CLM vs. LLM: To Build a Walled Garden or Not?

Using Google Deep Research to Help Me Understand the Pros and Cons of Custom Language Models

AI-Created TL;DR Version of the Post

Most marketing teams are using AI tools, but leaders often don't know what data is being shared, risking exposure of sensitive information. Custom Language Models (CLMs) offer a secure, tailored alternative, but come with higher costs and implementation challenges. To explore the trade-offs, I used Google Deep Research to help me understand the pros and cons. CLMs can be a smart option for balancing AI potential and data privacy, but careful consideration is essential.

Full Matt-Written Post

Earlier this week, I talked about starting small with AI. Today, I want to go big and discuss Custom Language Models.

Whenever I talk with marketing teams about their AI use, one thing immediately becomes clear: everyone is using ChatGPT in one way or another. This isn’t surprising – ChatGPT is a really simple, powerful and exciting way to engage with AI.

But the other thing that immediately becomes clear is that marketing leaders have no idea what kind of information their teams are putting into ChatGPT or any other AI tool. It could be something benign like a first draft of a blog post, but it could also be your organization’s strategic plan, with sensitive financial and competitive information, or protected patient or student information.

So what’s a marketer to do? You could do a deep-dive read into the data privacy and usage policies for every AI tool to understand what they’re doing with your data, but we know that isn’t realistic. And the tools themselves make it hard to figure out what’s happening with the data you put in. In ChatGPT, for instance, there’s a toggle switch in Settings that lets you turn off model training. But what does that actually mean? (And is it actually true? OpenAI’s record on copyright protection isn’t great.)

One potential solution I’ve been thinking about for a while is creation of Custom Language Models (CLMs) that give all the benefits of a large, well-trained language model plus the comfort of knowing that your sensitive data won’t end up in the public domain.

What is a CLM you ask? I’m not a technical person, so I think of it as a walled garden; a large language model that’s been walled off, data protected and trained on an organization’s specific information, brand and context. See the graphic I created in the image slider below. 👇

Ideally, this allows brands to use AI models freely, knowing not only that their data is safe, but that the CLM actually knows your brand – it isn’t giving generic solutions that could apply to any business.

But of course there are drawbacks to the idea, like the cost and time of building a custom solution. Like any buy vs. build discussion, there are pros and cons to each side.

So I decided to use the latest flashy AI tool – Google Deep Research – to help me sort it out. If you haven’t tried it yet, Deep Research is a tool that, once prompted, scours the web for relevant information and comes back to you with a comprehensive research report. It’s pretty incredible. (Also check out my post about Google Notebook LM 🤯)

Here’s the prompt I used:

“I’d like to see a research report on the pros and cons of custom language models vs off-the-shelf products for marketers in healthcare and higher ed. Please include elements like brand management, data privacy, cost, implementation effort, etc. Do you need additional info from me to complete this report?”

And here’s the "research plan" the system gave back to me for approval before beginning work:

1. Find information on custom language models and off-the-shelf language models.
2. Find information on the pros and cons of custom language models for marketers in healthcare and higher ed.
3. Find information on the pros and cons of off-the-shelf language models for marketers in healthcare and higher ed.
4. Find information on brand management, data privacy, cost, and implementation effort for both custom and off-the-shelf language models.
5. Find case studies or examples of how custom and off-the-shelf language models have been used in healthcare and higher ed marketing.
6. Based on the information you find, create a research report that compares and contrasts the pros and cons of custom and off-the-shelf language models for marketers in healthcare and higher ed.

This is how I envision the structure of a CLM.

This is the prompt I gave Google Deep Research and the research plan it gave me for approval before it ran off to do its thing.

This is the beginning of the report Deep Research created on the pros and cons of Custom Language Models (CLMs). It took about 10 minutes to complete the report.

Deep Research referenced 68 separate resources to create the report. I didn't give it any of these materials. It found them entirely on its own.

I gave Deep Research the virtual thumbs up and off it went. After about 10 minutes and a review of nearly 70 separate sources (none of which I gave it), it came back to me with this report. It’s quite good, and captures many of the pros and cons of each option.

So what I’m left with after all of this is that CLMs could be a good option for brands that are looking to take advantage of the potential of AI while mitigating the risks. But, as with any other decision involving technology, strategy and data privacy, there are a lot of factors to consider.

I also learned that Google Deep Research and other tools like it that will inevitably follow are incredibly powerful partners to individuals and teams. If you haven’t already, give it a road test and see how it can help you answer some of your stickiest marketing questions.