The problem with frontier models

Who would you rather hire for your business?

The genius that's capable of remarkable works of creativity (when they show up)?

Or the solid 9-5 workhorse who turns up day after day while patiently listening and learning from every lesson you give?

It depends on the role right? For most roles, we generally want reliability and trustworthiness.

So what does this have to do with AI automation you might ask?

Frontier models — your unpredictable geniuses

If you don't work with AI day to day, then it's easy to think the AI world consists of ChatGPT, Claude and Gemini.

These are the frontier models. They are the most "intelligent", capable of huge leaps of creativity. Why wouldn't you "hire" these to handle work in your business? In the eye of the AI hype storm, companies rush in and gleefully build their first proofs of concept on top of their API's. Everyone rejoices.

Fast-forward six months however and most these proof of concepts are well...still proof of concepts. Everything keep breaking. Reliability seems a million miles away. Trust starts fading. C-Suite starts getting angry.

How can they see such glossy social media posts and yet their company cannot reliabily automate a simple workflow? It can lead to a break in trust between the technology team and the business.

What the hype machine on social media doesn't tell you is that behind the scenes, these models are constantly changing as labs frantically battle to be top in benchmarks mostly irrelevant to your business.

It makes them extremely brittle for what most of us need — simple, reliable workflow automation. With a frontier model, what works today might break tomorrow as GPT5 revision 199 moves to revision 200 overnight.

When you look at the above illustration you might being to ask the question - "How did we ever expect this to work when building on top of these foundations?".

So, what is the solution?

Enter the small language model

At VerticalAI we help companies automate repetitive voice workflows. Think of those annoying time consuming tasks like "invoice lookup" and "stock query".

This puts us right at the coalface of customer interaction. The tasks are not complex but they require the certainty and stability that frontier models cannot provide. So, we pivoted to small language models. These are the less intelligent cousins of the household names like ChatGPT and Claude. They don't posseess the same creative range but that suits us just fine. That lack of range makes them great at being drilled in doing one task really well.

We take all of our customers historic calls and use the data to fine-tune an open-source model. We then simulate all your historic calls against it to ensure that the model handles it as well as your staff would.

And when it's live? No API changes to worry about. The model is bundled with your product. It gets updated when YOU want. The only benchmark that matters to it is the one you decide. And the kicker? It can run in our customers cloud or on-premise servers, ensuring their valuable data and IP is securely locked up.

Less hype, more old-fashioned engineering.

---

At our consulting practice in Perth, Western Australia, we help companies analyse workflows, find the best candidates for automation and reliably set about executing. Reach out to us for a chat.

Email us at: hello@verticalai.com.au

Visit our website at: https://verticalai.com.au