The transfer problem, in one sentence
Under the GDPR, sending personal data to a server outside the EEA is a
restricted transfer – and if that server is run by a US company, a US
law called the CLOUD Act can force it to hand that data over, no
matter where in the world the data physically sits.
Why "it’s hosted in the EU" is not the answer
The CLOUD Act (Clarifying Lawful Overseas Use of Data Act) compels US
providers and their parents to disclose data in their "possession,
custody, or control" wherever it is stored – a Frankfurt or Dublin data
centre included. So pasting EU customer data into a US-hosted AI service
is a Chapter V transfer, and the usual fixes only get you halfway there.
Schrems II killed the Privacy Shield and told you a Standard
Contractual Clause plus a transfer-impact assessment must still show the
destination law won’t undermine your protection – which is hard to prove
against the CLOUD Act. It is why supervisory authorities keep asking who
could be compelled, not just who promises not to. The residual
US-access risk is the bit no clause fully erases.
Add the practical exposure and it compounds: contracts, support tickets,
HR files, patient notes and client documents uploaded into US AI SaaS
are personal data leaving the EEA. That triggers processor terms
(Article 28), a record of processing, often a DPIA – and now a CLOUD-Act
analysis on top.
How local AI removes the problem outright
Run the model on your own premises and your own hardware and the
transfer analysis simply never starts. The data is not sent to a third
country; there is no US provider in the chain; nobody outside your
building has "possession, custody or control" of it. No restricted
transfer means no Chapter V chapter to satisfy, and no CLOUD Act hook
for a foreign authority to pull on. You stay the sole custodian. The
question stops being "can we paper over US access with a contract?" and
becomes "does anything ever leave the room?" – and the answer is no.
What that unlocks for a legal or compliance team
Confidential documents – NDAs, commercial contracts, M&A material,
regulated records – can be put through a capable model for summarising,
redlining, searching and drafting while staying inside the jurisdiction
that governs them. Where you genuinely need extra capacity, route it to
EU-controlled infrastructure under EU law, and keep the sensitive tier
local. You pair that with the GDPR basics you owed anyway (lawful basis,
data minimisation, purpose limitation) and with your EU AI Act
duties, which apply no matter where the model runs – local hosting is
not a substitute for those.
The honest caveat
Local AI is not legal advice, and it does not make a company
GDPR-compliant by itself. It structurally eliminates one specific,
hard-to-solve risk – the cross-border transfer and the third-country
access that follows it – which is exactly the risk US-hosted AI cannot
fully contract away. For teams that have been blocked from using AI on
real customer and client data for that reason, eliminating the risk in
hardware is far cleaner than managing it in paperwork.
We install and tune private, on-premises AI on your own machines across
Perpinyà, Girona and Barcelona, so customer and client data can meet a
real model without ever meeting a foreign subpoena. Ask:
privatelocalai@proton.me.
