Charles Camp

AI Architect building production AI systems. Agents, RAG, knowledge architecture & evaluation. ex-Credit Suisse.

[Charles Camp @CharlesCamp_]

Sovereignty means the other party cannot do something.

Not that they promised not to.

Every law firm and financial institution using third-party AI models has made a sovereignty trade they may not fully understand.

Sending sensitive documents to OpenAI or Claude with a data processing agreement is not sovereign AI.

It’s trust.

And trust is not architecture.

Think about energy.

If your country imports energy from a neighbor, an agreement doesn't make you energy sovereign.

It makes you dependent.

If the neighbor changes policy — or simply lies — you find out when the power stops.

AI infrastructure works the same way.

Your client list. Your deal flow. Your litigation strategy. Your compliance gaps.

All processed by a model you don’t control.

The cloud tradeoff itself is legitimate.

Running infrastructure is complex and expensive. For many firms the operational burden isn’t worth it.

But paying OpenAI for tasks that open-source models already perform is a different decision.

The capability gap has mostly closed.

The sovereignty gap has not.

The real question was never cloud vs no cloud.

It’s whether the system processing your most sensitive data needs to be someone else’s.

[Charles Camp @CharlesCamp_]

Six months before Covid, I asked to work remotely. The answer was no.

That refusal wasn't the reason I left. But it made the problem clear.

I wanted control over three things: location, time, and what I worked on.

At Credit Suisse, I got to lead projects instead of just executing them. Still, the broader direction came from above. Someone else set the constraints.

Freelancing changed that.

Location became a choice. I can spend a few months in Madrid while continuing to work with US clients.

I have much more control over my time.

And I choose the missions I work on. I get to sit with founders and executives from day one, defining the strategy and architecture instead of inheriting someone else's decisions and building within them.

Years later, those are still the three things I value most: where I work, when I work, and what I work on.

Remote work was never really the goal. Having control over my work and where it lets me live was.

[Charles Camp @CharlesCamp_]

One of the easiest ways to make an AI system fragile is to keep improving its prompt.

Sounds backwards.

But prompts often evolve like this:

If X → do this.
Never Y unless Z.
Except when A → do B.
Every failure adds another instruction. Every edge case adds another paragraph.
Eventually, the prompt stops guiding the model and starts becoming the application.

That's the problem.

Business logic written in natural language is hard to test, version and enforce.

A better separation:

Interpretation → LLM
Business rules → code/workflows
Domain knowledge → knowledge layer

The LLM should handle what requires judgment. Not everything that's convenient to put in a prompt belongs there.

[Charles Camp @CharlesCamp_]

Giving an AI access to your company's knowledge doesn't give it your company's expertise.

You can connect the documents, build the RAG pipeline, and retrieve the right policies.

The system still doesn't know how your best people decide.

Experienced employees know which evidence matters, when an exception is justified, when signals conflict, and when something should be escalated.

That's not knowledge. That's judgment.

And you don't capture it by adding more documents to a vector database.

You have to decide where that judgment belongs:

Rules → deterministic decisions
Knowledge structures → relationships and context
LLM → interpretation
Humans → decisions the system shouldn't make

The hard part of enterprise AI isn't giving the model more knowledge.
It's designing how that knowledge becomes a decision.