AI Risk, Governance & Compliance: A Practical Guide for UK Firms
Home » AI Risk, Governance & Compliance: A Practical Guide for UK Firms

AI Guides

AI Risk, Governance & Compliance: A Practical Guide for UK Firms

AI governance in the UK is a patchwork of regulator expectations, data protection law and good security practice, not a single rulebook. This guide sets out what applies today in plain English: the SRA warning notice, data protection, the EU AI Act, ISO 42001, human oversight, prompt injection and professional indemnity insurance.

Book a free AI Capacity Assessment
Take the AI Readiness check
UK legal practice

The short version

  • There is no standalone UK AI Act. Existing law and your regulator’s rules apply, and regulators are using them.
  • Accountability stays with the firm. The SRA’s August 2026 warning notice says so directly.
  • The largest practical risk is AI your people already use that nobody has recorded.
  • Prompt injection may never be fully fixed, so limit what an agent can see and do.
  • A one-page policy, a permissions table, a log and a named owner cover most of what a small firm needs.

In this guide

AI governance UK: where the rules stand

The UK has no standalone AI Act. The government has favoured a pro-innovation, regulator-led approach, and law firm RPC’s June 2026 overview confirms there is currently no standalone AI legislation. A long-promised AI Bill has been repeatedly delayed.

That does not mean there are no rules. It means the rules that apply are the ones you already work under: your professional conduct obligations, UK GDPR and the Data (Use and Access) Act 2025, confidentiality, contract and negligence. Regulators including the ICO, the NCSC and your professional regulator are applying those rules to AI now.

In practice, good governance comes down to being able to show four things: which AI tools are used, on what data, with whose sign-off, and what record exists afterwards. The rest of this guide covers each piece.

What your regulator expects

On 17 August 2026 the Solicitors Regulation Authority published a warning notice called Misuse of AI. It reported 42 reports of potential AI misuse between July 2025 and July 2026, with investigations continuing into inaccurate legal citations, weak supervision and confidentiality breaches.

The central message is that using AI does not change professional standards. Solicitors remain accountable for their work however it was prepared, and firms are expected to have effective governance, systems and controls. The notice also warns that both free and paid AI tools can pose risks, depending on their terms, settings and technical set-up, and it refers to an Upper Tribunal decision, UK v Secretary of State for the Home Department [2026] UKUT 81 (IAC), which raised concerns about client documents being entered into an open AI tool. The SRA notes that sharing confidential client information this way can waive privilege permanently.

The SRA is not telling firms to avoid AI. It is telling them to supervise it. The Bar Standards Board took a similar line earlier in 2026, warning that free AI tools will generally be unsuitable for legal work.

If you are an accountancy practice, consultancy or another regulated or client-confidential business, the same logic applies through your own professional body’s rules on competence, confidentiality and supervision. Check what your body has published and treat the SRA notice as a clear statement of the direction of travel.

The one-page AI policy

If staff are using AI on client work and nothing is written down, you already have an AI policy. It is just an unwritten one. The NCSC’s September 2026 guidance on shadow AI makes the same point from a security angle: bans are unlikely to work, so organisations should adopt realistic policies, find out why people turn to unapproved tools and give them secure alternatives.

A workable policy fits on one page and answers five questions:

  • Which AI tools are approved, and for what kinds of task?
  • What client or personal information must never be entered into them?
  • Who reviews AI-assisted work before it leaves the firm?
  • How is AI use recorded?
  • Who is told when something goes wrong, and how quickly?

The point is not paperwork. It is that a client, an insurer or a regulator who asks how you use AI gets a clear answer in minutes. Policy gaps are common: a 2025 TaxCalc survey of 215 accountants found half of firms had no formal policy on generative AI tools.

Data protection and confidentiality

UK GDPR still applies in full when personal data goes into an AI tool, and you remain responsible for it wherever it goes. For any tool or agent that touches client data, you need to know where the data is processed and stored, whether it is used to train a model, how long it is kept, which sub-processors are involved, and what the contract says. Where processing is likely to be high risk, a data protection impact assessment is the right tool.

The rules on automated decisions changed this year. Section 80 of the Data (Use and Access) Act 2025 came into force on 5 February 2026 and replaced Article 22 of the UK GDPR. The old regime was close to a prohibition on solely automated decisions with legal or similarly significant effects. The new one gives organisations more room to use automation, provided safeguards are in place, such as giving people information about the decision and a way to contest it, and ensuring meaningful human involvement where it matters.

The ICO published draft guidance on automated decision-making on 31 March 2026 and consulted on it. It covers when the rules apply, individuals’ rights, and the procurement of AI tools, including vendors that offer opaque, black-box decisions. It was still draft guidance at the time of writing, so check the ICO’s site for the final version.

For most small firms the practical rule is simple: keep client data to the minimum a task needs, use tools with contractual protections you have actually read, and do not let an AI make a significant decision about a person without a human who can change the outcome.

The EU AI Act: does it reach a UK firm?

The EU AI Act can reach organisations outside the EU, for example where the output of an AI system is used in the EU. If you advise EU clients or handle EU residents’ data, it is worth a conversation with your adviser. The timetable has also moved, so here is where it stands.

  • Since 2 February 2025: banned practices, and a duty on organisations to ensure staff have sufficient AI literacy.
  • Since 2 August 2025: obligations for providers of general-purpose AI models.
  • High-risk systems: under the EU’s Digital Omnibus on AI, stand-alone high-risk obligations (for example AI used in hiring, credit and education) were deferred from 2 August 2026 to 2 December 2027, and obligations for AI embedded in regulated products to 2 August 2028. Some reports say the amending regulation has been in force since 27 July 2026. Confirm the current position before relying on any date.
  • Transparency duties: most were left in place, although some watermarking obligations move later.

Most small UK firms using AI for admin, drafting and enquiry handling are not running high-risk systems in the Act’s sense. But the Act’s habits, which are classifying risk, keeping records and ensuring human oversight, are a sensible template whether or not it applies to you. This is general information, not legal advice.

ISO 42001: a useful framework, an optional badge

ISO/IEC 42001 is the first international management system standard for AI, and certification bodies describe it as the only AI framework that can currently be certified by an independent third party. It sets requirements for establishing, running and continually improving an AI management system. It applies to organisations that develop or use AI, and it is designed to sit alongside ISO 27001, the information security standard.

Certification typically involves a documentation review, an on-site assessment and surveillance audits. For a firm of 30 people or fewer, formal certification is rarely needed on day one. Our view is that the standard is most useful as a checklist: assign roles, assess risks, set controls, monitor, and review. Consider certification when clients or tenders begin to ask for it.

Human-in-the-loop oversight that works

“Human in the loop” only means something if you can say exactly where the human sits. We build agents around three levels of permission, and every action at every level is logged: what was done, from which data, who approved it, and when.

  • Level 1, act alone: low-risk, reversible tasks such as filing a document against the right record or sending a reminder whose wording you have approved.
  • Level 2, draft then ask: anything that goes to a client or changes a record. The agent prepares it and a named person approves it.
  • Level 3, never alone: anything involving judgement, advice, or money leaving the business. The agent can gather the facts, but a person decides.

That table of permissions is something you can read, change and show to a regulator or an insurer. It also fits the ICO’s emphasis on meaningful human involvement: a person who simply clicks approve on everything is not providing oversight. See our plain-English guide to AI agents for how this works inside a build.

Prompt injection and security

Prompt injection is when instructions hidden in content, such as an email, a document or a web page, trick an AI system into doing something its owner did not intend. The NCSC warned in December 2025 that it should not be treated like SQL injection. A language model has no inherent distinction between data and instructions, so prompt injection may never be fully mitigated. The NCSC’s advice is to reduce the likelihood and the impact, and to be wary of vendors who claim they can stop it altogether.

The NCSC’s September 2026 shadow AI guidance adds an important point for agents: an attacker who exploits one gains the same data, services and privileges that the agent legitimately has. That makes the design of the agent’s access the main defence.

Practical controls that follow from this:

  • Least privilege. Give each agent access only to the data, in the systems, for the job it does. Nothing more.
  • Separate reading from acting. Treat anything an agent reads from outside, including inbound email and uploaded documents, as untrusted.
  • Approvals on outbound actions. Anything that sends information out or changes a record needs a person, at least until the record shows it is safe to relax.
  • Logs and alerts. If you cannot see what the agent did, you cannot spot misuse.
  • Test with hostile inputs. Include documents and messages that try to redirect the agent.

The NCSC also observes that if an application cannot tolerate any residual risk, it may not be a suitable use case for a language model. That is a useful question to ask before you automate anything.

Professional indemnity insurance

Professional indemnity policies were mostly written before AI was in daily use, and the wording may say nothing about it. Commentary in 2025 and 2026 points to two concerns. First, broad cyber-related exclusions can be read to cover software-related losses, including AI. Second, PI cover generally responds when a client claims for negligent advice or a defective service from the firm, but may not respond where the loss stems from wrong information supplied directly to a customer by an AI system. Trade bodies have also been warning about so-called silent AI exposures across policies.

Whatever the policy says, the client will look to the firm they instructed, not to the software supplier. Sensible steps are:

  • Tell your broker which AI tools and agents you use and what they do.
  • Ask in writing whether any current or proposed exclusion affects them.
  • Check the wording at each renewal.
  • Keep your policy, permissions table and logs, so you can show the insurer how the AI is supervised.

This is general information, not insurance advice. Your broker is the right person to confirm your cover.

A 30-day governance starter plan

You do not need a programme to start. A realistic first month looks like this.

  • Week one: find out. Ask your team, without blame, what they have used AI for on client matters this month. Whatever comes back is your starting point.
  • Week two: write it down. Draft the one-page policy and an approved-tools list. Name one partner or director as owner.
  • Week three: set the controls. For any agent or workflow you plan to run, write the permissions table, decide what is logged, and agree who approves what.
  • Week four: close the gaps. Speak to your broker, run a data protection check on each supplier, and brief the team.

If you would like help working out where to start, the free AI Capacity Assessment is designed for exactly this.

Frequently asked questions

Is there an AI law in the UK?

There is no standalone AI Act as of mid-2026. Existing laws and your regulator’s rules apply, and regulators are using them. Watch for developments, but do not wait for new legislation before setting your own controls.

Do I need ISO 42001?

Not by law, and rarely at the start. It is a useful framework, and certification may become worthwhile if clients or tenders ask for it.

Can my staff use free tools like ChatGPT on client work?

Treat free public tools as unsuitable for client or confidential information unless your firm has approved a secure process. The SRA warns that both free and paid tools can carry risks depending on their terms and settings, and that privilege can be waived permanently.

Who is accountable if the AI gets it wrong?

The firm. The SRA says solicitors remain accountable for their work however it was prepared, and clients will look to the business they instructed.

Related reading

Find out where your firm stands

Law firms and accountancy practices can book a free 30-minute AI Capacity Assessment. Everyone else can take the free 8-question AI Readiness check. Both are no-obligation.

Book a free AI Capacity Assessment
Take the AI Readiness check

Sources

This guide is general information, not legal, regulatory or insurance advice. Rules and guidance change; check current positions with the original sources and your adviser.

More guides

Next steps

Related articles