Issues With Applying AI To Your Business

A lot of businesses struggle to find the most cost-effective solution. You'll have organisations like Microsoft, and even ChatGPT itself, offering products that are much easier to use if you're not technical, but that ease comes at a price, because it drives up the overall cost of using the model.

Here's the simplest way to think about it: the cost of using a model is something you genuinely can't avoid. That part's fixed. But on top of that, there's a whole extra layer of cost, the middleware, essentially, and that's where businesses end up bleeding money. It's worth actually stopping to ask whether you need that extra layer or service sitting in the middle at all.

There's so much hype around these tools right now, and at the end of the day, all most of them really are is another application sitting between you and the model. Sometimes that layer genuinely adds value. But more often than not, it's an extra cost you're paying simply because you don't have the knowledge or expertise to go direct to the model yourself. And the gap isn't small: most "AI wrapper" products mark up the underlying model cost by 3–5x, and in some consumer tools that markup runs as high as 20–60x what the AI itself actually costs to run.¹

Security is a huge issue for AI, full stop. AI leans heavily on plugins and add-on features to become more powerful: a database to store the information you feed it, an MCP, some kind of open-cloud setup with a dozen agents running around. The problem is that a lot of these plugins and tools are compromised, or riddled with security flaws. Recent independent audits found nearly half of tested MCP servers vulnerable to command injection, and over 80% carrying flaws that could let an attacker traverse into files they shouldn't be able to reach. One industry audit put the average security score of the MCP servers it tested at just 34 out of 100.²

That's exactly why the knowledge and expertise to actually vet these tools matters: to know whether the thing you're using is secure or not. Part of the problem is that it's become so easy to build a tool that people with no real expertise are churning out solutions, leaning entirely on the AI to build them. And sure, AI can build something. But in a lot of contexts you still need real expertise to judge whether that solution is secure, whether it introduces flaws, or whether it leaks data, because the consequences are still very real, and they don't land on the AI. They don't land on whoever built the plugin either, since that person typically carries zero liability. That liability lands on the business implementing the solution, or the business itself, and analysts are already projecting that access-control gaps like these will be behind the majority of successful attacks on AI agents over the next few years.²

Then there's effectiveness. One of the biggest issues here is simply the sheer quantity of solutions out there. It's not hard to find an integration or a plugin for practically anything anymore, because it's become so easy to make them, people are pumping them out like it doesn't matter.

That creates a real problem: working out which one is actually the most effective for you. Say you're looking for five different plugins that connect to your email. The one that rises to the top is usually the one with the most hype, the best marketing. Not necessarily the best fit. That's exactly why expertise matters here too. It's what makes it possible to actually work out what's the right solution for your specific problem, rather than just the loudest one.