The issues with a ton of these use cases is direct tying to avoid middleware, which is absolutely NOT something one should be doing with AI. At best AI should be used as a speech to text processor that has the optional ability to provide a formatted output, like say JSON. Past that, you are doomed to deal with issues that AI will not handle correctly.

That’s the key issue, once you have to have some logic layer in between, the cost of these things starts flying northward and the savings you’ll get from replacing around 8,000 minimum wage workers is nowhere near what you though you would be saving. So these companies believe that they’ll just bake in the logic into the system prompt. Which, LOL, gets you $15,000 priced combos.

The Taco Bell one is actually about as well as it should handle for a lazy deployment. The issue is that Taco Bell isn’t ready for everyone in the world to be bypassing their AI system and the AI system doesn’t have any good fall backs. There’s just a thin layer that says when something goes outside of parameters to trigger a human being, that’s striking too cautious a balance.

AI isn’t a replacement for humans, it’s no where near that ability at the moment. But if you think about say an IVR system, as long as you are keeping parameters incredibly narrow, you can fit AI somewhere in there. But that rarely justifies the cost, especially if you’ve already got a well developed IVR that you’re still dealing with.

That’s the biggest thing, there’s a lot of promise in these things. But we’re not there. We are not anywhere close to these things replacing 100% humans taking orders. If you’re a place that’s not yet built out an IVR, the AI agent might be something to look into, IF you have programmers on your team that can do the heavy lifting for the logic layers. In short, treat these things like a Small Language Model plus serialization, that’s about as far as they are good at the moment. Trying to toss logic into these things is just going to induce pain and failure along the way. And if that doesn’t save you enough money to justify the super high cost per credit, then don’t do AI. There’s a very small application for AI in public facing things at the moment.

The only thing I’ve seen AI at the moment be really good at is RF signal processing. It wouldn’t surprise me if down the road we saw some sort of model embedded for boosting Wi-Fi signal processing. But this kind of application with Taco Bell et al. No. At least not right now. Or not enough to justify it to most companies at this time.