Fine-Tune, Don’t Reinvent: How Teams Can Actually Use AI Without Burning Out

Most companies don’t need to build their own AI models.

Seriously. Unless you’ve got a massive R&D budget, a deep bench of ML engineers, and a few hundred GPUs lying around, it’s just not worth it.

Sol Rashidi, MBA, a longtime exec in data and AI, laid this out perfectly using a “three-layer cake” analogy in a chat with Joe Reis at GOTO Conferences.  And honestly, it’s one of the simplest, most useful ways I’ve heard AI strategy explained.

Here’s how it goes:

Layer 1: Build from scratch (bottom layer)

This is where you train your own foundational model, like what OpenAI or Google does. Sol’s take? Most companies have no business doing this. It’s expensive, messy, and really easy to get wrong. She’s worked with tons of big enterprises and says she’s never once seen a strong use case for going that deep. So unless your business is AI, this layer probably isn’t for you.

Layer 2: Fine-tune a model (middle layer)

This is the “sweet spot”. You take an existing model, say GPT or Claude, and fine-tune it using your company’s data. You’re not starting from zero; you’re customizing something powerful to work in your world. This is where companies can actually move fast and get results. You still need the right people and tools, but you don’t need to build the engine, you’re just making sure it runs well in your car.

Layer 3: Use existing tools (top layer)

If there’s already a tool that solves your problem, use it. No need to build a customer service bot from scratch if one already exists that plugs into your CRM. Sol made a great point: if someone’s already solved it better than you could, just adopt it. Save your energy for problems that are actually unique to your business.

She calls this approach “Frankenstein stitching”, basically, cobbling together components that already work instead of starting from zero. It’s not about being scrappy, it’s about being smart. I've seen this work well, even in regulated industries, she added. It's less about building every part, more about combining pieces in a way that makes sense for your use case.

But here’s the catch and this part doesn’t get talked about enough:

If your team doesn’t understand AI, none of this works.

Sol’s approach to upskilling is super practical. Don’t start with a formal course or a certification. Start by using AI tools in your own life. Have ChatGPT help you write a tricky email. Use an AI calendar assistant. Try a summarizer for your inbox.

Then, once you’ve seen how it works, dig into what’s behind it. How is it trained? What are the limits? What are good prompts? That curiosity goes further than any textbook.

She also recommends free courses, but only after you’ve gotten your hands dirty. It’s like learning to cook: reading recipes doesn’t help much until you’ve burned a few things and figured out what medium heat actually means.

Playing with Prompts has Transformed My Understanding of AI

I’ve been playing around with prompts myself, just testing how small tweaks change the output. It’s wild how much better the results get once you understand how to talk to these models. 

That led me down a rabbit hole, trying prompts for brainstorming, summarizing dense docs, even mocking up feature ideas. The more I experimented, the more I realized I wasn’t just learning how to use AI, I was learning to think differently. My learning’s accelerated, but more than that, my vision for what’s possible with this stuff has completely expanded.

That kind of learning? It sticks. It’s practical. And way more effective than sitting through a 10-hour course with no real-world application.

Don’t Reinvent the Wheel

So if you're figuring out your company’s AI approach, here’s the playbook:

• Don’t build what already exists.

• Focus on the middle: fine-tune a model that works.

• Let people learn by doing, not by reading.

• Stitch solutions together, don’t try to make everything from scratch.

And above all: keep it useful.

Flashy demos are fun, but results come from real people solving real problems. That’s where the value is.