People throw the term "data sovereignty" around to brag about how much control they have over their data.
But if your data sits in a storage container in Microsoft 365 or AWS, how much control do you really have?
Have you ever just not paid the provider bill and seen what happens?
I remember one time we didn't pay our Salesforce bill on time and our account was suspended.
The entire marketing team couldn't do anything until we swiped the credit card and got the account turned back on.
There was no option to download our data. No temporary access. Nothing.
Just a generic message saying our account was suspended.
And I remember thinking, what happens if you never pay?
Maybe you eventually get an export of your data in a bunch of garbled CSV files. Who knows?
That's the thing about cloud infrastructure.
You may own your data, but the provider controls the infrastructure your data depends on.
That's already a problem for organizations that need absolute control.
It gets much worse with AI.
Putting data in Microsoft 365 or AWS is mostly passive. The data sits there until you retrieve it.
AI is different.
Data can travel back and forth through prompts, conversations, RAG systems, embeddings, tool calls, logs, evaluations, and other parts of the AI stack.
And depending on the provider, the plan you're using, and your configuration, some of that data may also be used to improve models.
So now we're not just talking about where your data is stored.
We're talking about where your data goes.
But there's another problem with AI that I think gets overlooked.
You don't necessarily control the values and principles that influence how your AI behaves.
The model provider has already made those decisions for you.
So you can lose more than data sovereignty.
You can lose value sovereignty.
Your organization has its own policies, principles, risk tolerance, and objectives.
Why should those be defined by the company that provides your AI model?
That's why, if you need maximum control over your data and the behavior of your AI agents, you should consider running your own models in your own environment.
And this is where SAFi comes in.
You can run the models yourself.
You can define the boundaries yourself.
And you can use SAFi to govern the reasoning and actions of your agents at runtime, according to the policies and values you define.
Hosting your own AI models might not be as hard as you think.
And getting SAFi running takes about 10 minutes with the Docker image.
