The challenge
Same account, same question, a different answer every run
Jordan Van Horn, Monte Carlo's co-founder and COO, has asked the same question since the day they started with us: how would a revenue team run if it were AI-native from day one? His answer was ambitious: a system where agents do the work and people direct them. Monte Carlo is running it in production. The hard part is not building agents. It is getting dozens of them to give the same answer about the same account, run after run, without the bill climbing with every question.
"Every AI practitioner has a CFO asking about spend and users asking whether they can rely on the answers. Endgame is how we solve both."
The solution
Every agent reads the same account facts, from Endgame
The GTM Hub, where the agents do the work
They call it the GTM Hub: one governed place where the whole revenue team works with AI. An AE opens a cockpit for her accounts; a CSM sees his renewals, with account health built from every call and email; a manager inspects the same accounts. All three work from the same facts.
Behind those cockpits sit 53 production agents, picked from roughly 300 candidates. They do the account research, pre-call prep, outreach drafts and account-health audits that used to be analyst work.
The Hub's agents run on Claude, but what they know about an account comes from Endgame's graph, through its MCP server. David Leyden, Monte Carlo's Head of AI for GTM, designed it that way once he saw that Endgame's value was the context, not the interface. And it isn't limited to the Hub. Monte Carlo's SDRs live in Nooks, their dialer, where Joe Nicholls, Senior Manager of Revenue Technology and GTM Engineer, switched Endgame's MCP on too. SDRs get the same account context inside the dialer, and Nooks automations check a contact with Endgame before it enters a sequence.
"That's the beauty of Endgame. I can use the context layer at the right time to ask the right questions."
Why Monte Carlo bought the context layer instead of building it
Monte Carlo built its business on data observability. It knows more about data than most of the companies selling to it, so it had to ask whether to build the context layer itself. Jayne Tamboia, Head of Sales Enablement, asked it out loud: if Claude can connect to Gong and Salesforce itself, what is Endgame for?
So they checked. David's team built test agents that read straight from Salesforce and Gong, and their CTO, Lior Gavish, talked it through with ours, Kyle Wild. What they found: getting account data ready for agents means working out which records are the same company, which calls belong to which deal, what the last email meant, and who on the buying committee still works there, then doing it all again tomorrow because it changed overnight. The test agents cost more tokens, gave worse answers, and didn't hold up at scale. And Monte Carlo's engineers build Monte Carlo's product. Nobody wanted them building and running a second one, with its own data feeds, governance and security.
"The context graph is the secret sauce: account reconciliation and context, better than you would ever do it yourself."
They put their team on the agents and the cockpits instead, and on designing the revenue team they wanted. What David's team went on to build is one of the most ambitious things we have seen a GTM team do.
They didn't build it alone. What kept Monte Carlo on Endgame was as much the people as the graph. Endgame has spent years on making GTM agents effective, and its forward-deployed engineers meet with David's team every week to work through whatever comes next. Recently: how to structure an industry brief so agents can read it, whether a health score belongs in Salesforce or the graph, and which model a given workflow needs. In David's words, it's "not just the tool, it's the collaboration, the thought leadership, the willingness to lean in and help us." Monte Carlo pushes us harder than any customer. We like it that way, and it's how we want to work with everyone.
The impact
Consistent, accurate answers, lower AI costs, full visibility
Every agent gives the same answer, proven on live accounts
When Monte Carlo took the Hub to production, they wrote a rule: agents read account context from Endgame and may not swap in another source on their own. For a company that sells trust in production AI, writing a vendor into the rulebook is the strongest endorsement there is.
The rule buys two things: the same answer every time, and a token bill that doesn't grow with the fleet.
But they didn't take it on faith. On five live accounts, David's team had agents write the same pre-call brief two ways, several times each: once from the graph, once from the four most recent Gong calls plus Salesforce. The graph runs came back identical every time: same risks, same buying committee, same next steps, a citation on every claim. The raw runs disagreed about who was meeting whom and who ran procurement.
The graph agreed with itself every time. The raw connections didn't.
Five Monte Carlo accounts, one pre-call brief each, written three to five times per path. Endgame graph vs the four most recent Gong transcripts plus Salesforce.
| Attribute | Endgame graph | Raw Gong + Salesforce |
|---|---|---|
| Consistent across runsSame brief every time | 17 of 17 runs agreed Same status, risks, committee, and next steps every run. | Drifted in 2 of 5 accounts Named different people for the same roles from one run to the next. |
| Every claim citedTraceable to a source | Source on every claim Each fact traced to a call, email, meeting, or document. | No per-claim citations Nothing to trace a fact back to. |
| Memory beyond 60 daysOlder context carried | Multi-month history Procurement posture, build-vs-buy, pricing thresholds, tech stack. | Blind past ~60 days Sharp on the last few calls, nothing from before them. |
| Knows what it doesn't knowStale sources recognized | 27 cited facts, zero from calls No calls in 3+ months. The brief came from recent email, Slack, and docs instead. | Stale call read as current Its newest call was months old. Built the brief from it, unaware newer context existed. |
The reason is what the agent reads. Ask one about a renewal and, from the graph, it gets pre-extracted facts, each tied to the sentence it came from and to the right account and people, not three transcripts and a hope that the model finds the right paragraph. Without the graph, every run rebuilds the account from scratch, and the pieces land differently. As David puts it: "If it's in the graph and somebody asks the question ten times, they get the same answer ten times."
At dozens of agents and thousands of sessions a month, that difference is a CSM changing a renewal plan because an agent read the account differently on Thursday than on Monday. David's summary: raw connections "don't give wrong answers, they give windowed answers, priced per re-read."
50 to 600 times fewer tokens
Everything an agent reads is text the model pays for, and most of a transcript or email chain is noise. The graph strips that out once, when the data comes in, so Monte Carlo pays for it once instead of 2,500 times a month.
Every raw brief pays twice: to find the account, then to read it
Five Monte Carlo accounts, one pre-call brief each. Endgame graph vs the four most recent Gong transcripts plus Salesforce.
Payload the model reads per brief
Time to find the account, per brief
Full account history is extrapolated from the five accounts, not run. Gong pages 100 records at a time with no search endpoint, so every raw run re-finds the account before it reads anything.
The test put numbers on it: a run on the graph read about 1,200 tokens; the same run on raw connections read about 62,000, fifty to one, and that was only the last few calls. To match what the graph knows, a raw agent would have to read every call, pushing the gap closer to 600 to one.
Those are estimates from five accounts, but the direction is clear. Monte Carlo's own write-up puts a production agent run at about 78 cents, and the bill grows with the number of questions asked, not the history behind each account. David's version is shorter: "The more we use it, the more we save, and the better the output gets."
Leadership can see accounts, deals, rep and agent activity
"I can see the accounts and the deals, and I can see what every person and every agent is actually doing with them. For a company that sells observability, that's the bar. It's how we run the revenue team now."
The Hub's signal and customer-health views are built from Endgame context, so when Jordan asks how a renewal is going, the answer comes directly from the source, not from whoever touched it last. And because every rep and agent works from the same graph, leadership can watch the whole system work: what people ask, what agents did with the data, where friction shows up. They sell observability for a living. Of course they instrumented their own revenue org.
Monte Carlo's answer to Jordan's question is in production: agents do the work, people direct them, and everyone is aligned on every account while the AI bill stays "pretty low," per Jordan himself. The rest of the market is still finding its way. Monte Carlo is already there.