I recently jumped on the Agency Success podcast to talk about how agencies can actually use AI in their day-to-day work. Not the “50 automated Instagram posts a day” version of AI. The practical, “I have client work to do and not enough hours” version.
This page is a companion to that episode. Everything we talked about, the tools I mentioned, the framework I use with folks, and some links so you’re not frantically Googling while you listen.
Two ways this plays out for agencies
When I talk to agency owners about AI, I see it creating opportunity in two directions.
Internally, it helps you do more with what you’ve got. Less context switching, faster debugging, content you’ve been meaning to publish for six months finally getting out the door. I think of it as adding capacity without adding headcount.
Externally, it’s a new advisory layer with your clients. Your clients are starting to ask about this stuff. The same way you became their trusted partner on WordPress or SEO or their email strategy, AI is the next thing they’re going to want guidance on. And you don’t need to be an expert to start. You just need to be a few steps ahead.
Crawl, walk, run, fly
This is the framework I walk folks through when they’re figuring out where to start. It’s not proprietary or anything (it’s been around forever), but I think it maps really well to how AI adoption actually works in practice.
Crawl: embed AI invisibly. Use tools that leverage AI under the hood without requiring you to prompt anything. You’re getting value from the technology without changing how you work. A good example here is using something like Granola or Fathom to transcribe all your calls. You install it, forget about it, and suddenly you have this searchable archive of every client conversation.
Walk: learn to prompt. This is where you start using the off-the-shelf tools directly. ChatGPT, Claude, Gemini. You’re learning how to write good prompts, figuring out when to use which model, and starting to develop a feel for how these tools actually think.
Run: personalize the tools. Same off-the-shelf tools, but now you’re feeding them context about your specific situation. Custom GPTs, Claude Projects, uploaded docs. So when you open Claude to work on something for a client, it’s not a blank slate. It already knows who the client is, what you’ve been working on, the backstory.
Fly: build custom tools. This is the fully custom stuff. Agents, automated workflows, tools built specifically for your needs. I recently built myself what I call a CMO agent. It has access to my email subscribers in Kit, my Google Search Console data, and my content plan. Every day it pings me and says something like “hey, looks like we got a couple more subscribers but we’re still away from our goal. Given our content plan, we should probably publish this today.” It keeps me honest about actually promoting the stuff I write, which is historically where I fall off.
“Getting to the point where I’m building a custom agent to be my CMO has been a two or three year progression. I don’t recommend skipping steps. Even though all these new tools make it easier, it’s still helpful to have that base level understanding.”
WordPress-specific use cases
These are the things I’m showing people all the time. No custom builds required. You can start doing these today.
Debugging: screenshot your plugins, paste your error logs. I do this constantly. Take a screenshot of your entire plugins screen, grab the WordPress version and theme info, paste in whatever error logs you’re seeing. AI can cross-reference all of it and narrow down conflicts in minutes. I walked through a real example on the show where a colleague and I were debugging a jQuery conflict between a plugin and theme. The AI spent about five minutes looking through the plugin code, the theme code, all her customizations, and nailed it. Thirty minutes total, start to finish. That would have been a three-hour Zoom call, minimum, and the site was actively launching that night. So it was pretty critical.
Performance: translate PageSpeed reports into fixes. Throw your Google PageSpeed Insights report into Claude or ChatGPT and ask it to walk you through each recommendation one by one. Whether that’s installing a plugin, writing a little code snippet, or just changing a setting. The nice thing is it meets you at whatever technical level you’re at. A lot of those PageSpeed recommendations are really technical, and it’s hard to go from the recommendation to an actual fix on your site. This bridges that gap.
Coding: one ticket per sprint. This is what I recommend to agencies that are writing a lot of code. Just pick one well-defined ticket each sprint and try to solve it entirely with an AI coding tool. Don’t open your code editor. The first time, you’ll say “I could have done that in 30 minutes and it took me two hours to get this assistant to do it.” But then your brain starts to zoom out. You start thinking less about “how do I write this function” and more about “what’s the spec, and what context does the AI need from me?”
Client advisory: teach them the basics. Most clients are even further behind the adoption curve than you are. Even walking someone through how Claude Projects work, or showing them that this is a different tool than Google where you don’t just type a search query, that’s valuable. Once you’ve gone through the crawl and walk stages yourself, you can turn that experience into something you offer your clients.
Tools mentioned in this episode
The current stack as of February 2026. This will change. These are what’s working well right now.
- Claude — My primary AI for writing, reasoning, and general work. The Projects feature is great for persistent context. claude.ai
- Claude Code — Command-line coding tool from Anthropic. My tool of choice for building projects and debugging WordPress. anthropic.com
- ChatGPT — Custom GPTs for personalized workflows. I bounce here when I’m hitting a dead end with Claude. chatgpt.com
- Gemini — Google’s AI, with strong coding capabilities and integrated image/video generation. gemini.google.com
- Nano Banana — Google’s image generation and editing model inside Gemini. Great for hero images, mood boards, quick prototyping. I’ve also used it as a surprisingly powerful image editing tool. gemini.google/overview/image-generation
- Google Veo — Video generation. I use this to animate storyboard concepts I’ve created in Nano Banana. deepmind.google/models/veo
- Suno — AI music generation from text prompts. Full songs with vocals and instrumentation. My first year of college I was a music composition major, so it’s been interesting to watch these models develop. suno.com
- Granola — Transcribes calls without a bot joining your meeting. A great “crawl” stage tool. granola.ai
- Kit — Email platform (formerly ConvertKit). I use this in my CMO agent setup for subscriber tracking. kit.com
- OpenClaw — Open-source personal AI assistant that lives in your messaging apps (Telegram, WhatsApp, etc.). This is still bleeding edge for me. It consumes a ton of tokens, and I’m still figuring out the balance of giving it useful memory without overwhelming it. But I spent about an hour walking it through my gym’s website and now it can reschedule sessions for me, which is pretty cool. openclaw.ai
Key takeaways
Start where you are. If you’re using AI at all, you can start sharing what works with clients. A short video, a blog post, a quick Loom walkthrough of how you use it day to day. That’s enough to position yourself as someone who knows what they’re talking about.
There’s a huge gap between demos and real work. It’s very tempting to see someone on social media who’s built this automated pipeline and think “I need to do that.” But there’s a massive gap between seeing a cool demo and knowing how to apply it to the stuff you actually have to do for your clients every day. Bridging that gap is where the real value lives.
AI instinct is a real skill. Unlike traditional software where you go to the pricing page and it lists all the features, these models are pretty squishy. They’re good at some things, bad at other things. And sometimes the things they’re good at seem really hard and the things they can’t do seem really easy. The only way to develop judgment about what works is direct experience.
Code is where AI shines brightest right now. Code is in a unique position of being both a little bit creative (there are lots of ways to solve a problem) and easily verifiable (does it run? does it pass the test?). So these models are getting really, really good at writing code. The loop from idea to working product is getting remarkably short.
Context switching is the hidden killer. If a client emails you about a site you haven’t touched in a week, there’s this huge ramp-up cost to get back into it. AI tools that already have your project context can dramatically cut that down. It’s like having a dedicated project manager on every engagement.
If your AI usage isn’t failing occasionally, you’re not pushing hard enough. I genuinely believe this. If I’m not hitting a wall once or twice a week, it means I’m not trying to do interesting enough things. The failures are where you learn what these tools can and can’t do.
Want to talk through your situation?
Whether you’re just getting started with AI or you’re looking for someone to help you stay on the cutting edge, I’d love to hear what you’re working on.
No pitch. Just a conversation about where you’re at and where AI fits in.