The LLM Is the Dumbest Part of Your AI Company

I was on a product call with my dev team the other day - we're building out Galadon, our AI sales chatbot platform - and one of my developers said something that stopped me cold.

We were talking about the roadmap. Where does the product go after launch? What do we build next? And I said it out loud, almost without thinking: the LLM is the most basic part of what we're building.

He agreed immediately. And then we both just sat there for a second, because we realized what that actually meant.

Most founders building AI products right now are doing the exact opposite of what we just described. They're writing fundraising decks around the model. They're naming their companies after AI. They're designing their brand around the chatbot interface. They're spending enormous energy optimizing the one layer of their product that is - at this point, almost by definition - a commodity.

This post is about that trap. How founders fall into it, why it's so easy to fall into, and what the actual work looks like if you want to build an AI company that doesn't get replaced by a GPT-5 update.

The LLM Isn't the Product. The Plumbing Is.

Here's what we're actually building with Galadon: integrations, CRM tools, sales workflows, lead capture logic, onboarding sequences, training systems. A chatbot that doesn't just answer questions - it closes deals, qualifies leads, and responds to DMs and cold email replies automatically. The kind of thing that runs while you're asleep.

How much of that is the LLM? Almost none of it.

The LLM is the piece that takes an input and generates an output. That's table stakes now. What makes the product actually work - what makes it so that a user doesn't cancel after 30 days - is everything else. The integrations. The onboarding flow. The pre-trained workflows. The support infrastructure. The system that makes it easy for a non-technical founder to spin up a chatbot in ten minutes without needing to write a single line of code.

That's the real product. And it has almost nothing to do with AI.

When we look at where Galadon is going, we're modeling it more like Crisp - a full customer communication platform - than like a ChatGPT wrapper. Live chat, integrations, CRM-level tools, sales tracking. Most of what we'll end up building won't even be AI. It'll be the boring infrastructure that makes the AI actually useful.

But here's the thing: most founders building AI products right now aren't thinking this way.

Why Everyone Is Building the Wrong Thing

The AI wave hit fast. And when something moves that fast, everyone rushes to stake their claim on the most visible part of it.

The visible part is the chatbot. The model. The "AI-powered" headline. So that's what people build. They spin up a GPT wrapper, give it a name, build a landing page, and start pitching investors on their AI company.

I get it. I applied to YC with Galadon - partly as a legitimate shot, partly as a content moment. When you're in it, the AI angle feels like the story. It's what's exciting. It's what gets the tweet likes and the investor meetings.

But here's the problem: you're not the only one who figured out how to call the OpenAI API.

My dev made this point in a way I keep thinking about. He said that LLM app development is a trend that's really only been going for a short period of time. People are still figuring out the use cases, still figuring out how to scope these projects. But the window where "I built a chatbot" is a differentiator is closing fast. The model layer - the part everyone's branding around - is the part that's becoming a commodity first.

There are no-code tools that let someone spin up a basic chatbot in an afternoon. The reason our build took months wasn't the AI. It was the architecture underneath it - the system that lets users build their own custom chatbot without touching code, configure it through a guided multi-step workflow, integrate it with their existing tools, and deploy it without needing a developer on call.

That stuff is hard. And it's hard specifically because nobody wants to build it. It's not sexy. You can't put it in a pitch deck headline. But it's the reason users don't churn.

The Chatbot Creation Problem (And What It Taught Me)

Let me give you a concrete example from our own build.

We were working through the onboarding flow for Galadon - the sequence a new user goes through to create their first chatbot. Goal, sources, appearance, notifications, integration, installation. Six steps.

The question came up: what happens if a user tries to skip ahead? What if they're on step two and they click over to step five?

My instinct was to just let them click. Keep it open. But my dev pushed back - and he was right. There are workflows that depend on specific steps being completed first. If you skip sources, the chatbot has nothing to pull from. If you skip the goal setup, the whole conversation logic breaks. So the steps need to be enforced.

We landed on a simple solution: each step only becomes clickable when the previous one is complete. And if someone tries to jump ahead, a small inline alert fires - "complete previous steps before continuing" - nothing aggressive, just a nudge.

That's a tiny UX decision. Thirty seconds of conversation. But it determines whether a new user actually completes setup and gets value from the product on day one, or bounces in the middle of onboarding and never comes back.

That decision has zero to do with the LLM. It has everything to do with whether the product retains users.

This is what real product development looks like for AI companies. It's not model selection. It's not prompt engineering. It's the boring stuff - the flow states, the validation logic, the guardrails, the copy in the tooltip. That's where retention gets made or lost.

What "AI Company" Actually Means in Practice

Here's something worth thinking through: is Galadon an AI company?

In one sense, yes. The chatbot uses a language model. There's an AI layer in there. But if you removed the AI entirely and replaced it with a rules-based decision tree, would the core product still work? Yeah. Probably. For a lot of use cases.

What users are actually paying for is the outcome - more leads, qualified conversations, automated follow-up, faster onboarding. The AI is how we deliver that outcome. It's not the thing they're buying.

This distinction matters because it changes how you build. If you think you're selling AI, you optimize the AI. If you think you're selling outcomes, you optimize for the thing that produces outcomes - which, in almost every case, is the surrounding infrastructure.

When we talked about where Galadon goes from here, we talked about sending a million cold emails and having the chatbot respond to every reply automatically. That idea - autonomous outbound at scale with automated response handling - is legitimately powerful. But the AI is a tiny part of it. The hard part is the email infrastructure, the response classification, the handoff logic, the CRM sync. If you want to learn more about the outbound side of that equation, I laid out the framework in the Top 5 Cold Email Scripts - start there.

The point is: the AI makes it smarter. The infrastructure makes it work.

The "Text to Chatbot" Trap

There's a version of the AI chatbot idea that's been floating around for a while: what if you just described what you want, and the AI built the chatbot for you? You say "I want a chatbot that generates leads for my coaching business, my website is X" - and the whole thing gets configured automatically.

That's a compelling pitch. Webflow is doing something similar with AI-generated website layouts. Other tools are doing it with forms, with content, with emails.

But here's what most people don't realize: the hard part of that isn't the language understanding. It's the back-end that takes what the AI figures out and maps it to actual product configuration. The prompts, the API calls, the data structures, the validation. All of that has to exist first before the "just describe what you want" layer is possible.

In other words, you have to build the boring thing before you can build the magic thing. And most founders are pitching the magic thing before they've figured out the boring thing.

We're almost there with Galadon - where if you just fed it a one-line description of your business and your goal, it could configure a chatbot for you automatically. But that's only possible because we spent months building the underlying architecture that makes the configuration logic work. The AI-powered creation experience is the last feature, not the first.

What This Means for Your Build Strategy

If you're building an AI product right now - whether it's a chatbot, an agent, a content tool, a lead gen system - here's the question you need to answer honestly:

If OpenAI made GPT-6 available for free tomorrow and it was 10x better than what you're using, would your product still be defensible?

If your answer is no, you've built a wrapper. Not a product.

The companies that survive the next few years of AI commoditization aren't going to be the ones with the best model. They're going to be the ones with the deepest integrations, the most friction-filled switching costs, the most proprietary workflow logic baked into their product. The stuff that's genuinely hard to replicate - not because the AI is special, but because the surrounding system took years to build.

That's what creates retention. That's what creates margin. That's what creates a business that doesn't get one-shotted by a ChatGPT update.

Think about Crisp. Think about Intercom. These are not "AI companies." They're communication platforms. They've added AI features. But their moat is the integrations, the CRM layer, the support tools, the workflow automations - the decade of boring infrastructure they built before AI was even a conversation. That's why a new chatbot startup can't just come along and replace them. The model isn't the moat. The plumbing is.

The Marketing Problem

Here's the other side of this that most people ignore: your marketing is also probably over-indexed on the AI.

If your whole pitch is "AI-powered" - if that's the headline on your landing page, the hook in your cold emails, the angle in your content - you're competing on the most crowded possible positioning right now. Every product is "AI-powered." It means nothing.

What actually converts is specificity. Not "AI-powered chatbot" - but "a chatbot that closes more deals for coaching businesses, set up in under ten minutes, no code required." That's a specific offer that a specific person can say yes or no to. The AI is implied. It doesn't need to be the pitch.

This is the same principle I've been teaching for years on cold email. The offer has to be specific enough that someone in your target market immediately sees themselves in it. "We grow Instagram accounts for realtors" beats "social media marketing" every time. The same logic applies here: "we set up a lead capture chatbot for your coaching business" beats "AI-powered conversational marketing platform" every time.

If you want to sharpen your offer and your outreach at the same time, grab the 7-Figure Agency Blueprint - it's built around this exact framework.

The Real Opportunity in AI Right Now

Here's the honest take: we're early. The LLM application wave has only been going for a short period of time relative to where it's going. My dev said it clearly - by the end of next year, there's going to be a peak in LLM app development. Huge demand. Lots of money flowing. Lots of companies trying to build enterprise tools without really knowing how to scope them.

That's an opportunity - but not for the reason most people think.

The opportunity isn't to be the company that built the best AI. It's to be the company that built the best infrastructure around the AI. The deepest integrations. The most specific use case. The product that a buyer in a specific vertical can't imagine running their business without, because it's woven into every workflow they have.

The AI layer is going to get cheaper. The model is going to get smarter. What's going to separate winners from losers isn't the intelligence of the model - it's the stickiness of everything surrounding it.

Build the boring stuff. Build it well. The magic follows.

Where We're Headed With Galadon

To be concrete about our own roadmap: the chatbot is getting there. The model functionality is nearly complete. What we're focused on now is the stuff that makes users stay - the usage dashboards, the weekly and monthly analytics, the video training for each workflow step, the support chatbot trained on the Galadon knowledge base so users can get answers without waiting for a human.

None of that is AI in the exciting sense. All of it is infrastructure in the boring sense. And all of it is what will determine whether a user who signs up in month one is still a customer in month six.

That's the game. And it's a very different game than most founders think they're playing when they start building an "AI company."

If you're working through lead generation and outbound to support a product like this - whether you're building or selling - the Best Lead Strategy Guide is worth reading. And if you want to use tools to build your prospect list while you're building your product, ScraperCity's B2B database and the email finder are what I use - alongside Apollo and other tools in the stack - to build targeted lists at scale.

The LLM is the dumbest part of your AI company. Build accordingly.