SRA AI obligations library

Every SRA AI obligation, mapped to the evidence a firm must produce

The SRA has written no AI rules and has announced no intention to. It regulates outcomes, not tools, so every existing obligation applies unchanged when the work is done by a model. This is the full library we built from that position: each control traced to the instrument it comes from, quoted in full, and marked for how much weight it carries.

0Controls
0Distinct obligations
0Governance domains
0Red-line controls

On the sources

Not every source carries the same weight

A Code paragraph and a paragraph in a Risk Outlook report are not the same kind of thing, and a library that flattens them is misleading. Every obligation is marked with the status of the instrument it comes from.

BINDING59

Binding rule· obligations

The Principles, both Codes of Conduct and the Rules. A breach is actionable in itself.

GUIDANCE6

SRA guidance· obligations

The formal guidance series. The SRA may have regard to it when exercising its regulatory functions; a departure needs justification.

COURT10

Court authority· obligations

Judgments, binding to the extent of their ratio. The sharpest standard currently applying to AI-assisted work.

INDICATIVE14

SRA publication· obligations

Risk Outlook reports, Innovate resources and SRA blogs. Evidence of regulatory expectation, not a rule you can be sanctioned under directly.

ADJACENT7

Non-SRA regime· obligations

Obligations outside the SRA's arrangements that bite on the same conduct: the ICO, the EU AI Act, the PII minimum terms. A breach can also be an SRA breach.

The twelve domains

Everything the SRA expects on AI reduces to twelve domains

Each domain is anchored in binding text and sharpened by guidance and case law. It maps to a heading in the SRA's own five-principle AI taxonomy, and answers one plain question a firm should be able to answer about itself.

The library

Explore every control and the obligations behind it

Filter by domain or by the kind of instrument, or search the full text. Each control links to the obligations it rests on, and each obligation back to the controls that rely on it.

Showing 70 of 70

The credibility test

What is often asserted, and is not true

Every one of these circulates in the market. Each is wrong, and each is checkable in under a minute by a COLP. A library that repeats them discredits everything else it says, so we set them out plainly.

The SRA has issued an AI warning notice

False. None of the SRA's warning notices concerns AI or technology.

The SRA has published AI guidance

Nearly false, and newly true in one narrow respect. The only formal-guidance text addressing AI is the Effective supervision guidance as updated on 12 June 2026. Everything else is a Risk Outlook, an Innovate resource or a blog.

The SRA ran a thematic review on AI

False. The AI publication is a Risk Outlook report, a different and lighter instrument. Thematic reviews exist on other topics.

SRA rules require AI disclosure in every retainer

False. The word used is should, with the method expressly left to the firm. Always is used only for direct client-facing AI interfaces.

Firms must offer a non-AI alternative

Overstated. The Compliance tips page gives this as advice and a question, not an obligation.

There is an SRA guidance page on outsourcing

False. No such page exists. The requirements are distributed across the Codes.

There is an AI incident reporting obligation

False. The reporting guidance says nothing about AI, technology or cyber. The general serious-breach test applies.

The SRA mandates AI training

Not supportable as a technology-specific rule. It runs through the general competence obligations, which do bite.

There is no SRA sandbox

Now misleading. The SRA joined the Advisory AI Growth Lab on 8 June 2026, with legal services the first sector; applications opened on 3 August 2026 and close on 27 September 2026. It is a DSIT-led cross-economy sandbox in which the SRA participates, not an SRA-owned one.

A civil procedure rule requires AI disclosure to the court

False. Only the Upper Tribunal (IAC) judicial review claim form carries a standing certification. The Civil Justice Council is leaning against new requirements for professional drafting.

AI exclusions are entering solicitors' PII

Not found in the UK market, and impermissible in the minimum-terms layer. The real development is that underwriters have added AI questions to proposal forms.

The LSB issued a 2025 or 2026 AI policy

Not located. The LSB's AI material dates to April 2024.

There are 1,600 or more UK AI hallucination cases

That is a global tracker figure. UK trackers suggest around 67 as at July 2026, itself unverified.

The DUAA abolished Article 22, and took effect on Royal Assent

False on both counts. Articles 22A to 22D replaced it, the special-category restrictions survive, and the automated-decision changes commenced on 5 February 2026.

Questions

The library, answered.

How verification, governance, and the agents work inside an SRA-regulated firm.

Still have questions? Talk to the founder
No, and it has announced no intention to. The SRA regulates the outcomes a firm achieves, not the tools it uses, so its existing Principles, Codes of Conduct and Rules apply in full when the work is done by a model. This library decomposes that position into 70 controls and the 96 distinct obligations they derive from.

From the blog

The obligations in practice

What the regulator can ask for, and the checklist behind it.

See which of these your firm can evidence

The 10-minute readiness check maps your firm against these controls and returns a prioritised, source-cited gap report your COLP and COFA can act on. Or talk it through with founder Daman Kaur, who reads every reply. Your data stays in the UK or EU.