At Ramp, 300 to 400 sales reps can build their own tools using shared customer data. The central team learns from what they build and brings the best ideas back into its app. Keyan Sarrafzadeh explained how this works at the GTM + AI NYC Lightning Talks.
Notion takes a similar approach. Its team keeps customer information in shared hubs, and reps use that information to build agents for their own work. Julia Biedry Gonzalez described this as "raise the tide for all ships". The team supplies the information without trying to design every rep's tools.
Chris Prinz from Modal summed up the idea as "centralize the data and decentralize the agents." The details differed across the talks. Ramp records why agents use each tool. Notion links summaries back to their sources. Modal checks the quality of the data before putting it to work.
Deepline hosted the event at Ramp HQ in New York on June 30, 2026. Speakers from OpenAI, Ramp, Notion, Modal, Attention, and Deepline gave six talks over 91 minutes.
The talks
- Jai Toor of Deepline tested data providers against examples of the results he wanted.
- Keyan Sarrafzadeh of Ramp showed how hundreds of reps build tools using the same customer data.
- Chris Prinz of Modal explained how he organizes and checks customer data before agents use it.
- Julia Biedry Gonzalez of Notion showed the customer hubs that sales reps use to build agents.
- Bryant McCombs of OpenAI explained where he uses fixed rules and where he reviews an agent's work himself.
- Jacob Fleisher of Attention tested rep-to-buyer matching against historical deal data.
Watch the full playlist
Watch all six talks on YouTube
Jai Toor, Deepline
"Your ideal data waterfall is unique to you." Jai's answer to the question of which data provider to buy was to test providers against the contacts and companies you need.
A data waterfall tries several providers in sequence until it finds the missing information. Jai gives a coding agent examples of the results he wants, then has it test different providers against those examples. He reported reducing a four-week evaluation to 30 minutes.
"You define the outcome of what good looks like. You let the systems programmatically find the patterns."
He also combines public and private data for each use case. The results still surprise him. "I'm honestly like surprised it keeps working at this point."
Video not loading in LinkedIn or your in-app browser? Watch it on YouTube.
Keyan Sarrafzadeh, Ramp
"How Ramp let 400 reps build their own custom GTM tools." Keyan showed how Ramp gives reps room to build without making each person gather and maintain their own customer data.
Ramp has a shared app and an MCP server. MCP is the connection that lets a rep's coding agent use Ramp's data and tools with the right permissions. Reps build what they need, and the central team brings useful ideas back into the shared app.
Keyan described the goal as "we want to empower everyone to build their own tooling / And then bring those learnings back / Into the centralized solution."
Each time an agent uses a tool through the MCP server, it has to record why. Those explanations help Ramp understand what reps are trying to do and which tools they need.
Ramp only builds in-house when it expects a competitive advantage. As Keyan put it, salespeople are the most rational users on earth because there is a direct link between their work and their paycheck.
Video not loading in LinkedIn or your in-app browser? Watch it on YouTube.
Chris Prinz, Modal
"How Modal built one GTM data pipeline to rule them all." Chris wants agents to work from the same prepared data. He summarized his approach as "centralize the data and decentralize the agents."
He keeps the data processing in one dbt project, where he can check and debug it together. His second rule was "build a state machine not a router." A state machine records which stage a job has reached and what can happen next. That makes the sequence explicit instead of relying on agents to keep handing work to one another.
Chris has worked on data quality for years, "which makes me very fun at dinner parties". His advice starts with reading the data yourself. Pick 100 rows at random and check them by hand.
He also compares providers by the cost of finding a prospect who fits the business, rather than by the cost of a row. Cheap records do not help if they describe the wrong companies. As he put it, "Garbage in and garbage out is what people say around machine learning systems. It's really the same here."
Video not loading in LinkedIn or your in-app browser? Watch it on YouTube.
Jai's Deepline talk referenced the same idea from the other direction, so Chris's pipeline talk and Jai's waterfall talk pair well.
Julia Biedry Gonzalez, Notion
"How Notion built customer hubs: one shared context layer for reps and GTM agents." Julia's team gathers customer information into shared hubs. Account executives can then use that information in agents they build for their own work.
Julia's line was "good context is the bottleneck for accurate fast and efficient agents." Her team handles the shared information without trying to anticipate every rep's questions or tasks.
The hubs collect data ahead of time, so an agent does not have to fetch it all when a rep asks a question. As Julia put it, "we're kind of like converting GPUs to CPUs here."
The agent also has its own access permissions. In her words, "you can literally think of it as another collaborator in notion, so it's like a person with a seat." When it writes a summary, each claim links to the original record. A rep can follow the link and check what the summary says.
Video not loading in LinkedIn or your in-app browser? Watch it on YouTube.
Bryant McCombs, OpenAI
"Run your GTM through Codex." Bryant described using Codex to work through Salesforce tasks until colleagues began coming to him for Salesforce help.
He uses the agent to prepare work, then decides which steps need fixed rules so they run consistently. He described this as adding determinism, saying, "you have to find the highest leverage points in a workflow and figure out how do you add determinism to those."
He also stressed personal accountability, saying, "regardless of how many agents you have running for you, ultimately it's you." He wants the agent to prepare work for his approval. "I don't want the signal, I want the action to have already taken place and then I want to know whether or not I should approve that action."
Video not loading in LinkedIn or your in-app browser? Watch it on YouTube.
Jacob Fleisher, Attention
"Routing leads on interests, not territory." Jacob wondered whether matching buyers with reps who shared something with them would help close deals. He described his hunch as "at the end of the day what I've learned is people buy from people they like."
He tested it against past deals in Salesforce. One rep, Noah, has a UK passport and a higher close rate in countries where he holds citizenship. Jacob combined Salesforce, Attention, and Deepline data in Claude Code to find those matches, then added them to the team's sales process. He reported a 30% increase in win rate and a 25% reduction in time to close.
He opened with the story behind his buzzed head. He told his CEO he would buzz it if they raised a Series B, and the CEO said "no, when."
Video not loading in LinkedIn or your in-app browser? Watch it on YouTube.
Watch all six talks
The GTM + AI NYC playlist has all six.