Overcoming AI Myths by "Looping" LLMs

Like most people back in 2022, I used ChatGPT as a glorified search engine. I loved the lack of ads and simple chat experience, and later the instant visualization of ideas my brain conjured up, like the image above. At the time my team marketed AI products to customers, but we never used LLMs to actually do our jobs.

Then a smart friend sat me down and showed me something that changed everything: how to deploy an LLM with purpose. Not just tossing out half-baked prompts and hoping for magic, but actually steering it toward specific use cases and outcomes.

In less than a day, it was working far better with me because there was mutual understanding. Today, I use AI every day. Alongside my "legacy apps," it's helping me solve problems I've faced my entire career.

After countless conversations with leaders and doers, with very few feeling great about their rollout, I decided to write this. If you've been putting off using LLMs at work because you are too busy, or want to address leadership's frustration with AI failing to deliver results, I hope this helps.

This post had specific use cases for product/marketing/sales teams for each myth, but it ballooned to well over 2000 words. I would love your feedback on my next piece “The AI Guide to Supercharge Product Launch Teams” I’m writing instead.

DM me or comment on this post with a use case and I’ll share an early version with you.

Start Thinking in Loops

Before we dive into the myths, let me explain what I mean by "loops." Most people treat AI like a vending machine: insert prompt, get answer, happy, done. Not happy? Blame LLM. If AI were a colleague, it wouldn’t think you were very collaborative.

The reality? LLMs perform best working with people iteratively, not for people in isolation.

Each "loop" is a continuous feedback cycle where humans and AI collaborate to get smarter together. You train it, it helps you, you refine it, it gets better. Behavior like that is hard to change individually, but at an organizational level implementing these loops is quite a big shift.

The same use cases I've worked on for decades still exist and can be solved far better with AI. Build what customers need, personalize yet stay on message, sell our products and our vision. But LLMs require a completely different behavioral shift from traditional databases and systems. Let’s dig in.

🤖 Myth #1: "LLMs are generic and feel robotic"

Reality Check: Help AI learn your style, and it becomes surprisingly personal.

LLMs like Claude, ChatGPT, and Gemini aren't built to be specific unless you give them instructions, knowledge, and workflows. Generic AI reflects generic inputs. This explains why MIT researchers noted we're not seeing the business transformations everyone expected.

The Style Loop

Claude's Style feature lets you upload your writing, and it'll start writing like you. I fed it this post, and it summarized my style as "craft tech narratives that blend strategic insights with personal, conversational storytelling." Not bad, right?

Pro Style: Have Claude create a comprehensive style prompt from all your best writing. Put that in custom instructions and see if it can write your autobiography.

The Brand Loop

I can't remember meeting a product-marketing-sales leader who didn't struggle with getting teams on brand and message.

Your first big AI win? Help your LLM absorb your brand voice, tone, and messaging. We built a Brand/Messaging/Content agent for our clients at Launch Engine. Getting these three artifacts right becomes your LLM foundation.

Take your time here. This can be a foundation for everything you build later.

Brand Loop: Guide creation from single prompt to be edited, then added as agent.

😫 Myth #2: "AI doesn't understand our business"

Reality Check: Most companies don't give it a chance to.

If you just bought licenses and handed them out, your LLM has read your website (hopefully not the version from six months ago) but has zero clue what you actually do. If employees connect it to their desktop or Google Drive, it's looking at old products and outdated messaging.

The Learning Loop

Train your LLM like you'd train any new employee. Market nuances. Business model. Products. Roadmap. Customers. Competitors. Departments. Give it access to your onboarding materials, and have it tell you what you're missing.

The Expert Loop

Train your employees and use the relationship between your experts to push LLM performance. People learn from experts. So does AI. Create continuous feedback loops that improve market insights, product learnings, and customer knowledge over time.

Make sure it understands how teams work together so it integrates into actual workflows. We've created a taxonomy that provides backbone for helping AI understand clients' products, releases, competitors, and teams with consistent labels and relationships.

Learning Loop: Researching a company and labeling to understand business

💨 Myth #3: "The outputs are inconsistent"

Reality Check: True, but so are your prompts and inputs.

Employees provide wildly different quality levels in prompts. We move too fast, give AI half our thoughts, and expect it to read our minds. Also, don't expect generic AI to sift through terrible content. In 20+ years, I haven't met a single company that says their content is in great condition.

The Instruction Loop

Instruct AI to use chain-of-thought, then interrupt it for required inputs.

If someone asks for "research on [competitor]," have the AI ask clarifying questions like "what is this for?" before responding. If you've trained AI on your business, it can know who that company is and generate the right depth based on who's asking.

The Content Loop

What about content struggles? Good news—LLMs are excellent at helping you curate your best content and uncover outdated information. Have AI search for outdated content, and experts promote the best to a review group.

Instruction Loop: Chain of thought based on generic prompt.

🦄 Myth #4: "LLMs hallucinate and plagiarize"

Reality Check: This is true, but manageable.

All major LLM players have been trying to improve hallucinations, some are even degrading over time. The challenge? We trained the “giant plagiarism machine” to please us by curating and regurgitating the equivalent of fifty thousand Wikipedias in a matter of seconds.

It's up to each company and their users to acknowledge and manage this.

The Trust Loop

Focus your LLM on trusted content and ensure employees know it's their job to make it better. Set LLM temperature low (limits hallucinations but also creativity). Instruct it to focus on certain sources and show all sources to users. When it makes something up, have it tell you and use it as a placeholder for humans to fix.

The key for original, trusted outputs? Build verification into your workflows and training. Don't just hope for the best. Train employees to check sources (they can be hallucinated too) and build back in authenticity so you don't sound like every competitor.

Trust Loop: Creating limitations, placeholder stats and citing sources

📱 Myth #5: "LLMs are just another work silo"

Reality Check: LLMs are the biggest opportunity we've had to solve digital fragmentation, ever. The biggest failure in LLM deployment is becoming yet another silo.

At the same time, AI can also create fragmentation with unsanctioned apps accessing your IP and sharing it with competitors. Departmental apps popping up that were built on the LLM you already own, which you can’t properly deploy, and you pay twice for.

The Experience Loop

Use your LLM rollout to unite employee experience across apps.

LLMs were built on APIs and play nicely with other systems. When you deploy a use case or AI agent, integration strategy is incredibly important. Start with widely used apps like Slack, Google, Email and phase in integrations.

If you like adopting tech early, check out MCP (Model Context Protocol) from Anthropic. It's revolutionizing how we connect systems—read/write Notion docs, send Slack messages, all with language commands. Turn it on and buckle up, it’s still a bit bumpy.

Experience Loop: Integrating apps within LLM using Model Context Protocol

😞 Myth #6: "AI will make jobs obsolete"

Reality: Eliminating roles is still a human decision, even if heavily swayed by the bottom line.

For now, humans aren't just “in the loop”—we created it, use it, and should direct the loop.

If only 1% of companies are mature with AI and seeing varying results, firing people prematurely is like betting the company’s future without knowing the odds. Letting people go who have your domain and company knowledge will not get you to maturity faster. Creating a culture of fear around AI will not help people use it better.

There are no workarounds, this is reality, but I know this: even though AI makes things faster, people with AI are your leverage, innovation engine, competitive differentiator, and flywheel to create market momentum.

I'm hosting a small set of events in SF to discuss, and hopefully steer this shift in the right direction, please reach out if you're interested in joining.

Moving Forward with LLMs

To recap, when an LLM falls short of expectations, it's usually because:

• Employees have licenses but weren't trained

• The LLM wasn't properly deployed or tuned

• You're using AI-powered tools with restrictions

• You have countless AI apps that don't work together

• You're asking LLMs to solve problems they can't actually solve

You can start small with an LLM deployment. Your next product launch is perfect for a significant improvement with a small team, for rollout at launch. Remember: better is better.

Through continuous feedback loops between your experts and AI, you'll create a system that increases in value exponentially. The teams that embrace AI as a collaborative partner rather than a replacement will have a big competitive advantage in the years ahead.

PS: This will be outdated fast. Since I started writing, there have been dozens of major AI announcements. We'll need to evolve much faster to keep up.

PPS: I used this loop methodology to write this post:

1. Spoke to Claude project about "6 myths" idea while walking (Human → AI)

2. LLM wrote 1000 words, in my style, with citations in under 2 minutes (AI → Human)

3. I reworked it, gave revision/instructions back to LLM (Human → AI)

4. LLM caught confusing pieces, we both worked out kinks (AI → Human)

5. I realized I wanted more research, LLM scanned new topics (Human → AI)

6. I went way too deep, turning this into a 2K word piece (Human → Human)

7. Claude helped me shorten, clean it up, write Linkedin copy (Human → AI)

That's the loop in action.