I hosted an AI roundtable last week with a handful of founders—SaaS, voice agents, hospitality tech, marketing platforms. These are people I've invested in or worked alongside, and they're all shipping product and closing deals right now. The question I put to the group was simple: Is AI anywhere in the part of your business that actually makes money?
The room got honest fast. What follows are the insights that mattered, shared under Chatham House rules.
Executive AI Roundtable
Conversations like the one behind this essay happen every week. I host a closed-door roundtable for founders and C-level leaders navigating AI strategy — no vendors, no pitches, just operators comparing notes.
Request an Invitation →Most Founders Are Playing With AI Instead of Deploying It
A lot of founders are building internal dashboards, automating Slack workflows, and calling it AI adoption. But nothing they've built touches revenue. I see this in my own portfolio—founders with real credibility in the AI space whose monthly revenue doesn't reflect any of that expertise.
The internal tinkering feels productive. It's not. It's avoidance dressed up as innovation.
The first question worth answering: is AI contributing to how your business makes money? Not to how your team communicates. Not to how you organize tasks. To how you generate and close revenue.
What's Working: AI That Books Meetings and Fills Pipelines
An AI SDR That Actually Closes
The sharpest revenue-facing example from the roundtable: one founder's team sends about 2,000 cold emails a day through Instantly. When a prospect responds, an AI agent takes over the entire conversation. It knows the prospect's location, business type, pricing, and common objections. It handles the back-and-forth, overcomes pushback, and proposes three meeting times on the next available day with a booking link as backup.
Meetings get booked. No human SDR involved. For the first three or so exchanges, the agent performs well—because prospects responding to cold outbound can only say about three or four different things. You map those responses, give the AI guidance on each one, and it handles the rest. No decision tree required. A well-loaded prompt with bounded expectations is enough.
Where it breaks down: after several exchanges, the agent loses context. It'll ask someone to confirm a time they already confirmed. At that point, a human steps in. But for high-volume appointment setting, the ROI is clear even with that limitation: the expensive humans focus on the sales side.
AI That Replaces Your VA and Your Fiverr Scrapers
Another founder replaced a multi-day outsourced research process with AI that finds target organizations online, pulls lists, cross-references them against their HubSpot CRM, and approximates ideal customer fit.
What used to take a VA days now takes minutes—and feeds directly into the pipeline.
To build this, they tested ChatGPT against Manus and found ChatGPT produced stronger results for their workflow. This underlines a common thread in AI: model performance can depend on the task. Don't assume your favorite model for task A will have the best performance for task B. For critical work, try multiple models.
Narrow Agents Beat Grand Systems
The founders at the table who are getting results build narrow agents that do one job well. Specific task, clear boundaries, defined handoff to a human.
A churn agent that opens a conversation when a customer clicks "cancel." Its only objective: find out why. It can make one save attempt—a discount for seasonal businesses, for instance—then hands full context to the customer success team. It never cancels anything. The save is a bonus; the real value is intelligence. A conversation reveals why customers leave in a way that a five-option dropdown never will. That data feeds back into product, pricing, and retention decisions.
An onboarding agent that replaces static signup forms with a real-time conversation. It captures what the customer wants to accomplish, what they're using today, and what problems they need solved. When a new signup goes cold a week later, you have enough context for a follow-up that sounds like you were paying attention—because the agent was.
One founder at the roundtable took this further. Their company builds voice agents, so they use their own product to onboard new customers. A prospect fills out a basic lead form, then drops into a full voice-or-chat conversation where the agent asks about their business, configures the product, provisions their account, and gives them access to test. Lead to active trial in one sitting, no human needed.
The common thread: pick a specific job, give the agent boundaries and context, define when a human takes over, and ship it. Then keep iterating.
Claude Code Gets Better Every Time You Use It (If You Set It Up Right)
Claude Code came up a lot in the roundtable, and the gap between founders who get enormous value from it and those who find it merely useful can come down how much they invest in structuring context. For valuable workflows you plan to repeat, don't become complacent. You can invest in tooling to improve that workflow and get more predictable results with less re-work.
The CLAUDE.md file is the starting point—a markdown file in your project that Claude Code reads at the beginning of every session. It's a persistent system prompt where you define business context, available skills, tone, architectural patterns, and things to avoid. Think of it as your standard operating procedure or work instructions for the project. The more you put into it, the less you repeat yourself and the sharper your output gets as you improve it.
Anthropic's CLAUDE.md Management plugin takes this further. After a session, you invoke a command that reviews what happened—patterns followed, commands discovered, quirks encountered—and proposes updates to your CLAUDE.md. Every session teaches the next one.
A few patterns that came up from the group:
Nested context for larger projects. A root CLAUDE.md can point to subdirectories; each has its own CLAUDE.md with area-specific instructions. One founder's team separates templates, services, and deployment configs this way.
A domain glossary. One founder built a docs folder with business logic, testing patterns, and terminology. These freed him from re-explaining his domain every session.
Chained post-session automation. You can add instructions to your Claude.md requesting it to automatically perform housecleaning at the end of each task. Push to GitHub, wait for CI, pull results, run the CLAUDE.md revision command. It's a feedback loop that compounds.
AI Can Make SEO and Marketing Much Cheaper, Even if it's Not Your Top Priority
I connected Claude Code to Google Search Console and built an SEO auditor. It pulled my search data, found page titles getting truncated in results, identified content gaps based on competitor research, and wrote a full SEO brief. It caught things I would have missed—and the whole thing took an evening to set up.
Tell Claude Code to audit your site, research your competitors, and identify what you could rank for. The output is specific and ready to act on.
Even if outbound is your bread and butter, think about what happens after you send that cold email. A lot of prospects will Google your category before they reply. If they find your competitors and not you, you've lost credibility before the first conversation.
On the marketing front: a founder recommended installing Corey Haines' marketing skills for Claude Code. They used the CRO skill on a signup page, simplified the layout based on its suggestions, and saw conversions improve. Free, practical, and worth the five minutes to set up.
A Working AI Executive Assistant
I built my own Claude Code executive assistant. It pulls my calendars through iCal feeds, monitors outbound email responses, manages my to-do list, and references a directory of markdown files—branding, personal info, quarterly goals, strategic priorities.
For delivery, I use Pushover, a webhook-to-push-notification app. My agent pings me when the morning or evening briefing is ready.
A founder suggested the next evolution should include intelligent triage. I can forward emails or ideas to myself with specific subject tags throughout the day, and the assistant will ingest, categorize, and prioritize them. They described a similar setup—tagging emails, having Claude Code skills sort and surface what matters most. It works because it turns scattered inputs into structured priorities.
Three Hiring Outcomes in the AI Era
The roundtable surfaced patterns I've been seeing across my portfolio:
The founder who stopped hiring. A contractor left right when AI agents became capable enough to fill the gap. This founder now ships more than they did with a full-time developer, at a fraction of the cost. The real advantage isn't price—it's speed. No timezone gaps, no context loss, no translation overhead. The iteration loop is nearly instantaneous.
The hire who couldn't adapt. A founder brought on someone with deep industry knowledge and design skills, expecting speed. What they got was pixel-perfect Figma mockups when the priority was rough prototypes and fast feedback. The tooling wasn't ready for a non-technical person to ship code, and the hire couldn't bridge the gap. It's becoming more important to hire people who feel comfortable leveraging AI to get their work done.
The hire who expanded what was possible. Another founder brought on an engineer with a background in deploying complex systems at enterprise scale. From day one, this person proposed approaches nobody on the team had considered—novel indexing methods, new agentic workflow patterns. They don't just use AI. They change the ceiling for the whole team.
Here's a point I don't see enough people making: the two most valuable skills in the AI era are reading comprehension and writing. LLMs run on reading and writing. The people who communicate clearly, evaluate output critically, and iterate on prompts with precision will outperform technical specialists who can't. I've seen strong contributors come from liberal arts, journalism, and creative writing backgrounds. Stop trying to outbid OpenAI and Anthropic for the same engineering talent. Look where nobody else is looking.
The AI Adoption Problem on Your Team
Two dynamics came up at the table:
People who don't see the value. Non-technical employees aren't resistant to AI—they just don't know where it fits. The leverage isn't obvious when you haven't experienced it firsthand.
People who can't stop building. One founder's CSM was so enthusiastic about AI that he was shipping new dashboards and tools daily—more than the founder had bandwidth to review. They were concerned that the CSM might not be focused enough on results.
The fix for both: define success by what gets achieved, not what gets built. Have your people report on outcomes against the goals you've set. If someone is cranking out AI-powered dashboards that don't connect to a business result, redirect them. If someone isn't using AI at all but is delivering, don't force it—yet.
Your job as a founder is to make the target clear and give people enough room to hit it. How they get there matters less than whether they do. You want to build a culture where employees have the latitude to work creatively on one hand, and can report their results without micromanagement on the other.
Keep AI in Its Place
I'll be honest: I catch myself at night wanting to queue up tasks for the AI to work on while I sleep. The pull to keep building is strong. Every founder in the room admitted the same thing.
But the most valuable thing any of us can do has nothing to do with AI. It's deciding where the business needs to go. What's the most important thing you can contribute this quarter? Usually it's driving attention to the business or opening a new revenue channel. Not building another internal tool.
Every founder gravitates toward their comfort zone. Engineers build. Salespeople sell. AI amplifies whatever your default impulse is—which means it makes it easier than ever to stay busy without being strategic.
Force yourself to step back and ask: what does this company need to achieve in the next twelve months that would actually change our trajectory? Sometimes the answer is tactical—more of what's already working. Sometimes it's bigger—an enterprise deal, a partnership, a repositioning.
Agents and automations are valuable. But only in service of a strategy. Never as a substitute for one.
Resources From the Roundtable
- Instantly — Cold outbound email with built-in AI agents for appointment setting
- Claude Code — Agentic coding; invest in CLAUDE.md to make it compound
- CLAUDE.md Management Plugin — Keep your project context current after every session
- Corey Haines' Marketing Skills — Free Claude Code skills for CRO, SEO, copywriting, and growth
- Import AI Newsletter — Weekly AI read by Anthropic co-founder Jack Clark; a journalist's take on the AI landscape
- Pushover — Webhook-to-push-notification service for AI workflow alerts
- Google Search Console — Connect to Claude Code for automated SEO auditing
I work with a small number of founders on AI strategy—where it creates real leverage, and where it's a distraction. If you're ready for AI to help you grow your business, let's talk.