I am the founder and CEO of one of these startups. The problem space is vast and complex. We have paying customers in production in the 5,000-50,000 employees range and in the 50,000-100,000 employees range. The latter ones have a more urgent need to address. Below 5,000 employees the need is not there right now. Fin-ops and compliance are major drivers for adoption of these control planes for AI right now. Gartner started tracking a new category called AI middleware that is sort of what you are referring to (runtime governance).

This space is probably much harder than you think. While enterprise AI adoption is an inevitable trend, and the demand for data sensitivity and privacy protection is very real, deeply integrating AI with actual enterprise business data comes with a massive number of hurdles.

Take the phenomenon of "Shadow AI," for instance: 1. When AI tools are used across different departments, how do you prevent sensitive data from leaking? 2. With departments adopting all sorts of disparate tools, how do you actually achieve centralized governance and control?

Although Microsoft has rolled out an AI governance framework, I think practical implementation remains an uphill battle. Their approach isn't hard to grasp—it essentially boils down to logging activities before and after an Agent executes tasks to enable traceability, but that doesn't solve the core problem.

So, if you want to build a product in this space, I believe it's going to be extremely difficult unless you can come up with a genuinely new approach that truly resolves these pain points.

As someone building in this space I would say adoption is slow, but interest is growing. We have no regulation in the US so it's hard to convince companies this is something they need to spend money on right now, especially when their mostly still struggling to define the ROI on their AI ambitions to begin with.

That said, the more the frontier models hack external systems and once the lawsuits start flowing the tune will almost certainly change. At least that's what I'm hoping :)

If you're curious in what we're working on you can check it out here: https://lyfe.ninja/

How do these governance systems work in practice when AI models and their behavior can change so quickly?