AWS re:Invent 2025 - How AI Agents are Changing the Way Revenue Teams Work (STP110)

🦄 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 - How AI Agents are Changing the Way Revenue Teams Work (STP110)

In this video, Rox CEO Shriram explains how their AI-native revenue operating system transforms enterprise sales productivity. He traces the evolution from Siebel's monolithic CRM through Salesforce's cloud era to today's fragmented landscape where sales reps spend only 20% of time actually selling. Rox addresses this by building a four-layer architecture on AWS: data layer, knowledge graph, agent swarm, and multi-channel delivery. Their multi-agent orchestration system deploys account-level AI agents that handle the entire revenue lifecycle from lead to renewal. Processing 400 billion tokens monthly as part of OpenAI's 1 trillion token club, Rox demonstrates real impact with customers like Ramp achieving 15% more six-figure deals and 90% reduction in meeting prep time. The company emphasizes their forward-deployed engineering approach over DIY agent builders, enabling consistent execution across revenue teams.

; 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

Introduction to Rox: The Revenue Agent Revolution and the Broken Revenue Stack

It's hopefully the last session today. I hope you all had a good re:Invent first day. Myself, I'm Shriram, I'm the co-founder and CEO of Rox. Rox builds revenue agents for the Global 2000. As you all know, we are in the midst of a huge fundamental platform shift that's happening. After the cloud, this is the complete transformation that is happening in enterprise technology. Just as coding agents 10x the productivity, we believe revenue agents will 10x productivity for enterprise revenue-facing teams. Over the next few minutes, I'll walk you through how we got here, how the existing revenue stack broke, and how AI native architectures are emerging and how Rox's AI agents are helping enterprises move from pilot to large-scale production.

All right, so let's get started. About ourselves, our mission is simple but ambitious. We want to secure and grow the world's revenue. We've assembled a world-class team. Myself, I'm from ex-AWS, and we have people from Google, Stanford, Berkeley, and so on and so forth. I was at AWS for seven years and built Amazon Aurora with a few people in the audience here. We launched Amazon Aurora here in 2014, and that was the first re:Invent I was here for. After that, now with Rox, we launched out of stealth last year in November 2024. After that, we have worked with over 5,000 organizations across different verticals like financial services, energy, healthcare, semiconductors, and sovereign AI stacks. We are also backed by Sequoia, General Catalyst, and GV. Our mission is to redefine the revenue systems from the ground up.

Now, let's have a little bit of history of what has happened with revenue systems. In the on-prem era, Siebel was the single monolithic CRM. It was very powerful, but it was very rigid, expensive, and incredibly hard to operate. Then with the advent of the cloud and AWS, Salesforce became the system of record, the CRM, along with Salesforce and HubSpot. But what happened was the system of record and the system of action actually split. The system of action became tools like Clari, Gong, Outreach, and so on and so forth. Now with the advent of Cloud 2.0, which is with the advent of data warehouses like Snowflake, Databricks, Redshift, and so on and so forth, the data moved from the CRM into the warehouse. And why did people move it? It's because they wanted to integrate it with their product usage data and construct dashboards on top. The workflows were still fragmented across all these point tools out there. Now we believe that there is a new bundling cycle that's happening where the system of record and the system of engagement are coming together in the AI native era.

Now this historical shift sets up the real issue. Revenue work is fragmented, manual, and incredibly inefficient right now, which brings us to why these teams are so ripe for AI augmentation. Now, as I explained before, there are three main reasons. One is that data is fragmented across multiple different point tools. You have CRM, you have support, you have product usage on the structured data side. You have your call recordings, your emails, your transcripts, and your documents, which could be in S3 or SharePoint or wherever on the unstructured data side. Now this data is completely fragmented across multiple different systems. Your workflows are still extremely manual and inconsistent. Reps have to take a long time doing manual research on accounts. They have to write and rewrite their emails and what they have to send to their customers because they have to pull context from multiple different systems, and they have to chase different teams to get that information to them.

All in all, what happens is, as you see in the chart here, this is from a BCG study, only 20% of the time is actually spent selling, and the remaining 80% is just manual drudgery and administrative work, which is completely ripe for AI augmentation. So this is why we believe a net new architecture matters, and we are at the precipice of this transformation. In the next few slides, we'll explain how Rox is actually taking advantage of this. As I explained before, we have built a fundamentally different AI architecture that is designed to unify this context and action across the entire revenue lifecycle. Rox is the first AI native revenue operating system that is built for this transformation, and it is across four different layers. We build everything from the ground up instead of doing anything bolt-on and having a separate tool for you. First, number one, you have the data layer.

Rox's AI Native Architecture: From Data Layer to Agent Swarm

At the foundation, you have the data layer where all the structured and unstructured data is accumulated across your CRM, support, and product usage. You have your documents and all that information on top of the data layer, which could reside on your side or it could reside on Rox's side.

We have built a context or knowledge graph or data fabric layer. Imagine this is like metadata that actually has pointers to your data and where your data lives. This metadata consists of entities and relationships. It actually extracts entities from unstructured information and documents and integrates it with the structured data out there. So what we have is metadata that contains a knowledge graph, and the agents that sit on top of this are grounded on the context that this layer provides, so agents don't hallucinate. They are actually grounded in the data, which is grounded in the metadata that has pointers to the data.

On top of this, we have the agent swarm. These are turnkey enterprise agents. These are not DIY agents or agent builders, and these agents run the entire revenue lifecycle from lead to close to renewals to expansions, churn prevention, and so on. Now this is all delivered to you across multiple different channels like iOS, Slack, web, and so on, where we want to meet the sellers where they live.

So this is a ground-up build all the way from the data layer to agents to end product. This architecture is what allows us to deliver consistent experience and scalable execution across your entire revenue organization. We'll talk about how we actually operationalize this. We call it the agent swarm. If you have a rep in your company, each representative has a bunch of accounts, and for each of their accounts, they have AI agents that are orchestrated by Rox.

Imagine for every account out there, there is a bunch of AI agents that are responsible for driving the entire revenue lifecycle. They orchestrate amongst themselves so that they can augment you. Imagine they're sitting in every conversation, augmenting and understanding what's going on with respect to that customer. For every account, for every rep in every organization out there, that's what happens at scale.

So now we can convert the insights into action. For example, let's give an example. Let's say you're trying to break into an account and there is a new funding round. By the way, that is research that is happening on the public data side. And then on top of that insight, you have to draft an email. Now the agents need to know that there is the context of you been trying to break into the account, and they're going to draft an outreach ready for you so that you, as a human in the loop, are going to come in and actually send that email.

You can also completely automate this so that the research to action or the insights to action is completely automated out of your way. Imagine doing this for deals that you're running right now and you're trying to figure out risks in the deal, and there are new risks that are surfacing because there is information that is happening on the public internet. There are emails that are happening, and then there are transcripts. Figuring this information out and then surfacing what is the action that you need to take is what these agents orchestrate along, which was not possible before. That's what this architecture enables.

Platform Architecture and Multi-Agent Orchestration at Scale

Now let's get a little bit one step deeper into the platform architecture. We built completely on AWS, and we have a clean separation between the customer's data plane and Rox's data plane. Customers can either share their data with us using secure data share. This is standard in every data warehouse like Redshift, Snowflake, BigQuery, and so on. Or they can actually pump the data into our warehouse.

We fully use all the AWS services out there, from intelligence to compute to content ingestion to orchestration, and so on. Now, of course, being on AWS gives us the reliability, the global footprint, and the security posture for all the Global 2000 deployments that we're after. On top of this is the secret sauce, which is the multi-agent orchestration.

We do not run a single general-purpose agent that is trying to do everything. That just does not work. For each workflow out there, you have an orchestrator agent that is going to take the workflow, break it into a sequence of steps that needs to happen, and for each of those steps, choose the right agent that is responsible for those steps. This breakage, this actual split of the responsibilities, is what gives us the quality out the door.

Now all of these agents share a common working memory, and they're pulled from the unified knowledge graph. The working memory enables agents to understand what went on before and what's going to happen later on, but then pulling from the knowledge graph helps it ground in the context and factual information instead of hallucinating something out there.

So the result is a predictable, consistent, repeatable execution across all your accounts in your enterprise that you can scale not just for one rep, but across all the representatives in the organization.

Now let's take a simple example. This is a huge query that the seller can have, and this contains information across earnings calls, press releases, job postings, product signals, leadership commentary, and so on and so forth. How does this work in practice? The orchestrated agent gets this request and is going to break it down into a bunch of research which comes in from private research, which is your product signals, and public research which comes in from the internet.

When the query comes in, it first goes to the guardrail agent. The guardrail is going to look at and see, hey, is this user even allowed? Does he have permissions to actually access this product signal? And second, it ensures that it's safe for work and not something that you want the agent to answer. Then the orchestrator breaks it into multiple research agents.

Now let's take a public research agent as an example. It's going to call the query expansion agent to actually expand your query to ensure that it is sufficient to get all the information, do a parallel web search, come back, and then synthesize it using a reasoning agent and finally format it using a formatter agent and then verify using a verifier agent. Now, anybody who has built AI agents in production knows that this is just the starting point of how you build workflows.

You have the invisible asymptote where you have to scale it across multiple different accounts, multiple different agents. You have to ensure model fallbacks and reliability. Models fail all the time at scale. You have to ensure that the model fallback. When you fall back to a model, you'll fall back to a different provider, so you need to ensure your prompts are working for that different provider and you have your evals and guardrails set up. And of course you have to optimize this at scale because you have to optimize for cost. You're not just burning tokens left, right, and center. So that is the invisible asymptote which is very hard to build if you're building internally because you need the necessary talent and understanding of how to build these agents.

Now this is a fun slide here. We are of course operating at scale as I mentioned before. We are already doing 400 billion tokens per month. We are part of the 1 trillion token club from OpenAI, only 20 organizations out there. So this is rare in enterprise AI. This is not experimentation. This is daily operational dependence from revenue teams for building their entire revenue life cycle from research to outreach to meeting prep to opportunity management to renewals and expansions.

From Pilot to Production: Enterprise Adoption and Real-World Impact

All right, so we talked about the vision, we talked about the motivation. It all looks great. Now, let's talk about the real adoption in the enterprise. So, everybody from all the revenue leaders expect Gen AI to become critical for sales, marketing, and services over the next few years. The vision is very clear: increase revenue per rep, growth without the headcount increase, and of course modernize my go to market.

But the reality is that AI is still bolt on. It is not integrated into the core workflows. The time to value is extremely slow, and there's a behavior change that is needed across your representatives to start using this technology. You guys probably followed the MIT case study where 95% of the Gen AI pilots actually fail in production. They don't give any business value. So let's dig deeper on why existing approaches actually struggle to deliver value.

If you look at the landscape today, there are three different options that you have. One is the incumbents. These are legacy CRMs. They're very powerful, but they're very rigid. They're very hard to configure, and they come up with a lot of technical debt. Their AI strategy is to actually have DIY or agent builders on top of their existing platforms. They expect your IT teams or engineers or representatives to configure these workflows, which they are not equipped to do so.

The next is you have bolt on. These are worse than before because now you have to buy multiple vendors for the entire revenue life cycle, and there's a huge integration burden for you. We believe just building from the ground up, Rox is positioned to take care for the enterprise AI transformation, and we'll be a long term transformation partner for this.

Now how do we do this? What is different amongst the ones that I've mentioned before and what I'm talking about now? So if you look at it, we do not rely on IT teams to build their own agents or prompt their own workflows. That does not work.

So what we have is we come up with our own forward deployed engineering team. That has a blend of ex-McKinsey and BCG consultants and software engineers whose only goal is to deliver measurable impact to you quickly.

So the process is simple and repeatable. First, we launch a focused pilot with measurable outcomes. Second, we train your existing teams. We also then train the trainers because they're going to be the people who are early adopters to actually understand and drive these workflows. The third thing is where the magic actually happens. If you look at it, sales is a power law. Ten percent of the people bring ninety percent of the revenue in any organization out there. So if you are able to copy those workflows across the entire organization, then you make the best better and the average better than what they were before. And this is very specific to every organization. There's no way you can build a general platform and then expect people to just adopt.

The next thing is then we have rollout across multiple teams and functions, and then we institutionalize a new revenue motion across the organization. This is how we actually take it from a stalled pilot into a core operating layer across the entire revenue platform. So let's talk about how it works in practice and case studies.

This is Ramp. Ramp, if you know, is the fastest growing SaaS company ever, and even at Ramp we were able to uncover fifteen percent more six-figure deals and ninety percent reduction in meeting prep time and twenty percent more in meetings booked. By the way, Ramp has the ultra dialed in go-to-market motion, and we were still able to do this. The key difference for us is consistency. Every rep gets the same research, same prioritization, same execution quality that is driven by the account level agents that I talked about.

Now across multiple customers we see the same pattern: faster meeting prep, higher conversion, better targeting, shorter ramp times, and consistent execution at scale. So this is how we believe AI agents are going to fundamentally transform the way revenue teams work. We are still at the very beginning of this transformation, and we're doing it one organization at a time. As I mentioned at the beginning of the talk, just like coding agents ten x productivity for engineering teams, we believe revenue agents will ten x the productivity of enterprise revenue teams by giving every account an always-on agent that operates with full context.

Thank you so much for your time today. Thank you so much for listening till the end. I know this is the last meeting for you of the day. I'll be right here. Any questions, I'll be more than happy to answer.

---

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