96% of UK law firms use AI. Most have no governance framework for it. | Insights

A figure emerged from the SRA’s thematic review in December 2025 that deserves more attention than it has received: 96% of UK law firms now use some form of AI.

That number is not a warning. It is a description of the status quo. AI adoption in UK legal practice is essentially universal. The question is no longer whether your firm uses AI — the question is whether you know how, by whom, and on what data.

The answer at most firms is: no.

What the 96% figure actually means

The 96% adoption figure comes from the 2025–2026 survey cycle, corroborated by the SRA’s December 2025 thematic review of AI in legal services. It covers law firms across all size segments — from sole practitioners to large commercial practices.

What it does not capture is the composition of that use. “Using AI” includes:

Managed, governed deployments on sovereign infrastructure, with documented processing purposes and applicable DPAs
Enterprise SaaS tools (Microsoft Copilot, Lexis+ AI, CoCounsel) with formal procurement and at least some governance documentation
Free-tier consumer tools (ChatGPT, Gemini, Claude.ai) used ad hoc by individuals on personal devices or firm machines
Browser extensions, document summarisers, transcription tools, and AI-enhanced search products used without any specific governance assessment

The distribution is not evenly weighted towards the governed end. The SRA’s December 2025 review found that only 1 of 36 Compliance Officers for Legal Practice surveyed could accurately outline all of their firm’s regulatory obligations in relation to AI. One. Of thirty-six.

The 96% adoption figure describes a sector that has deployed AI at scale without a corresponding investment in governance. The productivity benefit is being realised. The regulatory liability is being accumulated.

The governance gap

The December 2025 SRA thematic review documented a governance gap that the 96% adoption figure makes acute.

The gap is not primarily a matter of willingness. Most firms understand, in the abstract, that AI introduces data handling risks. Many have issued guidance to staff about using AI “carefully” or “responsibly.” A significant minority have formal acceptable use policies in place.

The problem is structural: acceptable use policies cannot govern Shadow AI.

Shadow AI — the use of consumer and ungoverned AI tools on client data outside any formal governance framework — has a defining characteristic: it is invisible to the firm. When a member of staff submits a privileged document to ChatGPT on a personal device, there is no audit trail. When a trainee uses a browser extension to summarise a client file, there is no processing record. When a paralegal uses a free transcription tool for a client meeting, there is no DPA.

An acceptable use policy tells staff what they should not do. It cannot tell you what they are doing. It cannot generate the audit evidence the ICO will request. It cannot prevent the data transfer from occurring. Policy cannot solve an architecture problem.

The SRA review found that 62% of UK firms plan to expand their AI use in the near term. Without corresponding governance investment, that expansion will compound the liability, not resolve it.

What the ICO expects

The Information Commissioner’s Office has been clear, if not yet universally heard, about what it expects from organisations using AI with personal data.

The Data (Use and Access) Act 2025 (DUAA) expanded the ICO’s enforcement toolkit in ways that directly affect any firm with a Shadow AI exposure:

Compelled independent audits. The DUAA grants the ICO the power to compel organisations to fund independent audits of their data processing practices — on the basis of mere suspicion of non-compliance, not proven breach. For a firm that has not mapped its AI tool usage, an audit notice is an existential compliance event.

Mandatory interview notices. The ICO can now compel named individuals — including senior partners and Compliance Officers — to attend formal interviews. SMCR-style personal liability is not confined to financial services.

Escalated fines. Maximum fines under the DUAA remain £17.5 million or 4% of annual global turnover. The ICO has stated publicly that AI-related data breaches are an enforcement priority. Fines of this magnitude have already been issued in sectors where the ICO has established enforcement precedent; legal services is the next frontier.

Submitting unredacted personal data — client files, witness statements, HMRC correspondence — to a consumer AI platform is not a borderline case. It is a direct, provable UK GDPR violation with a clear causal chain from the submission event to the data transfer. The ICO does not need to demonstrate intent. It needs to demonstrate that the processing occurred.

The five ICO aggravating factors most relevant to AI data breaches

The ICO has published guidance on the aggravating factors it weighs when determining penalty levels following a data breach. Five are particularly relevant to Shadow AI incidents. Rather than repeat the full analysis here, we direct you to our detailed treatment in the ICO enforcement article. In summary:

Volume and sensitivity of data involved — privileged client files represent the highest-sensitivity category
Duration of non-compliance — Shadow AI typically reflects ongoing, repeated processing over months or years
Whether appropriate technical controls were in place — the absence of any technical governance mechanism is an aggravating factor
Cooperative disclosure vs. discovery by the ICO — firms discovered to be non-compliant fare substantially worse than those that self-report
Senior management awareness — where Compliance Officers or partners knew or should have known of the risk, penalty levels are elevated

Shadow AI incidents typically score badly on all five. The use is ongoing. No technical controls exist. Senior management cannot credibly claim ignorance of published SRA and ICO guidance. The data is the most sensitive the firm processes.

Why acceptable use policies fail

The governance gap identified by the SRA review is not a gap in documentation. Most firms have, or can quickly produce, an AI policy. The gap is operational.

You cannot audit what you cannot see. An acceptable use policy cannot generate the Records of Processing Activities that UK GDPR Article 30 requires. It cannot document the data subjects affected. It cannot identify the AI vendors who have processed the data, let alone verify whether those vendors hold adequate DPAs.

You cannot enforce what has no technical control. When ChatGPT is accessible from any device with an internet connection, an instruction not to use it is an instruction that relies entirely on individual compliance. Enterprise software deployments where compliance depends on individual behaviour fail at rates documented by every relevant survey. The SRA review confirms the failure in legal services.

You cannot reverse what has already occurred. When a member of staff submits a privileged file to a public AI service, the data has left the organisation’s environment. Policy enforcement after the fact does not retrieve it. It does not prevent the ICO from treating the transfer as a breach. It does not restore professional privilege.

There is a further, counterintuitive consequence of AI bans and restrictive policies: they do not eliminate usage. They make it invisible. Every firm that has banned public AI tools has discovered the same thing — usage continues, it just moves to personal devices and unmonitored channels. The risk is not eliminated. It goes underground.

The courts have spoken

The judiciary has begun to apply consequences to AI misuse that firms had not factored into their risk assessments.

In the case of R. (on the application of Ayinde) v Haringey LBC (2025), solicitors were issued wasted costs orders and referred to the SRA after submitting AI-hallucinated case citations to the court. The court treated this as professional misconduct, not administrative error. SRA referral was automatic.

In Hassan v ABC International Bank (October 2025), the court imposed a costs sanction of £5,881 and made an SRA referral in connection with AI-generated submissions that could not be verified to actual law.

More significantly: UK judiciary guidance issued in October 2025 states that any information submitted to a public AI chatbot must be treated as “published to all the world.” Legal professional privilege is waived on submission. This is not a regulatory position — it is a judicial determination about the nature of the act of submission to a public AI service.

The practical consequence: a solicitor who submits a client’s privileged documents to ChatGPT to assist with drafting has, as a matter of judicial guidance, waived privilege over those documents. The SRA and opposing counsel can treat those materials as public.

The only architecture that eliminates the risk

The pattern in every regulatory action, every court ruling, and every SRA finding is consistent: the problem is not AI capability. The problem is architecture.

Firms that use public AI services have no technical mechanism to prevent data egress. They have no audit trail of what was submitted. They have no control over what the vendor does with the data. The governance problem is not solvable with policy because the technical facts do not support governance.

The only architectural model that eliminates Shadow AI risk entirely is one where the AI infrastructure itself is architecturally prevented from transmitting data externally.

Zero-egress sovereign AI achieves this at the network layer. The environment is built with no outbound connectivity — there is no path for data to reach any external AI service, not because policy prohibits it, but because the architecture provides no outbound route. There is no call to an external AI API. There is no prompt transmitted to a vendor’s servers.

When your staff use a sovereign AI workspace, they get the productivity benefit — the AI capability for drafting, research, document review, and analysis — without creating any data transfer event. There is nothing for the ICO to investigate. There is no privileged document leaving the organisation. There is no Shadow AI to govern.

For more on the architectural model, see our detailed overview of zero-egress AI.

What to do next

Step 1: Read our existing analysis of the Shadow AI compliance catastrophe — the full regulatory exposure analysis for UK professional services firms.

Step 2: Understand how a 48-hour Shadow AI audit works — a practical framework for assessing your current exposure before the ICO does it for you.

Step 3: Read our companion article on the CLOUD Act and data sovereignty — the legal architecture that makes any US-headquartered AI tool structurally exposed to foreign government access, regardless of where the data is stored.

Step 4: Talk to our team to walk through your organisation’s AI tool landscape, map the governance gaps, and see what a zero-egress sovereign deployment looks like in your environment.

Private AI Ltd (SC881079) acts as a Standing Data Processor under Article 4(8) UK GDPR, operating a Zero Standing Privileges architecture. We deploy sovereign, zero-egress AI environments for regulated professional services. In the Default Operating State, we hold no active connection and no credentials to client systems.