The Revenue Harness: Why Context Beats Compute

Julia Nimchinski:

That hour flew by. Thank you so much, what an amazing session, and staying with Erik, we now welcome Ishan, co-founder and CEO of Rox. What a pleasure, Erik. Back to you.

Erik Charles:

Thank you, thanks to Sean, thanks for joining me, you know, You know, we’re gonna have a little bit of fun with this. You know, every conversation at this summit, and I just… we spent a great hour hearing from several, several founders, you know, circling the same question. The models are commoditizing. you know, the… this is magic to hold it, how many tokens, what’s this cost, what’s the consumption, what’s the outcome-based pricing, and even build versus buy is once again back in the conversational path. So… You’re the co-founder, CEO of Rox. Before that, you were at New Relic, you know, you took their self-serve business from zero to $100 million, you did Pixie Labs, you worked on the Siri Knowledge Graph at Apple.

I was talking to Siri, over this weekend as well. I’ve been laughing about it, I’ve got a couple faculty who are staying at my house right now, and we got into this whole question, is at what point does AI actually have emotion and the ability to both understand our emotions and react with emotion and sound like it actually cares? So you’re running revenue agents, and you’ve got them at MongoDB, you’ve got them at RAMP, thanks for being here. We’re gonna… we’ve got 30 minutes, so I’ll just jump right into this, you know, so you talk about your stack of a warehouse-native data foundation, we all need our data lake, we need something to protect.

I mean, I’m sorry, you know, to run a business. The reliability hardness, and then the agents. Everybody else in the marks is talking about, you know, bolting on agents on top of legacy architecture. They talk about the agents, but you started… you start on the different side of the conversation.

Ishan Mukherjee:

Yeah, absolutely. First off, like, thanks for having me, always a pleasure to join this community. And yeah, keep the questions, yeah, as hard and as… and happy to be transparent and candid as we kind of walk through it. So, yeah, we’ve got 27 minutes, and I’ll try to be, crisp here. So. We started the business early 24, and we made the bet on warehouse native knowledge graphs and autonomous agents in late 23, so we were early. Alright, so we met… thankfully, the agent’s bet has panned out, and even my mom knows about it, and the warehouse native context graph is now, like, apparently the most popular term to use.

The question is, like, why did we do that instead of bolting on? It’s because while I held, like, a billion-plus number in New Relic. I was an acquired, kind of technical founder. I realized that most of the operational data that’s used to run revenue, whether that’s, the frontline account teams wanting to know how to grow customers, or leadership trying to drive forecasting and pipeline calls, that data was no longer in Salesforce. It was in the data warehouse, kind of surfaced through a Tableau or Looker dashboard. Mostly because most businesses now no longer do these upfront commit revenue contracts.

It’s more land and expand, more consumption. You need a real-time understanding of what’s happening with customers. So, when we thought about, kind of. agents like coding agents and support agents hitting their revenue organization, we felt like building that on top of a system of record that has already lost most of the context, like a safe but a dead-end road. So we decided to say, hey, like, could we build a warehouse-native revenue system? that aggregates all context in the customer’s warehouse, and we bring to bear all public context, and then agents build on top of that.

So, yeah, that was the core insight, and thankfully those two best have panned out.

Erik Charles:

So, so… So when someone tries to just, you know, run their agents off of those old data warehouses, that old data lake, whatever you want to call it, or data murky pond, we should probably coin that term at some point. What starts breaking first?

Ishan Mukherjee:

It’s very easy to wipe code and, in some way, create, like, a demo agent as a single-player mode. It’s extremely hard to build an agent that can run, like, a hundred million dollar, a billion dollar, like, renewal operation in production. without making errors, and where it breaks is the way you manage context, right? So, the… if you look at building something on top of the CRM, the CRM is a MySQL database, which has account mapping, account hierarchy, relationship data. It doesn’t have unstructured context, it doesn’t have emails. Actually, the full-body email, full-body context, text. Slack, none of that data is there.

So when you ask the agent to say, hey, I’m going into our expansion conversation, should this be a 5% expansion or a 7%? And then give me the, like, reason about the data, and then, And, and then give me, like, a rationale and a slide deck that I can present. You can do one, like, thing, but if you have to run it across the board, the quality is, we publish the public research, it’s extremely poor, and maybe 8 times more expensive than if you had all the data in one place and built a knowledge graph on top, right?

So… Kind of plain English, it’s very easy to build thin agents on traditional CRMs, where you tightly couple the agent to specific data files. It’s very hard to build an agent that That has universal capability across all the context is mostly because context management is extremely hard.

Erik Charles:

I mean, that’s interesting. So, your founding team, you know, is interesting. You’ve got Chris Ray, who came out of the Stanford AI Research on Data Prep, and, you know, my alma mater, so I’m very happy to see that. How has the background of the team you’ve assembled, as much as you’re willing to talk about the team, helped you build you know, that maybe a lot of traditional go-to-market software teams would… might not have attempted. I mean, pulling somebody straight out of the academy like that… And I’m sure you have other examples. That was just one that popped up when I was… I’ll freely admit, I was just kind of going through…

Ishan Mukherjee:

Great question. So Chris and I have worked together for maybe a decade plus. We were on the ground floor of Kubernetes, we were on the ground floor of Amazon Robotics and Siri. We build, I would say, somewhat important technologies. I would say this is the hardest thing to build, so if anybody thinks that this is easy, they’re wrong. Like, this is extremely hard, because you’re trying to build a… production-grade enterprise system… software system for the Fortune 500, when inherently your computer is probabilistic. So it’s an extremely hard system to build. So… The team, we’re about 45 people, all in San Francisco.

It’s an applied AI research and development team, so we do upstream research, like domain-specific research about how to do search, how to do art style generation, even upstream of the two big labs. And then we have a product team that actually builds out the product. Again, the product is, we have a context system, we’ve got an agent system, and the UI. So, all three were very vertically integrated stacks. So, when we compiled the team, the first was to get the platform team and the AI team right. So, Chris and I obviously collaborated for a decade-plus, Then Srinam, who’s our CTO, he was one of Chris’s students, he built Aurora and Kafka, so two, not small, successful technologies.

And we started off with building out the platform layer, which is, like, let’s figure out how do we get all the data in one place, how do we build the graph, it’s governed, secure, and, like, transactional. That was probably, like, we started early 24, so we built… that’s about two and a half years in the making. The agent system got on top. I would say the applied AI work, is more about systems thinking than it is about pioneering, like, AI research. So we have a pretty lean team who’ve been able to kind of, execute. How have we been able to go from Chris, me, Srinam, all the way up to, like, 45 folks?

I would say, like, the core thread is… we fundamentally believe and we realize that we’re building AI infrastructure that runs revenue for the world’s largest organizations. So what Stripe did for payments, or RAM did for finance, we’re doing that in these revenue organizations. So this idea of building infrastructure software that is making decisions and driving is kind of what’s pulling some of the best talent.

Erik Charles:

Okay. So we’ve got, I know, some CROs, logged on who are either listening right now, or they’re half listening, and then they’re gonna later come back to a transcript, or they’ll listen to a recording at 1.5 speed, while working out. Yes, I am totally stereotyping several… based on several CROs I know, but… So, I’m… I don’t even think we need to argue to them that context is a differentiator. They are… you can’t work in the revenue world without understanding that it’s a lot more complicated if it wasn’t, we wouldn’t even pitch, we would all just answer RFPs and either get business or not get it.

So… But the question is, is, like, what is proprietary context? What’s in your graph beyond… you already touched on this a little bit beyond just, say, what’s in the CRM record. Yeah, because you said there’s CRM, there’s Slack, there’s, you know, text, there’s LinkedIn messaging, there… and then… and I’ll even say. You know, how do you maybe even capture, say, trade show hallway conversations in your… in that context?

Ishan Mukherjee:

Yeah, absolutely. Well, I was one of the first customers of Einstein and Data Cloud, and so we understand technically those systems, like, really well. Traditional CRM just doesn’t have unstructured data. It was never designed for it. And, the act of customer relationship management, or doing commerce, or doing sales or marketing of services is… mostly about capturing these unstructured conversations in your primary text, maybe there’s, like, video. And people might think the CRM has a log of emails and chats. It actually doesn’t have anything in a format that you can use, right? So, first thing is, how do you index all unstructured internal context?

Emails, texts, notes, Slack messages? And then link that to your core entities, people, companies, products, events. So that’s the first part. Doing that, is traditionally hard and expensive, but LLMs have gotten really good at doing what’s called the entity resolution. We can throw a bunch of data, and it can figure out that these emails are with JP Morgan, this email is with Nike, and that’s the fourth part. The second is operational data, is when you’re actually driving these massive businesses. 60-80% of the revenue is from install-based growth. Existing relationships, buying more services and products. For that, you need time series data, which is billing, and support, and customer health.

And again, that was never well done in the CRM. So, those two things are what is, in some way, proprietary, that once linked, gives you really rich context. And then second is LLMs are phenomenal at it, because it’s their birthright. to be good at unstructured data and managed data, right? Like, it being, like, relational data is, like, fine. So I would say those are the two things where we started off with. So, now when you think about, like, an agent that’s somewhat sentient, like, our agents are benchmarked, like, 91st percentile, like. PhD, like McKinsey consultant at this point.

It has all your unstructured data science data. You do have to enrich it with public context, so having the agents go to the internet, do scaled web search at the company level, and then aggregate contact data and use that to enrich the graph is extremely important. So, that’s what I would say. Unstructured data, operational data, and then, like, public data.

Erik Charles:

So, what’s some of the stuff that you’re willing to reveal, admit to, that’s the stuff nobody thinks of in revenue systems that matters?

Ishan Mukherjee:

Yeah, Look, like, our experience comes from running public companies, holding a number, being in the engine room for a decade, decade plus, so CROs on the call, RevOps people on the call, you all know that there’s, like. Maybe 5-10% of the people who, kind of. lack of a better word, like, have the feeling of a quarter-to-quarter noose on your neck, you gotta hit a number. It’s a very intense responsibility, and it’s… and those folks are usually not on LinkedIn, kind of posting and posting a bunch, so we come from that kind of DNA. So, for us, the most impressive things is that agents are making your future-proof, quota-carrying employees dramatically more productive.

And then the number of people, or the ratio of people you need to support them has compressed by maybe, like. 5X, or maybe, like, 7X. So what do I mean by that? So we have a large Fortune 500 medical hardware company, that uses, us to basically run their entire revenue system. Our agents are now finding RFPs, feeding that to the account teams, so that the account teams can go in at the right time, the right person, with a very knowledgeable kind of talk track and presentation that’s generating about $100 million more of order volume. In a given month, right?

So that’s about domain-specific frontline sellers who are now executing at an elite level. Now, as they’ve been scaling, their revenue per employee has been growing. now, like, the question is, do you need BDRs and sales analysts and all the folks who support them? The answer is, yes, you need them, you need fewer of them. And the idea is, like, how do you train them to be… to go from Waymo, kind of Uber drivers to, like, Vaymo orchestrators. Like, how can they have, like, a swarm of agents that they’re using to orchestrate, so… which is doing a lot of work, kind of feeding through the account team?

So, again, like, these agent systems have to do work. we hold accountability in terms of driving revenue outcomes. The revenue outcomes are coming from really, like, real economy. Fortune 500, like, global 2,000 customers who are not just, like, tech companies in the East and West Coast. So that’s probably, like, the most interesting thing.

Erik Charles:

I do enjoy… one of my favorite things is… I’ll quote Trevor Loy, who runs a venture firm out of New Mexico. Excuse me, and his line has always been, you know, fish where nobody’s fishing. I love selling to people that aren’t just selling to another tech company or another, you know, the East Coast company. I love having old-school firms, you know, that… that are ignored, quite honestly, by a lot of the development that comes out of Silicon Valley and its equivalencies. So everyone’s data’s a mess. And we’re back to the, again, context, because I think that’s… I think this is very interesting.

One of the things that AI’s doing for us, more than anything else, I feel, is taking care of tons of horribly unstructured data, or dirty data. that is out there. I’m seeing data integrations, like a variable we use in AI to help us with the data integrations for managing sales commissions, and it has had a huge impact on how fast we can get up to speed. which is why I wrote this question this way, which is, you know. What does a company has… how far back can the company go in its data? How quickly can they assemble this with your system?

And, you know, until they have sufficient context that they can actually have a time to value?

Ishan Mukherjee:

Great, so we don’t get to stay in the room. We’re usually pulled in by the C-suite, the CRO, the CIO, the CFO, the CEO, and we don’t get to stay in the room if we don’t… we are not the fastest time-to-outcome vendor, because we compete with Agent Force head on. And then we try to accelerate, kind of, internal build, so we are the fastest time to outcompete. kind of vendor. In that case, it’s, it’s, we commit to weeks, and we try to get there in days. So, our platform can connect internal data, bring in public data, and create the graph, and have these agents start running. pretty much out of the box.

What we do is we might add custom integrations if there’s bespoke data sources, or kind of connect, tune the agents, but the platform is up and ready, so if you go to rocks.com, you can just start using it. I think it was the only agent platform that is fully self-service. It’s mostly because the way to think about it, Erik, is you obviously have a nuanced understanding, is you can either go recruit, McKinsey consultants and the best, folks and have them start working for you, or you can… in some lack of a better word, like, birth a rep, school them, train them, and it might be 9 months, RAM, might be 12 months RAM, and… And by that time, the boards moved on, right?

So that’s kind of our place. So, our core IP is to have the platform work out of the box, and its proof is… It’s open. Go to rocks.com and try it. And then once it’s out there, is how do we, with our kind of deployment team, make sure we take ownership of a number and drive outcomes? rubric is we want to be in a position where after somebody kind of starts working with us, by the next board meeting, we need to have a board reportable VIN. And that’s generally, like, how we are gold ourselves. And just to know, like, how these… we’re an agent company, so agent companies are not seat-based, they’re not license-based, we’re usage-based, so… and the… the… the commits are based on the customer seeing outcomes, right?

So, if… If we don’t hold the number and drive outcomes, we don’t get to stay and grow. So that’s in some way, like, is pretty value-aligned. But the hard part is to get the agents to actually do the work.

Erik Charles:

Now, I’m gonna go off-script here. So, you’re a usage-based, which, that’s what a lot of people are dealing with right now, you know, how do we price, how, you know, so it’s a token-based, you know, some sort of a, you know, to you, you know, rolls up. How close are you to outcome-based pricing?

Ishan Mukherjee:

Yeah. Again, we, just to give you context, we started to work on this problem in late 23, early… so Sequoia led the seed round back in February of 24. We started off with warehouse native agents with usage-based pricing. So, we’ve been at it for a while, and then through the two, two and a half years, the LinkedIn speak has gone from. every variant of pricing and packaging, and it’s really interesting. My general view, like, right now is that pure outcome-based pricing is not yet feasible in pre- and post-sales agent workflows. Because pre- and postal agent workflows are inherently open-loop, non-verifiable domains.

It’s not like you reach out to somebody, close the deal, and you can attribute value. There’s… it’s open-looped, there’s time lag. It’s not like filing a PR or doing a res… directing a support ticket, right? Like, it’s somewhat open-looped. So I think maybe we can get there where agents are talking to agents and they’re selling transactionally. We’re not there yet. So what do we do is… Our sales team goes in and sells on business outcomes and secures a budget based on ROI. They would say, okay, we will help you generate $100 million of RFP wins, or we will help you drive 5% uptake on your renewals, that’s going to have this much top-line impact, and in return, or on a 30 to maybe 50 XRI basis, we would love for you to kind of commit a certain volume of consumption.

Then the platform is used on a pay-for-use basis, right? So as you’re using it, you’re using up these credits. So that’s the thing, is value-based selling to secure budgets and outcomes, and those budgets used on a pay-for-use. Over time, I do see, I would say, mid-term. we’re getting to a point where outbound agents are getting to full autonomous. I would say it went from 0% to 40% of our actions are now fully autonomous outbound agents. Outbound agents reaching out to customers, booking meetings, creating kind of opportunities, which is pretty phenomenal. It started, I would say, February was the inflection point, and now It’s at scale.

I think if people come in and just want to use the outbound agents, we’re getting close to outcome-based pricing, but not there yet, but it might happen very soon, yeah.

Erik Charles:

I do think there’s… and I’m facing it at variable as well. It’s like, you know, how do I shift my pricing model from a per seat to a consumption? or even just, like, an active rep, or something like that. I think the… it’s gonna be… well, now… We keep on getting announcements, I see a text, or a Wall Street Journal piece, or even Wired will say, you know, so-and-so’s released a new model, their latest model. I was… it was mentioned, I was playing with Claude this weekend, and it was like, which one of these five models do you want me to do this with?

You know, you’re being lazy and doing a recipe, so I’m going to give you this model, because you really are just asking me to scrape the web, but other one, you know, when I was asking Claude about emotions, it went a different thing. So, how do you protect yourself, is where I was really going. You know… There’s… the labs are gonna keep on releasing smarter and smarter models. It gets easier and easier to build agents. Vibe coding’s a term, but, you know, I… you know… But, you know, there are people, like I said, there was a company I was speaking to recently that is creating their own CRM.

They’ve just decided to, you know. they’re already an AI user, like, why don’t we just use it to build a CRM, and we think we can have it build it, and it’ll still be cheaper than what we’re paying, you know, one of the… a certain big player out of San Francisco. So how… how do you protect yourself? What have you got that can hold its value that… it’s not gonna be an easy overnight. If I’m a Fortune 500 company, why am I not just telling my IT department to go and develop your product?

Ishan Mukherjee:

Yeah, makes sense. And then, just to replay badly, what protects me as, like.

Erik Charles:

Yeah.

Ishan Mukherjee:

Founder, CEO of Ross, yeah. I think, like, a lot of people, like, frame this as the moat question. Like, what’s the moat? Like, how can you… it’s a really good question. To answer that, like, we have to empathize with where we’re at in the technology cycle. Like, the marginal cost of building software has, like, really gone down. Like, people can… and build stuff to get up front. And then the horizontal capabilities is rising. Like, the big labs and what you can do is rising. So as a vertical domain, like, what we challenge ourselves is we have to do the full work.

Like, the entire work. Like, coding agents have to actually ship the whole product, support agents have to run the delivery center, we have to run revenue operations. As a vertical player, like, for us to stay, to be the leading player, and be dominant, like, we have to push ourselves to do the end-to-end work. So. ultimately, for a customer, like, the moat is you could go back, work with the incumbents, like the large San Francisco company with a skyscraper. Second is you could build, in-house with the IT team. Or do you want to push accountability to a vendor whose life depends on it?

And why is the accountability important? It’s when you have software doing the work. There is liability, there’s quality, there’s cost, the usual kind of considerations. Which is not a 5% engineering effort. Like, we were doing this as a public software company, it’s at a very minimum a 65% engineering effort. So, will companies deploy 65 engineers to run revenue systems? Will 65 engineers to run support systems, and finance systems, and legal systems, and coding systems? Yes, you could do that. I think… so where we’re in the cycle is the bars getting better, people can breathe stuff.

Now. there’s an exuberance that we’re gonna land on a steady state where people are going to decide, like, what to build versus not. We really focus on the segment of the Global 2000, which are hardware, manufacturing, life sciences, where, the outcomes are really, really high, the liability is really high. And, and then that’s where our moat is to essentially become, like, an embedded, kind of… human resourceful, where our agents are doing work, and we’re taking on the liability. I think that’s where most vertical players will go eventually, is essentially become, kind of, teams of those businesses, and… and… and time will tell.

I think on the moat piece, nothing is permanent. Like, we’re all riding this massive once-in-a-generation wave, and our job is to build the best surfboard and hire the right surfers, but… but the wave ebbs and flows, right? So these questions definitely evolve.

Erik Charles:

You won me with the surfing context, even though I’m a terrible surfer. I’m very good at falling off of a board and or trashing it because I, you know, go in the wrong direction and hit something. So you build a… so, you know, I’ve seen numbers of, like, you give time back to reps. Goodness knows I’ve used the line of, I give X hours per person. We’ve all built ROI models over our time in Silicon Valley of, we’ll give you this much time back. hump… you know, are you seeing that productivity, and what are companies doing with that quote-unquote time?

What are you seeing the reps do? If we suddenly give the reps, like, say, 8 hours a week back. Are they selling more?

Ishan Mukherjee:

Yeah, yeah, or they’re playing golf, yeah. That’s the question you get. I think selling on, hours saved per employee is just oversold. Like, if you look at the global, like, Microsoft sales team, they’ve been selling that to attach, Copilot for about two and a half years, so we’re just in a two and a half, three years sales cycle where buyers are fatigued with that value proposition. So we kind of, very consciously try to steer away from it, is… we try to anchor on top-line impacting metrics that are pre-existing and being discussed at the C-suite level.

It could be revenue per employee, it could be pipeline generation, it could be deal execution, like, there’s only 2 or 3 metrics that the board and the C-suite cares about. That is where we are seeing, obviously, massive outcomes. It goes from driving the whole RFP pipeline for people selling into the public sector, to really high-scale companies who are driving they’re directing maybe 5-7% of their payroll to us, and then we’re driving 36% of their net new business, right? So really thinking about those metrics are what matters to us. Like, if you’re not helping our customers grow revenue, then we’re not doing our job.

Erik Charles:

Yeah. I get it. How should people be thinking about turning all this raw data context into a knowledge layer, versus just… You leave it in the lake and let the AI dig through it.

Ishan Mukherjee:

I think if you leave it in the lake, the air, dig through it, you can get plenty far if you’re an early-stage company, or if you’re doing, kind of, a proof of concept. it can get brittle and expensive really fast, so ultimately you do have to build, like, a knowledge graph with governance and SSO, so what can they do? They can obviously try to build that on their own, or… and they use, kind of platforms like Rox, which have, in some way, solved it, and… and effectively, it’s enabling technology for us, so… so you get that out of the box.

We… we only… get paid when the agents are delivering, so the graph, the enrichment, all that goes in is gone in. I obviously haven’t worked on knowledge graph for about a decade. I think it’s easy to start, hard-to-scale problem, and probably not the most… it’s not core to every business, right? So, it’s a software problem, it’s a scale problem that’ll get solved by vendors.

Erik Charles:

Alright, very quick, because we’ve got about 1 minute left. Two years from now, what’s the data in life of an account exec who’s using your tool? Or whose company has your tool, perhaps?

Ishan Mukherjee:

Yeah, yeah, I think as revenue agents are increasingly pretty standard now, I think, you know, future-proof enterprise threat AEs are still very, sought after in the market. They’re probably supporting two times more customers than they are today. And then, the folks that they relied on, sales analysts, BDRs, RevOps engineers, like, they just need fewer of them. What it would mean is, as a percentage, you would have a lot more quota-carrying employees than before. And your deal cycles will be much faster than before, right? So, and I think the win rates are the best way, so it probably stays consistent, so, so leaner, more concentrated talent pools, like, just moving a lot faster.

Erik Charles:

Great, thank you very much. Ishan, co-founder, CEO of Rox, I appreciate your time, and the summit will continue now with Wade Foster from Zapier, Kady Srinivasan from Freshworks on moving from individual AI to an institutional AI. Julia, passing it to you.