The problem isn't AI. The problem is automating before you know what actually needs to be done.
This is the story we keep hearing. Founders see AI, get excited, and immediately try to automate everything. Then they wonder why they're bleeding cash and still don't have product-market fit.
Early Mistakes to Avoid
We've seen founders waste 10 to 15 hours a week setting up elaborate CRMs when they have zero customers. We've watched people let AI reply to customers autonomously in ways that made them cringe.
Here's what happens: you read about AI agents, you think "I can save so much time," and you start building. Legal agent. Finance agent. Marketing agent. You're solving problems you don't have yet.
Meanwhile your actual problem is that you need to talk to 50 more customers to figure out what they actually want. But you're too busy configuring agents.
The rule is simple. Do it manually first. Then automate.
What to Keep Manual and Founder-Led
Some things should never be automated early. Customer interviews. Prospect conversations before you have product. Customer success.
One founder still spends 75% of his time in customer accounts watching actual usage. Not delegating to an AI. Not handing it off to a junior hire. Actually watching how people use the product.
This is pattern recognition. You can't build it with an agent. You build it by being in the room.
Any interaction with someone you've spoken to before should never be handed to AI. If they've engaged once, they deserve you, not a bot. Outbound replies once someone responds? That's you. Known contacts reaching out? Also you.
You can automate the stuff around these conversations. But the conversations themselves? Those stay founder-led until you have real patterns.
Tools the Panelists Are Using
When founders do automate, here's what's actually working.
Claude as an orchestration hub. Not just for writing. For connecting systems, managing workflows, and acting as the control center.
Exa for ICP building. It pulls better context than most manual research.
PostHog for customer analytics. It's free until you scale, and it gives you the usage data you need to spot patterns.
LangFuse for agent tracing. When something breaks, you need to see what the agent actually did. This shows you.
The cost optimization approach is smart. Use expensive models like Claude Opus or GPT-4 when you're exploring and figuring out the workflow. Once you understand what needs to happen, switch to cheaper compute. You don't need the best model to run a process you've already mapped.
GTM Automation: What Scales Well
Once you know what you're doing, automation can be powerful.
One team cut onboarding time from 45 minutes to 3 minutes. They didn't automate the customer conversation. They automated the repeatable backend steps that happened after the conversation. Account setup. Data migration. Configuration.
Warm outbound workflows that pull context from HubSpot and Granola notes work well. The AI drafts an email based on actual context from past conversations. You review it. You send it. Still founder-led, but faster.
Customer upset detection that flags at-risk accounts in Slack is a win. The AI watches sentiment and usage patterns. When something looks off, it pings you. You handle the conversation.
One team built a self-healing bug triage agent that fixes basic bugs overnight. It doesn't touch complex issues. But the simple stuff? It handles it while you sleep.
Notice the pattern. These automations support the founder. They don't replace the founder.
Audience Q&A: Founder GTM Problems
Founders asked tactical questions. Here's what actually matters.
Learning to code with Claude: You don't need to be a developer. You need to be able to describe what you want clearly. Claude can write the code. You review it and test it. Treat it like pair programming with someone who knows syntax but doesn't know your business.
Transitioning from founder-led sales: Your first four reps build the playbook. They watch you. They shadow calls. They take notes. Once they've seen you do it 20 times, they try it with you listening. The playbook isn't a doc you write. It's what emerges from repetition.
Treating agents like untrained interns: This framing helps. You wouldn't let an intern send customer emails unsupervised on day one. Same with agents. They need guardrails, review, and iteration.
Timeboxing customer conversations: Set a timer for 25 minutes. Tell the customer upfront. Most people respect it. You get better signal when conversations are focused.
Connecting email and LinkedIn to Claude: Warm leads are already in your inbox. Connect your email and LinkedIn to Claude. Ask it to find people who've engaged with you before, commented on posts, or replied to messages. Then reach out personally. This isn't cold outbound. This is following up with people who already know you exist.
Recommended Resources and Closing Advice
Closing actions:
Audit your day for time sinks. Where are you spending hours on work that doesn't move the needle? That's where automation might help. But only after you've done it manually enough to know what good looks like.
Publish about your journey. Founders research you before every call. If they see you're thinking deeply about the problem they have, they'll want to talk. Your blog, your Twitter, your LinkedIn. It all matters.
And remember: automate after you've done it manually first. The AI needs to learn your actual voice and process. It can't learn from nothing.
Over-automating early wastes capital and time before you know what actually needs to be done. Stay in the conversations until you have real patterns. Then automate the repeatable stuff around them.
That's how you use AI without lighting money on fire.
Resources discussed at the Event
Books
- The Experimentation Machine by Jeff Bussgang
- The Power of Pull by Rob Snyder
- Outliers by Malcolm Gladwell
- How Big Things Get Done by Bent Flyvbjerg and Dan Gardner
- Atomic Habits by James Clear
- Go for No by Richard Fenton and Andrea Waltz
Newsletters
- Growth Unhinged by Kyle Poyar
- 100 Founders by Dave Rubinstein
Tools the panelists are using