Matt named it. Most people are walking past it.
Matt’s argument is simple. Influencers are busy comparing tools and reading last week’s AI news, content that expires before the video finishes rendering. Meanwhile, every “connect your email, connect your workflow” recommendation routes more of your business through someone else’s pipeline.
Big AI versus sovereign AI. Dependency versus ownership.Matt Murphy, @mattmurphyai
He also calls out who’s giving the advice: people who have never owned a P&L, sat in a compliance audit, or explained to a client why their data ended up somewhere it was never supposed to be. Fair hit. Admitting you have a prompt is step one. Admitting you don’t know where your data goes is step two.
Training is only one way the data leaves.
“They’ll train on it” gets the headlines. To be fair, some vendors don’t train on business or API data by default, and some do on consumer plans. Check the plan you are actually on. But training is one risk among several:
- Connector scopes. Clicking “connect” grants permissions. Read access to a full inbox is a map of your clients, your pricing, and your deals.
- Retention. How long are prompts, files, and logs kept? Who at the vendor can see them? Which subprocessors touch them?
- Contracts. In healthcare, if protected health information (PHI) touches a vendor, that vendor needs a Business Associate Agreement (BAA). No BAA, no PHI.
- Dependency. Price changes, model retirements, outages, and policy updates all happen on someone else’s schedule. If your workflow can’t run without them, you don’t own it.
Four layers. You can own every one.
“Sovereign AI” sounds like something only a government can afford. In practice, it comes down to four layers:
| Layer | Big AI default | Sovereign version |
|---|---|---|
| Model | Hosted model behind someone else’s API | Open-weight models running on your hardware through a local runtime such as Ollama |
| Data | Uploaded into a vendor’s workspace | Your own database and vector store, such as Postgres with pgvector. Documents never leave. |
| Tools | Third-party connectors with broad scopes | MCP servers you write and control, scoped to exactly what the task needs |
| Hardware | Rented, metered, shared | A workstation or small server with a modern GPU, more capable than most people assume |
We run this kind of stack ourselves. A single workstation handles document search, summarization, drafting, and agent workflows for a lot of day-to-day work. See Less chat. More decisions. and Before you connect an agent, read the tool list. for hands-on starting points.
Sovereign doesn’t mean anti-cloud.
Here is where we’d add nuance to the fork. Local models are excellent at a lot, but frontier models are still stronger at some hard reasoning and coding work. Pretending otherwise is its own kind of influencer content.
The answer is a routing rule:
- Sensitive data (PHI, client records, financials, pricing, strategy) stays on infrastructure you control.
- Non-sensitive work (public research, generic code, drafts with no client data) can go to a frontier API when quality matters.
- The decision is explicit. Data gets classified before it moves, not after.
That’s the real meaning of sovereign: you decide what leaves, on purpose, every time.
In healthcare, the bar is already set.
If you work with claims, members, or patients, the rules that already govern your reporting apply to AI too:
- Minimum necessary. The model gets only the data the task needs. Aggregate when you can.
- Access controls carry through. An AI assistant must never become a back door around row-level security.
- Audit trails. You should be able to answer who asked what, which data was used, and where it went.
- Agreements before data. BAAs and data-use terms are settled before the first prompt, not after.
None of this replaces your privacy, security, and legal teams. It gives them something defensible to review.
Run this against your own stack.
- Which AI tools are connected to your email, files, or CRM right now, and what scopes did you grant?
- For each vendor: do they train on your data on your plan, how long do they keep it, and who are their subprocessors?
- If you handle PHI, do you have a signed BAA with every AI vendor that could touch it?
- If your main AI vendor doubled its price or went down tomorrow, which workflows would stop?
- Do you have a written rule for which data can leave your environment, and does your team actually follow it?
Can’t answer at least four with confidence? Then an influencer probably chose your side of the fork for you.
The question is whether you chose it, or whether an influencer chose it for you.Matt Murphy, @mattmurphyai
What this note is, and isn’t.
What it is: an opinion and practice note expanding on Matt Murphy’s reel, based on our experience building with healthcare data and running local models.
What it isn’t: a product test, a benchmark, or legal or compliance advice. Vendor terms vary by plan and change over time, so check the current terms for the plan you use. Compliance decisions belong to your privacy, security, and legal teams.
Start with the original.
Original video and the “Big AI vs. sovereign AI” framing: Matt Murphy (@mattmurphyai) on Instagram. Shared with permission.
Want help building AI you own? Talk to JABSystems.