey, so we have to think about that and ask: where are we headed? The reality is there's tremendous momentum because people believe there's a huge opportunity here with AI, particularly agentic AI and conversational AI. Money is already being spent. I'm sure some of you go back to your workplaces and see your leadership asking how you're taking advantage of AI. Money is being spent, budgets are being shifted from headcount perhaps over to making decisions around buying tools or AI products, and asking how you're becoming more productive, how you're reducing customer effort or employee effort, and how you're improving customer satisfaction. These are real questions being asked, and we have to figure out how we get to answers that really help us in the long term.
Learning from Computing History: Building an Operating System for the AI Era
I'm going to take us further back from cloud and ask: how did we get started with computers five or six decades ago when we had those behemoths sitting in our workspaces, workstations, and mainframes? How did we get to where we are today? The reality is we started off with something expensive, onerous, and it wasn't really easy to figure out how it all makes sense, but we figured it out. We decided that to go big, we had to go small. We had to create an operating system, and we had to start to figure out how to build more programs on top of it, build workgroups, and make things shareable and more use case driven. We created a user experience, and that was our ticket to the kind of expansion we saw in the decades to follow.
If we start to think about that and ask what that teaches us, it teaches us that if we build the right paradigm, you can have tremendous purpose by thinking about the hardest problems and building on top of those paradigms. That's what we've done. We got more productive, more efficient, we got more creative, and that's the huge opportunity that's existed not just when mainframes became workstations for us, but then when the desktop came out, when cloud came out, and again here with AI. That's the same opportunity we have. It's an opportunity for us to define the operating system and then figure out how we use and develop that to become more productive, efficient, and creative.
How do we bring that computer mindset to AI? How do we bring an operating system to the forefront here? Let's think about it and ask: what would it look like? What would the dream computer look like? The dream computer would be able to take all your critical data and all your critical integrations and bring them to your most important surfaces. That would be the computer you'd want. It would be a computer that did not create false boundaries between your opportunities and your incidents, your tasks and your sessions, your issues and your cases, your events. It would really understand and respect all of it and wouldn't think that any of this information was foreign. It's all part of your business. These are all departments serving the same customer, working towards building the same set of products, am I right?
Introducing Computer by DevRev: A Unified Platform for Agentic AI in the Enterprise
That's where we want to go. How do we get there? Well, what we've thought about is creating a product that does just that, and we actually call it a computer. It's called Computer by DevRev, and it's exactly that. It brings your most critical data and your most critical integrations to all your most important surfaces, whether that's at the workstation, on the desktop, in a customer help support center, in a portal, on your website, perhaps on the mobile device. Wherever you need it, Computer is there. It's multiplayer, so it's not just there for you, it's there for your team and your organization, and it's also there for your customer. There are some really hard problems behind the scenes that need to be solved to do this really well. First of all, you need to understand how to unify information, and unifying information is not just about being able to export data. That's not what it's about. It's not about taking data from here to there. That's not the problem to solve. This is about creating memory for agents and AI for Computer really. This is about understanding how to bring information in a near real-time bidirectional way.
Because if it's bidirectional in near real-time, you can actually take actions on that information and have it reflect back in the systems where the information originated. So you can continue to coexist with the rest of your infrastructure, with the rest of your long tail of IT. It has to coexist , and for that reason alone, it needs to be bidirectional and near real-time.
It can't just bring in the data; it's got to bring in all the metadata. It's got to understand the metadata so it knows what the other options are in the dropdown menu and it needs to understand what the other radio button selection possibilities were. How else would you actually take action if you don't understand the field dependencies and the access provisions and the selection criteria and the variables at play?
That's where you want to be. You want to be in a place where you can bidirectionally sync information and understand the schema and the access provisions so that you can respect all of that as you create a system that can search, give you answers, and of course take action on that as well. When you bring it into memory, memory is going to have to handle both long-term and short-term context. It's got to be able to give you the most fast, accurate, and hyper-personalized answers.
That's the only way to create a personal computer. Of course, you can't just do that by mimicking the data as it exists in these disparate systems. You need to understand how to connect data between the work that you do, your persona, the customers you serve, and the product you create. This information and these inner relationships need to be understood deeply.
If you're going to do all that, you're going to need a series of foundational services on top of it. These are simple but powerful foundational services. You've got to be able to search, and of course what we know is that as we have gone deeper in search, it's a deep problem. You've got to be able to do it at scale. You need to understand not just the syntax and the keywords, but the semantics, and you've got to enrich this information.
It's got to be able to understand the deep annotations, the links between work and opportunities and customers and entities and identities. It's got to understand that really well, and it's got to be fast at the scale of GPUs. That's where we want to be. Of course, it has to be able to contextualize all the answers based on who you are so you're not just giving the same information to everybody. It's got to be unique and personalized.
Of course, it's got to have a data layer that's well suited to agentic AI. Agentic AI can't be built effectively on top of the data layer that we use to build client-server architectures or to build traditional client or cloud systems. It has to be a data layer that is native to conversational as much as it is to more conventional surfaces. SQL—what's old is new again. SQL is something that's grounded in facts. It's easy to advocate for because it's grounded in facts, so you can actually audit it, you can trace it, you can acknowledge it, you can permit it.
So you can conversationally access your information, and you know that the answer that you got was based on real facts because you have the understanding of the SQL that was working behind the scenes to actually provide that information. You can then take that and in a more traditional way access it to create a graph or a dashboard or to share that information with somebody else who can do more intelligence on top of it. This is the most universal way for you to communicate with your information in traditional ways but also in more conversational ways. That's what the future looks like. It's not abandoning SQL; it's actually embracing it as a unified data layer.
Of course, you also have to say yes to not just the data in memory but the data outside of memory. Just like computers had access to the Internet, so do these agentic computers. These computers have access via MCP so you can actually access tools from around the world and have plans that reflect not just the information inside your memory but all the information outside it as well.
This is what a modern computer looks like. It's built on all of the engines—the powerful engine of search that is both syntactic and semantic and powered by LLMs, but also SQL, the ability to ground facts but be conversational as much as you are conventional dashboards and conversational at the same time. Of course, MCP as well, and that's just on the access for searches and answers. But to take actions, you need to be able to build agents and skill agents because we're not going to be able to create thousands and hundreds of agents.
Who's going to name them and who's going to address them? It's too difficult. You've got to have a thoughtful mechanism to actually upskill these agents, and these agents cannot be put in everybody's hands to go create upskilled agents for the rest of the company. That's not going to work. You have to allow for citizen development. That means the ability for people to actually upskill an agent.
This capability is critical, whether for themselves, their team, their department, or the entire organization. It must work within these scopes. You have to allow for skills and agents to manifest at the personal level so that people can have autonomy to create the type of automation they need for themselves, perhaps for their teams, and then maybe for the company as well. Of course, the same principle reflects back into the enterprise. You cannot have the same agents advocating across all your customers uniformly. It has to be different and unique for every brand, for every region, for every channel. It is not a one-size-fits-all approach.
You need to be able to upskill agents in a way that is scalable, allowing you to create skills, borrow them from a marketplace, or create and define them in a no-code way. You must also deploy them in responsible ways at the brand, region, channel, team, or personal level. The computer has to be able to answer questions by accessing all of this information so you can answer questions like the one shown here. It is able to work through all of these tools you use, whether it is Jira or Notion , and it does not matter. It is able to access that information and provide an answer that is fully cited so you can continue with your work.
Of course, it not only synthesizes information but also has to be able to do reasoning on top of it. You want it to synthesize information, but you also want it to create a plan and reason about the information so you actually get an opinion. This opinion leverages the innate ability of frontier models to cluster and synthesize, prioritize, and classify information, which is super important. Here you can see that not only are we getting an answer that is grounded in the information across our enterprise, but it actually gives you opinions grounded in facts that are fully cited .
Of course, you could follow up those questions with real questions like what are the action items, and it will actually tell you. Not only are you able to unify information and use planning and synthesis to get reasonable responses, you can actually take actions. In this particular case, I am not only asking questions, I am actually asking for a meeting to be scheduled . This is the real synthesis of where this is going. You are going to have the ability to unify your information and have a teammate that can actually help you reason about what needs to happen. If you agree or disagree, you can ask follow-up questions with the same level of context. Finally, you can take actions to create work, create opportunities, or schedule work .
These are some of the examples, and putting it all together, ultimately it has to help you work more productively and faster than ever before. That is where we are headed. We have to be able to ask questions like what is our travel policy and not only expect an answer that is grounded in the information that exists. It is the exact information that makes sense for who I am and the department that I belong to. I can ask follow-up questions and say, what about this other region? Not for myself, but I want to know what the travel policy looks like in that other region , and it will actually tell you because it understands that that is not the region that is native to you . These are the kinds of experiences that we will come to expect from these modern computers.
Just to recap, these are not systems that are going to work at the boundaries of departments. They may help you in your departments, but they have to be able to go beyond those departments. You know what, I am not just a computer for your support team . I am a computer that also works equally well for the people that are building the software, people who are working closely with the customers, and people who are creating deep insights so that we understand where to go next. It has to work for everybody. It cannot create false boundaries between agents and microservices or applications . It has to coexist with all of that.
Just to keep it really simple, you can find us at booth 1204 . If you are excited and do not know where to go next, booth 1204 near Theater 2 is where we are located. Our founders, Deri Bande and Manoj, previously of Nanic's fame, left in 2020 to start this company . We have got a forward-deployed team and applied AI to help you with this journey. Of course, we do a lot of workshops, and Deborah of View helps you with understanding how to embrace this world as we make this transition from SaaS 1.0 to SaaS 2.0. Beyond that, we have hundreds of customers and partners. You see some of those logos, I am sure you recognize those, and of course we are fully compliant with GDPR . We have raised lots of money and so forth. Thank you so much for your time.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)