I spent some time last week with a CIO whose business operates in more than a dozen jurisdictions, and we barely talked about which frontier model to use. What we kept talking about was the deployment question: cloud or on-prem, which region, which cloud platform provider, and what happens to all the IP he builds on top of a platform? "Where to run" is becoming the harder call, especially for companies operating in multiple jurisdictions.
Many of his customers operate in countries that require data to stay within their borders, and the hyperscalers don't necessarily have a data center in every one of them. His existing data estate has significant tech debt and would be (too) expensive to move to cloud, so keeping it where it is and running models alongside it suddenly looks like the cheaper, faster, lower-risk path. That's why open source models like Llama and Gemma have quietly become enterprise-relevant: they let you run inside your own four walls when the regulator insists on it.
The deeper issue this CIO is worried about is lock-in. Low-code AI platforms are credible products that get you off the ground fast, but the agents you build, the prompts you tune, and the workflows you wire together don't come with you when you leave. In choosing a platform, you're not only choosing a model, you're choosing whose ecosystem your institutional knowledge lives inside. For a business operating in jurisdictions with different and changing regulations, that's not a hypothetical risk.
A number of powerful frontier models seem to be playing tag in terms of which one is 'the best' at any given time, however this choice is not necessarily as important as it seems. LLMs can be swapped out, you set them up behind an API endpoint and off you go. Also, the building of agents is being commoditized, what once may have been a complex task is easy now. However, what has not been made easy despite all the technological progress is the work of deploying these innovative solutions into a real enterprise.
Compliance, legal, cyber security, sovereignty, change management aren't optional, and they don't accelerate just because the model does. Without a focus on these, we'll keep seeing more Agentic AI programs and MVPs fizzle and not get to scale.
Planning ahead when launching a new agentic solution in a company is key; doing this early and in parallel with the tech development is an absolute prerequisite to success and ROI.