# Using AI to run founder-led outbound for 10 YC companies Canonical URL: https://pumpgtm.com/blog/ai-founder-led-outbound-system [← All guides](/blog) # Using AI to run founder-led outbound for 10 YC companies AI made the thinking part of outbound cheap. The hard part is the system around it: one lead universe, shared prospects, copy measured by meetings. Namanyay Goel · September 9, 2026 · 8 min read Contributors: Namanyay Goel, Aditya [Original source](https://x.com/NamanyayG/article/2097783972302557668) · 2026-09-09 · Adapted and expanded for PumpGTM AI made the thinking part of outbound much cheaper. Researching an account, understanding what a company sells, finding an angle, turning a founder's notes into a decent message, summarizing a reply, drafting the follow-up: all of that now takes seconds. But generating text was never the hard part. The hard part is running the system around it, and that is where most founder-led outbound still breaks. I run PumpGTM, and we run GTM for 10 YC companies. As a YC founder myself, I had to figure out how to sell while shipping, fundraising, onboarding users and answering every sales conversation personally. Nobody prepares you for that part. This is what we learned, first about where AI actually saves time, then about the system we ended up building because the tooling did not exist in one place. ## Where AI consistently saves time in outbound Three places, and they are not the ones most tools advertise. ### 1. Turning account signals into a messaging angle We give the model everything we know about an account and ask one question: why would this specific person care right now? We are not asking for a LinkedIn message yet. We are looking for the reason a conversation should exist. If that reason is weak, twenty personalized versions of the message will not fix it. This is the step people skip, and it is the step that decides whether the rest of the sequence works. ### 2. Writing variants instead of one perfect message We rarely trust a single outbound message. We would rather run three hypotheses at once: - one leads with the pain - one leads with the existing workflow - one leads with a result Then we send all three over the same period and measure what happens. Not opens, and not whether someone on the team likes the copy. Replies and meetings. We built the A/B testing into the sequence itself so nobody has to run it by hand. ### 3. Handling replies faster The most useful and most underrated tip: reply fast to every question that comes in. A founder running outbound gets a mix like this in one afternoon: - interesting, how much does it cost? - we already use a competitor - not this quarter - can you speak to my head of sales? These should not be treated as the same reply. Our AI classifies the intent, summarizes the conversation so far and drafts the next response. The founder spends time deciding, not re-reading eighty threads. That is roughly 30 percent of the system. The other 70 percent is everything needed to make those decisions happen reliably. ## The system we ended up building Running founder-led GTM, we kept hitting the same wall. The tooling was fragmented: one place for prospecting, another for LinkedIn, another for email, another for reply management, and a spreadsheet claiming to know which campaign produced the meeting. So we built one system around the actual motion. ### Start with a market, not a sequence Most outbound products begin with "upload a list." We wanted to begin one level earlier: who should even be on the list? We built the concept of a lead universe, the full set of people who could plausibly buy. Ours currently holds 26,477 potential buyers across the markets we are testing. They come from competitor customers, specific company categories, [people engaging with a founder's posts](/blog/turn-social-engagement-into-qualified-prospects), Sales Navigator searches and existing account data. Then we turn markets into plays: founders at seed to Series A, sales leaders at competitor customers, heads of growth in fintech, RevOps leaders at 200 to 2,000 person companies. The play defines who we are trying to reach and why. The sequence comes later. ### Every lead keeps its source This sounds boring until you try to answer: where did this meeting actually come from? We have 2,446 leads in the system right now. For every person we can see where they came from, which play discovered them, which sequence they entered, which LinkedIn account contacted them, which mailbox emailed them, their latest activity, whether they replied and whether a meeting was booked. Once you run several campaigns at the same time, this is the difference between attribution and guessing. ### LinkedIn and email work on the same prospect If Jordan connects with someone on LinkedIn and that person replies, the email campaign should know. If someone replies by email, the LinkedIn follow-ups should stop. If Priya is warming up a new LinkedIn account, the system should know which prospects can safely be assigned to her. Our founder sequence right now: 707 invites, 399 connections, 314 messages sent, 71 replies, 26 meetings booked. Sending is distributed across several LinkedIn accounts and mailboxes. [Multichannel only works when both channels operate on the same person](/blog/linkedin-and-email-outreach-same-lead-list). ### Measure copy by meetings We are testing three message variants inside the sequence, named playbook, compare-notes and direct-ask. Each is tracked against people reached, reply rate, meeting rate, the baseline and statistical confidence. One variant sits at a 23 percent reply rate and a 9.1 percent meeting rate. Another sits at 13 percent replies and 2.1 percent meetings. That is far more useful than debating copy in Slack. We also do not declare a winner after twelve sends. The experiment tells us when there is enough data to decide. ### The inbox is ranked by what needs a decision The workspace has separate views for "needs you," "engaged" and "meetings booked." Each reply is classified as meeting intent, question, objection or relevant later. A founder asking about pricing does not get buried under forty people saying thanks for connecting. You open the conversation, see the earlier messages, see why the system classified it that way, edit the suggested response and send. This is where AI earns its keep again: it has the conversation context instead of a blank prompt. ### Expose the whole thing through MCP This is the part I think matters most now. I barely look at my own dashboard anymore. We added an [MCP layer](/mcp) so the GTM workspace can be used directly from an AI assistant. Instead of exporting a CSV, uploading it to ChatGPT, explaining every column, asking a question and repeating it all tomorrow, you ask the system directly: - show me replies with meeting intent - which sequence generated the most meetings this week? - find leads waiting for review - compare the three message variants - which accounts visited the site after replying? The model is not replacing the GTM system. It is operating on top of it, which is a much better job for an AI. ## What changed for us Last seven days across our own outbound: 539 invites sent, 354 new connections, 276 messages, 98 replies, 38 meetings booked. Those numbers come from a system where prospecting, LinkedIn, email, replies, experiments and analytics all share the same records. AI helps us research faster, draft faster and process replies faster. It only becomes seriously useful once it knows what is happening inside the actual sales motion. If you are a founder still running outbound between product calls, [start at pumpgtm.com](https://pumpgtm.com/). You enter your website, PumpGTM proposes the first prospects and the reason for each one, and you approve before anything goes out. If you would rather see the setup first, [reach out](/contact) and I will send the walkthrough. ## Questions founders ask about AI outbound Can AI write my cold messages for me? It can draft them, and it is good at variants. It cannot tell you whether the reason for the conversation is real. Do the angle step first, then [review the draft against the anti-slop list](/blog/review-ai-written-cold-messages). How many message variants should I test at once? Three is enough. More than that spreads the sends too thin to reach a confident result at founder-scale volumes. Do I need email and LinkedIn from day one? No. Start on the channel where your buyers already reply. Add the second channel only when both can see the same prospect record, otherwise you will double-message people. What is a lead universe? The full set of people who could plausibly buy from you, before any sequence exists. Plays are slices of it, defined by who you want to reach and why. Sequences come after that. Why measure by meetings and not replies? Reply rate rewards curiosity. Meeting rate rewards fit. A variant with fewer replies but more meetings is the one you want. ## Put the guide to work Explore PumpGTM for finding relevant buyers and planning your outreach. [Explore PumpGTM](/?utm_source=pumpgtm_blog&utm_medium=owned_content&utm_campaign=gtm_guides&utm_content=ai-founder-led-outbound-system) ## More guides [Turn LinkedIn post engagement into qualified prospects](/blog/turn-social-engagement-into-qualified-prospects) [The Corgi outbound playbook: one lead list, LinkedIn plus email, and follow-ups that do not quit](/blog/corgi-gtm-playbook-linkedin-email-follow-ups)