What Can Agents Do For SMB Owners?
The owner of a marketing agency who took the first round of my AI class was on the 21 Hats podcast last week talking about some of the results from an agent workflow we built.
The workflow scrapes cold leads from a lead source, scores candidates against her Ideal Customer Profile, enriches the lead with research, and delivers thirty well-targeted and documented leads to her each week.
There are no ads. No sales team to hire. This all comes from an AI workflow we hacked together in ChatGPT with some connectors.
The results are amazing: Within a week or two of setting this up she’s:
Brought in two new paying clients with more to come.
Saving 30 hours a week. That's 30 hours of work that still gets done but just not by her or her employees and it takes 2 minutes.
The work didn't disappear. It moved. And the best part is, no one is getting laid off.
In fact she’s hiring staff to serve the new clients.
The Six Levels of AI Usage
I’ve been writing about small business and AI since last Spring and now after running my first training class, I’m getting a better picture of where people are and the levels of AI maturity. Here’s how I’m thinking about it now:
Level 1: Using Chatbots. Most of us already do this but people need to diversify their basic ChatGPT. It’s better to use Claude for writing and coding. Gemini for planning. ChatGPT for for search or Deep Research. The key here is learning how to prompt the AIs more effectively (almost no one does this well) and to start building basic agents like a custom GPT in ChatGPT or a Claude Project.
Level 2: Building AI workflows. This is all about building agents and then chaining them together. This lets you automate repetitive work. Everyone in my last class built custom GPTs in ChatGPT that pass information to each other, with an orchestrator managing the whole thing. You also learn how to begin connecting other applications to your GPTs and ChatGPT using the built in Connectors, MCP features, or Zapier.
Level 3: Building Basic Web Apps and Sites. The next level is understanding introductory programming tools like Lovable, Replit, Bolt, etc. These tools allow you to build websites and basic web apps using plan English prompts. Now you are vibe coding websites and apps you can publish to the web. For example, we use a spreadsheet with messy data about food trucks in New York, their offerings, and their locations. We cleaned up the data and built a web app that showed an interactive map of the trucks and their menus.
Level 4: Using Coding Agents: This is where real magic happens. You start using Claude Code or Codex from Openai with a Github Codespace or in your computer terminal. Now you can build almost anything. This is true vibe coding and with the release of Claude’s Opus 4.5 model last week, I think we are close to amateurs building fully scalable web applications.
Level 5: Building More Sophisticated Agents and Tools to Improve Agents. Now we can start making the agents smarter and we’ve freed ourselves from custom GPTs with Claude Code and Codex. We learn about basic agent models like LangChain, Google Vertex, etc. We use antagonist and user advocacy agents that challenge the main agent to get better performance. We also look at building agents that can ingest metrics to improve their output. For example, I have an agent that reviews all my Linkedin post metrics and makes recommendations for improving the Linkedin post agent.
Level 6: Building Tools that Build Agents and Apps. This is the cutting edge that really only took shape over the last month. You can now adopt other people’s best practices to improve your AI coding. A good example is Claude Code’s Skills and Plugins features. Don’t know how to do testing for your app? Go find a programmer with 30 years of experience who built a testing skill and download it from Github. Now you can do what he does.
You're still doing work that an AI workflow could handle in 2 minutes. Not someday. Right now.
My AI cohort is built for owners—not developers, not your IT guy. You'll leave with working systems, not slide decks.
Next cohort starts in last January. Limited to 12 seats. For more information, sign up here.
My Challenge To My Students
Back to my marketing agency owner. Now that she is bringing in clients and saving 30 hours a week, we need to start improving the agents. Which leads converted and which didn’t? Can we learn something after we’ve gone through two months of leads that can help the agent workflow?
That's the difference between a system that works and a system that compounds. Right now her workflow runs. But it doesn't learn. It doesn't get better on its own. It doesn't flag what's working and adjust.
That's Level 5. Level 6 would be figuring out how to optimize the way the agent works. Can we scale it to 50 leads? 100? Can we build agents that do most of the outreach work she and her team are still doing manually. And that's what my advanced class is designed to help you reach.
Why This Matters Now
Most existing companies are at Level 1 and maybe 1% are at Level 2 but younger entrepreneurs are going to start at Levels 4, 5, and 6. They don't have legacy processes to unlearn. They don't have employees who've "always done it this way."
If you're running a business built before ChatGPT launched, you're competing against founders who build agents from Day One to scale their productivity.
The question isn't whether to use AI. It's whether you'll keep doing work that should be delegated to systems—or start designing the systems instead.