The essentials in 5 points
- The real question for a regulated professional is not “which model is best” but “where does my data go?”.
- With cloud providers, that question has no satisfying answer: the moment data leaves the machine, it escapes your control.
- Two independent safeguards: keep inference local, and redact personal data before any possible transmission.
- The quality-versus-confidentiality trade-off is gone: consumer 128 GB hardware runs frontier-level models.
- The right architecture is local-first: most queries served on site, the complex minority routed to the cloud with data already masked.
- Local AI also means leaving the token economy: hardware bought once, minimal marginal cost, and you stop paying twice, in money and in data.
The short answer
For a lawyer, an accountant, or a clinician, the question that matters is not which AI model performs best. It is where the documents entrusted to them go.
A client file, an accounting record, a medical report: the moment this data leaves the machine to be processed, it enters a provider’s chain. The professional loses direct control over what protects it.
Sovereign AI turns the problem around. You do not try to promise that “it is secure”, you guarantee that the data goes nowhere, which is verifiable by design. Two mechanisms make this possible: running the model locally, and redacting personal data before any transmission when part of the processing goes through the cloud anyway.
The rest of this article explains how these two safeguards fit together and what they change for professional secrecy and GDPR.
The question cloud vendors don’t like
“Where does my data go?” Put to a cloud AI provider, that question never gets a fully reassuring answer for a professional bound by confidentiality. The provider will document encryption in transit, the location of its servers, its contractual commitments. All of this is useful, but none of it changes the central fact: the data has left the machine.
Once gone, it depends on a chain the professional does not control. Recent incidents and proceedings around the major AI services regularly remind us that the content of exchanges can be retained, analyzed, or shared more widely than the user imagined.
For a law firm or a healthcare professional, this risk is anything but theoretical: professional secrecy is directly at stake. Confidentiality sold as a commitment remains a commitment. Confidentiality guaranteed by the fact that the data does not move is of a different nature.
Two safeguards to take back control
Two mechanisms, independent and combinable, guarantee that data stays under control.
The first is to keep inference local. The model runs on a machine or an appliance the organization owns, within its walls. The data never passes through a third-party server. It is the strongest confidentiality argument there is, because it rests not on a promise but on a physical impossibility: what does not leave cannot leak.
The second is to redact personal data before transmission. In cases where part of the processing goes through the cloud, names, addresses, national ID numbers, and other identifiers are masked upstream, before anything leaves the machine.
The remote model then sees only anonymized text. This filtering makes an occasional use of the cloud acceptable without exposing the people concerned.
The local machine or appliance
The idea of an AI running within the company’s walls long ran into one objection: local hardware would be too weak to run a serious model. That objection no longer holds.
A market study shows this model already exists: consumer hardware with around 128 GB of memory runs frontier-level models locally, the ones that rival the best cloud services. The principle is a dedicated machine or appliance, installed on site, on which the model runs without ever opening an outbound connection with the data.
For the organization, this changes the nature of the offer. Local AI is no longer a degraded compromise accepted out of caution, it is comparable processing capacity, with confidentiality added. The professional no longer has to choose between working with a good tool and protecting their files.
Local-first, cloud as backup
All-local or all-cloud is a false dilemma. The right architecture is local-first, with a fallback to the cloud for the minority of queries that justify it.
A study conducted by a US university estimates that around 70 percent of queries can be served locally. These queries, the most common ones, never leave the machine. The remaining 30 percent, more complex, can be routed to a more powerful cloud model, but only after personal data has been masked.
This combines the confidentiality of local processing on the vast majority of cases with the power of the cloud on the minority that genuinely needs it, without ever exposing identifying information.
This architecture has a double benefit. It protects data by default, and it reduces dependence on a single cloud provider, whose service, pricing, or terms can change.
Leaving the token economy
Cloud AI is pay-per-use: every request consumes billed tokens, and cost tracks adoption. The more your teams use it, the higher the bill. This model makes you pay twice: once in money, once in data, since your content goes off to feed a third party’s chain.
Local inference reverses that equation. The hardware is bought once, the marginal cost of a request comes down to electricity, and usage can grow without the bill following. Your data no longer pays anyone anything. AI becomes equipment, not a subscription.
What it changes for professional secrecy and GDPR
For a regulated profession, local inference shifts the compliance burden. If the data never leaves the premises, the question of transfer to a subprocessor, let alone outside the European Union, no longer arises in the same terms. Professional secrecy is no longer protected by a contractual clause but by the technical architecture itself.
On the GDPR side, the logic is identical: the less data circulates, the smaller the risk surface. Redacting personal data ahead of any cloud call reinforces this principle of minimization.
One caveat is worth naming to the client. Local does not mean automatic compliance: you still have to secure the machine, document the processing, and keep in mind that a local model is not audited by default.
Security has to live in the infrastructure and the procedures, not only in the choice of model.
For the hands-on technical side (processing pipeline, building blocks), see also Confidential AI: Data Stays On-Premises.
Where to start
The right entry point is a scoping exercise: which data is genuinely sensitive, which AI uses touch it, and what share of the processing can stay local. This review leads to a target architecture sized for your needs and budget, with a progressive migration plan.
Colombani.ai designs this kind of sovereign architecture (local inference, personal data redaction, controlled hosting), specifically supports regulated professions bound by confidentiality, and trains teams in these practices through the Local AI course, Qualiopi-certified and eligible for OPCO funding. The expertise is certified by Anthropic (Claude Certified Architect).