Loading page…
A warehouse-native agent platform that does the grunt work of enterprise selling, refuses to charge for licences, and two years after founding has 35-plus B2B customers including OpenAI, Nvidia and MongoDB.
backed by Sequoia Capital, General Catalyst and Google Ventures · founded 2024 · priced on meetings booked, pipeline generated and opportunities closed rather than seats · no revenue, ARR or funding amount disclosed
"We are here to sell work." Rox refuses to charge for seats, usage or consumption, and instead gets paid on the same numbers the customer's own reps are comped on — reply rate, meetings booked, opportunities closed. The deliverable is the product; the software is just how it gets made.
Fewer bars = easier, cheaper, or faster for an AI-assisted solo builder. Editorial judgments based on the case details.
Rox builds what its cofounder calls revenue agents: software that does the grunt work of enterprise selling end to end, rather than one more dashboard a rep has to log into. Shriram Sridharan — Ram — is cofounder and CTO, a former founding engineer on Amazon Aurora and an early engineer at Confluent. The company started in 2024 on two bets. The first was that every large enterprise is now AI-transforming whether it wants to or not. The second, and the more specific one, was that the data those agents need has already left Salesforce and Zendesk for the warehouse, because consumption-based contracts forced companies to tie product usage to account records and build a customer-360 view somewhere central. It did not begin as an agent company. Rox started as Rox Data Corp, aggregating data out of a customer's scattered business systems and selling a knowledge graph and a governance layer on top. The agents came later, sitting on that foundation — and Ram is blunt about why the ordering saved them. Without the data context, he says, "we are one of those startups that got killed by Claude Code." The model is someone else's; the context is yours. Two years in, the host's introduction puts more than 35 B2B companies on the platform, including Ramp, OpenAI, Confluent, MongoDB and Nvidia, with Sequoia Capital, General Catalyst and Google Ventures behind the company. No revenue figure appears anywhere in the interview. What does appear is the commercial posture, which is the genuinely unusual part: Rox will not sell seats or consumption, and takes its money against meetings booked, pipeline generated and opportunities closed.
The first obstacle in the Fortune 2000 is not price, it is the data.
The quick-reference card — acquisition channels, replication playbook, risks & traps, and our verdict — is members-only.
Join the Unicorn Club — $5/moCancel anytime, in one clickPaste into Claude Code / Codex and get a working version of this product end-to-end.
Two production-grade prompts per idea — a full build spec and an SEO growth plan — plus 10 tailored build specs a week. Members only.
Join the Unicorn Club — $5/moCancel anytime, in one clickEvery claim traces back to a listed source — open any of them and check the number yourself.
Data credibility: A founder interview on a sponsored B2B podcast, with no financial disclosure of any kind. No revenue, ARR, contract value, funding amount or headcount appears anywhere in the episode. The traction claims — more than 35 B2B customers, the named logos Ramp, OpenAI, Confluent, MongoDB and Nvidia, and the investors Sequoia Capital, General Catalyst and Google Ventures — come from the host's introduction rather than from the founder, though the founder accepted the introduction and discussed three of those customers himself. The customer outcomes are vendor-described: the 'multi-millions' of RFP opportunities at WSP, and the claim that MongoDB's BDR outbound is measured on reply, bounce and open rates, are Rox's own account of its own results with no customer present to confirm them. The episode is produced by a podcast agency that sells podcast production and is itself the sponsor, so it is a promotional format. Two names are uncertain in the auto-captions: the customer rendered as 'Binder', and the cofounding professor rendered as 'Chris Ray'. The product name is spelled 'rocks' throughout the transcript; the host says the domain aloud as rox.com. The auto-captions also render 'Claude Code' as 'cloud code' throughout, including inside the founder's line about being killed by it.
The first commercial product was not an agent at all. Rox Data Corp sold the unglamorous layer — data aggregated out of a customer's scattered systems, a knowledge graph over it, governance around it.
Three loops reinforce each other. The open self-serve product is the first: keeping it public forces the team to build something a stranger can understand, which stops the enterprise version becoming what Ram calls a Frankenstein of bespoke integrations, and it drops inbound like WSP into the pipeline for free.
Ram's own version of that stretch lasted a decade. He came to the United States wanting to start a company and spent roughly ten years unable to, held in an immigration queue that required him to stay employed — eight of those years at AWS building Aurora, then Confluent.
Rox is an agent orchestration layer sitting on the customer's own data. Underneath it is a knowledge graph stitched from three kinds of input: private structured data from Salesforce and other business systems; private unstructured data such as call transcripts, emails and sales enablement material; and public unstructured data — funding news, published RFPs, signals scraped from the open internet. A governance layer sits above that, and the agents above the governance. The agents come in two modes. Foreground agents answer a request: build this QBR deck, prep me for this meeting. Background agents read the calendar, deliver the prep the rep should have done, and update the data systems afterwards so nobody opens the CRM tab. The span is deliberately end-to-end across the revenue lifecycle — inbound lead quality, outbound, qualification, closing, expansion, churn prevention — because the pitch is the finished work rather than a productivity uplift. Ram's framing is that coding agents and support agents improve the bottom line, while revenue agents aim at the top line. Three customer examples appear. MongoDB uses it for top-of-funnel automation, personalising BDR outbound at scale from warehouse signals plus public data. WSP uses it for RFP automation: detecting Canadian public-sector requests from public data, filtering them against business rules and drafting the response, in a process previously run by a large team reading government websites by hand. And a company the transcript renders as Binder uses research agents to find domains a competitor has stood up, infer where that competitor is selling, and run a takeout campaign into those accounts.
Rox sells outcomes. Ram is explicit that the company chose not to charge for seats, usage or consumption, and instead binds itself to the metrics a revenue organisation already lives on: reply rate, bounce rate, open rate, meetings booked, pipeline generated, opportunities closed. For the buyer that converts an AI purchase from a bet into a line item with a return attached. For the vendor it means revenue only arrives when the agents actually perform. Above that sits the free self-serve tier, kept open on purpose. It is not a monetisation strategy — he says plainly that it exists for feedback and for keeping the product coherent — but it functions as the top of the enterprise funnel, which is how a 50,000-person firm arrived without a salesperson touching them. Our read: the cost side is where outcome pricing gets hard. Forward-deployed engineers sit with each customer to learn a workflow that differs at every account, which is a services cost carried against software revenue. That works while the customer count is in the dozens and each engagement teaches the product something reusable. Whether workflow learning generalises fast enough that the hundredth customer needs less hand-holding than the tenth is the question no interview answers, and it is the one thing a reader evaluating this whole category should watch.
The architecture is the business. Data lands in the customer's warehouse or lakehouse — Ram's founding observation is that this migration was already happening everywhere, not only in B2B SaaS, because consumption contracts made account executives dependent on product-usage data that lived nowhere near Salesforce. Rox builds a knowledge graph over it, adds governance, and runs agent orchestration on top. His argument for why that ordering is defensible: the compute is somebody else's model, so the data context is where the moat sits in any company of this kind. Two cost centres get named. The first is deliverability, handled internally, including rotating multiple sending domains. His view is that volume is not the lever — it is the timing and the content — and he gives the arithmetic behind signal-based prospecting: three replies from ten well-chosen people and thirty replies from a thousand cold ones come out at the same rate, with very different costs and very different reputational damage. The second is the forward-deployed engineering team, which is headcount-heavy by design. Our read: gross margin here will look more like a data platform than a seat-based SaaS. Inference and warehouse compute scale with usage, forward-deployed engineers scale with customers, and outcome pricing means a quarter where the agents underperform is a quarter the vendor eats. That is survivable at enterprise contract sizes and impossible at SMB prices, which is a large part of why they went up-market in the first place.
Nobody reading this is about to raise from Sequoia and sell agents to the Fortune 2000. The copyable piece is smaller and sitting in plain sight: the WSP use case. Public-sector RFPs and tenders are published on public websites, read by hand, in every country, and in that game the first credible response often wins. A solo builder can run the same loop at niche scale — watch a defined set of public tender sources, filter against one buyer's business rules, alert them the day something matches, and draft the first pass. No warehouse, no data connection, no security review, because none of the input is private. That is the most under-rated thing in the interview. Copy the land motion as well. Proving value on public data first is the answer to every enterprise security objection, and it scales down perfectly: show a buyer something true about their own market before asking for access to their systems. Change the pricing. Outcome-based contracts are survivable when investors are funding the gap between work delivered and money collected; a small operator should charge a flat monthly fee for the same work and keep outcome bonuses as an upside clause rather than the whole basis. And change the scope. Rox's end-to-end ambition needs forward-deployed engineers at every account. One workflow, one vertical, one repeatable deliverable is the version a two-person team can actually run.
Ram names his own biggest risk unprompted: a company in this space that owns no data context is "one of those startups that got killed by Claude Code." The horizontal models keep absorbing capability, and any agent layer whose only asset is prompt engineering gets flattened. His defence is the knowledge graph over the customer's own warehouse, which is a real one — but it also means the product's value depends on the customer's data being good and connected, while the same interview explains that customers refuse to connect data until they trust the vendor. The open product is a deliberate exposure. Competitors can see it and copy it, and his answer is that being out-executed would be their own fault. That is a confident position rather than a defensible one. Then the structural risks. Deliverability sits on top of email providers whose spam policies can change overnight. The CRM incumbents are shipping their own agents and own the system of record. Outcome pricing means a customer whose numbers dip has a ready argument about the bill. And forward-deployed engineering is a services cost that can quietly turn a software company into a consultancy. Finally the evidence. Nothing in this interview is a revenue figure. The customer count and investor list come from the host's introduction, and the customer outcomes — "multi-millions" in RFP opportunities at WSP — are the vendor describing its own results on a sponsored podcast.
The four founder playbooks — first customers, first dollar, the flywheel, the silent stretch — are in the Unicorn Club, on this idea and on all 1,403 others. The story, the numbers, the sources and every later section stay open to everyone.
Join the Unicorn Club — $5/moCancel anytime, in one click