Foundation
UK GDPR and the Data Protection Act 2018
The UK General Data Protection Regulation, sitting alongside the
Data Protection Act 2018, remains the primary statutory framework
governing the processing of personal data in the United Kingdom.
Where a regulated organisation processes client personal data through an AI
service, the framework applies in full — the technology does
not displace the obligations.
Controller and processor roles
An organisation using an AI service to process client data is the data
controller. The AI provider is, in most configurations, a data
processor — or, in the case of consumer AI services where the
provider determines the purposes of processing (training, model
improvement, telemetry), arguably a co-controller. Establishing
which is which is a foundational compliance question, and it is
frequently unanswerable for consumer AI tools.
Article 28: written contract with the processor
Article 28 requires a written contract between controller and
processor setting out subject matter, duration, nature and purpose
of processing, type of personal data, categories of data subjects,
and the controller’s obligations and rights. The contract must
bind the processor to confidentiality, security, sub-processor
authorisation, audit rights, and assistance with data subject
requests. Standard consumer AI terms of service do not satisfy
Article 28.
Article 32: security of processing
Article 32 requires the controller and processor to implement
appropriate technical and organisational measures to ensure a level
of security appropriate to the risk — including, where
relevant, pseudonymisation and encryption, the ability to ensure
ongoing confidentiality, integrity, availability and resilience of
processing systems, and the ability to restore availability and
access in the event of an incident. Demonstrating Article 32
compliance for a public AI service used on client data is
structurally difficult: the organisation has no direct visibility of the
processor’s controls.
International transfer restrictions
Chapter V of the UK GDPR restricts the transfer of personal data
to third countries unless an adequate level of protection is in
place — through an adequacy regulation, appropriate safeguards
(e.g. the International Data Transfer Agreement or the UK Addendum
to the EU Standard Contractual Clauses), or a derogation. Where
client personal data is processed by an AI service whose corporate
group sits in a jurisdiction without an adequacy regulation, a
transfer mechanism must be in place — and the organisation must be
able to evidence it.
Why consumer AI platforms create structural compliance problems
Consumer AI services are configured for individual users, not
regulated organisations. They typically lack: an Article 28 processor
contract on terms a controller can rely on; a documented
sub-processor register; transparent data-flow documentation;
contractually enforceable data residency; the ability for the
controller to evidence Article 32 measures; and a defensible
international transfer mechanism. Each of these is, individually, a
compliance problem. Together they describe an environment in which
the organisation cannot demonstrate accountability under Article 5(2).
How sovereign self-hosted architecture supports compliance
A sovereign, self-hosted deployment narrows the processor chain to
two named entities: the infrastructure provider and the integrator.
Both are contractable on UK GDPR terms. The infrastructure sits
within a jurisdiction the organisation has chosen. The data flow is
documented at the network level, not inferred from a privacy
policy. The architecture is built to support — not guarantee
— UK GDPR compliance: the residual obligations remain with
the organisation as controller, but the structural blockers that consumer
AI tools create are removed.
New Statute
The Data (Use & Access) Act 2025 (DUAA 2025)
The DUAA 2025 amends and supplements the UK data protection
regime. For regulated organisations using AI on personal data, three
aspects warrant particular attention: a strengthened technical
audit power, recalibrated enforcement mechanisms, and a sharper
focus on the documented, demonstrable basis for automated
processing.
Compulsory technical audits
The Information Commissioner’s Office may compel an organisation
to commission and fund an independent technical audit of its data
processing activities, including AI processing. Where an organisation cannot
produce documented evidence of which AI services its staff use on
client data, what categories of personal data are processed, and
what controls apply, the audit becomes a discovery exercise on the
organisation’s balance sheet.
Enforcement mechanism changes
Penalty calculations, procedural timeframes, and the categories of
breach attracting enhanced enforcement have all been revisited. The
headline figures (the higher of a fixed cash cap or a percentage of
worldwide turnover) remain consistent with the pre-existing UK
regime, but the procedural runway available to an organisation under
investigation has been shortened.
Implications for organisations using uncontrolled AI tools
The DUAA 2025 does not introduce a separate ‘AI offence’.
What it does is render the absence of a defensible AI governance
posture more visible, more quickly, and at greater cost. Where
staff are using consumer AI services on client data without
organisation-level controls, the organisation holds the regulatory exposure —
and increasingly bears the audit cost of evidencing or refuting it.
A defensible audit trail
Documented sovereign infrastructure produces an audit trail by
design. The data flow is fixed at the network level. The processor
chain is two named entities. The access pattern is logged on
organisation-controlled infrastructure. Where the regulator asks ‘what
AI processes your client data, and how do you know’, the organisation
has a written, evidence-backed answer.
Regulator Posture
ICO Enforcement on AI
The Information Commissioner’s Office has, since 2023, made AI
enforcement a priority workstream. Public guidance, published
opinions, and reprimands point to a consistent posture: the organisation
remains accountable for personal data processed through AI, the
standard of evidence required is rising, and the ‘the AI did
it’ defence is not available.
What the ICO expects
The ICO has made clear that organisations using AI on personal data
should be able to evidence: a lawful basis under Article 6 (and
where relevant Article 9); a Data Protection Impact Assessment
(DPIA) where the processing is likely to result in high risk;
documented Article 28 contracts with AI processors; an audited
sub-processor chain; and demonstrable Article 32 measures.
Generic terms of service and unwritten internal practices do not
satisfy these expectations.
DPIA expectations
For most AI use cases in regulated work — in particular
those involving client-confidential material, special category data,
or systematic processing — a DPIA is the expected starting
point. The ICO’s guidance treats the absence of a DPIA in
high-risk AI processing as itself a compliance issue, separate from
any underlying data incident.
Cross-border AI data transfers
The ICO’s position is that international transfers of personal
data to AI services follow the same Chapter V analysis as any
other transfer. A US-headquartered AI service processing UK client
data triggers a transfer assessment, regardless of where the
provider claims its servers are located. Adequacy, appropriate
safeguards, or a derogation must be in place — and the organisation
must be able to evidence the assessment.
Jurisdictional Architecture
The CLOUD Act — why geography is not jurisdiction
The US CLOUD Act 2018 (Clarifying Lawful Overseas Use of Data Act)
establishes a single operative principle: jurisdiction follows
corporate ownership, not geographic server location. A US court can
compel any US-incorporated technology company to produce data held
anywhere in the world — including data sitting in a UK data
centre. AWS, Microsoft Azure, and Google Cloud all operate within
the UK. Their parent companies are US corporations. Every prompt,
every document, every conversation processed by those platforms is
subject to potential US court process served on the US parent
— and the demand may arrive under a gag order that prevents
the provider from notifying the UK client that it happened.
This affects every US-headquartered AI vendor operating in the UK
— without exception. The specialist legal-AI tools are no
different: however polished the product, most run on US-incorporated
model providers and cloud platforms underneath, which is what
determines CLOUD Act reach — not the brand on the box. The
‘sovereign cloud’ variants of the major US platforms
— marketed at a premium — do not resolve the problem.
Their parent companies remain US corporations. The CLOUD Act applies
to US corporations regardless of where they have physically located
the data. No contractual arrangement, Customer-Managed Keys
configuration, or sovereign cloud product can override the
extraterritorial reach of a valid US court order. This creates a
direct conflict with UK GDPR Articles 5, 32, and 48
— and it cannot be mitigated by DPA language alone.
Public confirmation by US technology executives
The position is not contested by the providers themselves. Senior
executives of major US technology companies have confirmed, in
public statements and in testimony before US legislative bodies,
that their organisations comply with valid US legal process for
customer data — including data held overseas. Microsoft has
publicly defended the CLOUD Act framework and the company’s
obligations under it. Where a US authority issues a lawful order to
a US-incorporated provider, the provider produces the data; the
UK-located customer is not a party to the proceeding and may not
even be notified.
The architectural answer
Private AI deploys exclusively on infrastructure whose
owner is outside US jurisdiction — a UK-registered,
UK-operated provider for UK deployments, and a European-owned,
EEA-operated provider for EEA deployments. Neither has a US parent
that a US court could compel under the CLOUD Act to produce your
data. Your client data, on dedicated infrastructure with no outbound
path, sits beyond the reach of a US order served on an American
corporation — because there is no American corporation in the
chain holding it. For the organisation, its clients, and its DPO,
this is the architecture the CLOUD Act simply does not reach.
Read the full legal analysis →
Professional Conduct
SRA — the Solicitors Regulation Authority
For solicitors and law firms regulated in England and Wales, the
SRA Standards and Regulations apply in full to AI use on client
matters. The SRA has, in published thematic work and guidance,
made clear that its existing principles — competence,
confidentiality, integrity, and the duty to act in clients’
best interests — bind AI use without modification.
Thematic review findings
The SRA’s thematic work on technology and AI in legal practice
has identified a consistent pattern: AI adoption is widespread, AI
governance is patchy, and the firms most exposed are those whose
fee-earners use consumer AI services on client matters without firm
oversight. The regulator has signalled that it will take an
increasingly assertive supervisory posture in this area.
Conduct obligations on AI use
A solicitor remains personally accountable for work product
generated with AI assistance. Confidentiality obligations under the
Code of Conduct apply to information disclosed to an AI service
exactly as they apply to information disclosed to any other third
party. Where the AI service’s terms of use are inconsistent with
the firm’s confidentiality undertaking to its client, using
that service is not a defensible position.
The ‘no AI did it’ defence does not exist
The output of an AI tool, used in legal work, is the work of the
firm and the individual who relied on it. Errors, hallucinations,
fabricated citations, or breaches of confidentiality flowing from
AI use are attributable to the firm, not the tool. Defensible AI
governance, in the SRA’s framing, is a written, enforced, and
auditable position on what tools may be used, on what data, by
whom, with what review.
Scottish Profession
Law Society of Scotland
For solicitors regulated in Scotland, the Law Society of Scotland
and the Standards of Conduct govern professional practice. The
Society has issued AI-specific guidance and continues to develop
its position as adoption accelerates. The underlying duties —
particularly confidentiality and competence — are not
materially modified by the technology, but the practical
application of those duties in an AI context is a live area.
Confidentiality
The duty of confidentiality owed to a client under the Standards of
Conduct extends to any handling of client information —
including disclosure to a third-party AI service. Where the AI
service’s terms permit the provider to retain, use, or train on
submitted content, the firm must be able to demonstrate either
that those terms are not engaged for the firm’s use, or that
the client has given properly informed consent.
What the Society expects firms to demonstrate
Practically: a written AI use policy; documented review of the AI
tools sanctioned for use on client matters; a defensible position
on data residency and processor jurisdiction; awareness of the
interaction between AI tooling and the firm’s confidentiality
obligations; and, where AI is used in advice, an evidenced human
review step before output is communicated to the client.
Cross-Border Reach
EU AI Act — implications for UK and Scottish organisations
The EU AI Act is a regulation of extraterritorial reach. UK organisations
are not automatically out of scope; the Act applies to providers
and deployers of AI systems where the output of the system is used
within the European Union. For UK and Scottish organisations with EU
clients, EU work, or EU-based personnel, scope must be assessed
on the facts.
Scope for UK organisations
A UK organisation using an AI system to produce output that is used in the
EU may, depending on configuration, fall within the deployer
obligations of the Act. The relevant questions are where the
intended output is used, what role the organisation plays in the AI value
chain, and which risk classification applies to the system.
High-risk classification for legal AI
Several use cases relevant to professional services —
including AI used in administration of justice and certain
employment-related decisions — are within the high-risk
category. High-risk classification triggers a substantial set of
obligations: risk management systems, data governance, technical
documentation, record-keeping, transparency, human oversight,
accuracy and robustness, cybersecurity, and conformity assessment.
Practical obligations
For organisations within scope: maintain documentation sufficient to
demonstrate human oversight of the AI system; ensure logs are
generated and retained; assess and document the data used; and
implement appropriate technical and organisational measures to
manage the identified risks. Each of these is materially easier to
evidence on infrastructure the organisation controls than on a consumer
AI service.
How sovereign architecture supports compliance
A self-hosted deployment produces the documentation, logging, and
oversight artefacts the Act expects as a side-effect of normal
operation. The data used by the system is the data the organisation chose
to expose to it. The model and configuration are versioned. The
access pattern is logged. None of this guarantees compliance with
a regulation as broad as the AI Act — but it removes the
structural blockers that uncontrolled tools create.
National Direction
Scotland’s AI Strategy 2026–2031
Published by the Scottish Government in March 2026, the AI strategy
for 2026–2031 articulates a national position on AI adoption,
governance, and investment. For regulated data-sensitive organisations
in Scotland, two aspects warrant attention: the direction of
public-sector procurement, and the signalling effect on what
‘responsible adoption’ is taken to mean by Scottish
stakeholders.
The strategy frames AI infrastructure as strategic national
capability. For organisations whose clients include Scottish public bodies
or Scottish-incorporated businesses with public-facing obligations,
the practical effect is that ‘sovereign, documented,
governed’ is becoming the assumed posture rather than the
exception.
UK Direction
UK Sovereign AI Fund
Separately from the Scottish strategy, the UK Government launched
the Sovereign AI Fund in April 2026 — a commitment in the
order of £500m to back UK-based AI compute, models, and supporting
infrastructure. The fund is a UK-wide instrument and sits under
Westminster economic policy, not the devolved Scottish strategy.
Read together, the two interventions point in the same direction:
the UK as a whole is moving towards sovereign AI as the default
posture for sensitive workloads. For regulated organisations, the relevant
inference is not the headline figure but the trajectory —
public-sector procurement, regulator expectations, and client
due-diligence questionnaires are all converging on the same set of
assumptions about where AI infrastructure should sit and who should
be able to compel access to it.
Insurance
Professional Indemnity Insurance
Professional indemnity insurers are now asking AI questions at
renewal. The questions vary, but the underlying inquiry is
consistent: what AI tools does the organisation permit, on what data, with
what governance, and what residual risk does the organisation retain.
What a defensible answer looks like
A defensible PII answer is documentary: a written AI policy; a
named, sanctioned AI environment for use on client matters; a
described data-flow; a known processor chain; and a position on
shadow AI use that is enforced rather than aspirational. Insurers
assess risk, not aspiration.
How sovereign infrastructure supports renewal
A sovereign deployment produces, by design, the artefacts an
insurer is asking for: deployment architecture, sub-processor
register, DPA with the integrator, data-flow documentation, and
access logs. The renewal conversation moves from ‘tell us how
you control AI risk’ to ‘here is how AI risk is
controlled’.
