“Where does the Nile begin?”
For most of recorded history nobody in Europe knew. Then, in the span of a hundred years, four men answered, each with a different place. One came up through Ethiopia and said it was there. One went in from the east coast and said it was the lake he had reached. His companion walked three weeks further north and said it was the next lake. The fourth went looking far to the south, along a river he followed until it killed him, and that river turned out to be the Congo.
The Nile has one source, the explorers identified four, and the difference between them was which way each man had walked. They were all correct in what they saw. But each was wrong about what he hadn’t seen, and none could tell the difference from where he stood.
Sales teams have their own version of the Nile question.
“Where does this deal stand?”
The answer depends on who you ask. The rep says one thing, the manager hears another, the CRM says a third. The truth is in there somewhere, spread across the CRM, the recorded calls, the email threads, a Slack channel, the meeting notes and the shared documents. And no one person has read all of it.
AI was supposed to solve this. Connect an agent to all the systems, ask the question, and let it rip. But what does the agent do? Like the intrepid explorers searching for the source of the Nile, it picks a way in, wanders around, and stops. It then reports from its final location, confident of what it saw and unaware of what it did not. Ask again and it goes in through another path, stops somewhere else, and arrives at a different answer. So how in the world is a rep expected to rely on what the agent tells them?
We call this the wandering agent problem.
How an agent wanders
When an agent is asked where a deal stands, the first thing it decides is where to begin. Then, based on what it sees in that first step, it decides what to look at next, and based on that, what to look at after. At some point it decides it has seen enough, and it stops.
You can see this in the animation below. Go ahead and click ask. Watch the agent wander through raw records to answer the question. Then ask again and see what happens.
The answer depends on the route, not the evidence. That’s why connecting an agent to all the systems and letting it search doesn’t settle anything. It can’t. The agent has more evidence than it can see at once, so it only sees some of it. Which part it sees depends on the path it followed and where it stopped.
Give your AI a map
What settled the debate over the Nile’s source wasn’t yet another explorer with a better route. It was a journalist who, in 1875, surveyed a lake connected to the Nile and, by sailing all the way around it, confirmed only one river outlet. That is where the Nile begins. Thereafter, the answer was on a map. And with that map, nobody had to go and find the source themselves.
At Endgame, we build maps for GTM AIs.
The map we build today is a context graph: every account, person, deal, call and email is a point on it, and every fact is attached to the points it is about. Our machines read the calls, email, Slack messages, notes and documents connected to an account, each record in its entirety, and pull out the facts, along with the exact words each fact came from.
For every account, we compile the facts that matter most into cited sections, such as where a deal stands, who is on the buying committee and in what role, what’s in flight and what’s blocked, who owns each piece and when it’s due. Every few hours, we go back and redraw what changed.
When your AI has that map, it can look up what it needs, and it gets the same up-to-date information every time. No wandering around, no stopping randomly. Just the same path each time, and the same answer each time.
Experience the difference below. Click ask. The sources are read and compiled into the map first, and only then does the agent go in, by one short route. Ask again and the route is the same.
When your AI has a good map, it knows what to do. Everything it needs is readily available. Its answers are consistent and correct. Finally.
Making maps for Monte Carlo
David Leyden runs AI for the revenue team at one of our customers, Monte Carlo. With more than 50 agents in production across the entire GTM motion, he experienced the wandering agent problem firsthand. Before he trusted any of the outputs, he tested the agents by having them write the same pre-call brief for the same accounts, over and over. Some runs read the raw sources. Some read Endgame’s map. Then he compared what came back.
What he found could have been said of the Nile explorers: raw connections “don’t give wrong answers, they give windowed answers.” The agents’ views were true as far as they had looked, but silent about everything past the edge of their known windows.
His verdict on the version with the map, the Endgame version, was simpler. “If it’s in the graph and somebody asks the question ten times,” he told us, “they get the same answer ten times.”
He made it a standing rule. His agents read the map, and nothing else. The rest is in their case study.
Stop exploring. Start navigating.
For any agent you build or buy, find out who decides what it reads. If the agent decides at question time, it’s an explorer, every ask is an expedition, and someone has to check every answer that comes back.
In go-to-market, most questions are about the same few things: an account, a person, a deal. Most of the answers can be compiled before anyone asks. Sail the whole shoreline once, so nothing is skipped. Put what you find on a map that every agent reads. Then the only thing that can change the answer is the evidence.
If you’d like us to help build your maps, we’d love to hear from you.