Skip to main content

Resource: governance framework

AI Governance Framework for UK Regulated Firms

A practical AI governance framework for UK regulated firms: eight components, who owns each, and what the SRA, FCA, ICO and ICAEW have published.

By , Founder, The AI Consultancy (London) Ltd.

Published · Reading time approximately 16 minutes.

An AI governance framework is the set of roles, decisions, and records a firm uses to control how AI is chosen, approved, used, and watched. For a UK regulated firm of 20 to 500 people it can be small. This page sets out eight components, what the minimum version of each looks like, who owns it, and how it relates to what the Solicitors Regulation Authority (SRA), the Financial Conduct Authority (FCA), the Information Commissioner's Office (ICO), ICAEW, and the Law Society have published.

This framework is a starting point, not legal advice. Following it does not make a firm compliant with any rule, and nothing on this page claims endorsement, alignment, or accreditation from any of those bodies. Have the firm's own version reviewed by the people who carry its obligations: the compliance officer for legal practice (COLP), the Head of Compliance, the Data Protection Officer, or an external solicitor.

Last reviewed: 12 November 2026. The sources it relies on are listed at the end, and two of them were still changing on that date.

What is an AI governance framework, and does a UK regulated firm need one?

The Law Society's practice note "Using artificial intelligence in legal practice" (21 October 2026) gives a usable definition: AI governance is "the systems, policies, processes and oversight arrangements that manage how AI is used within an organisation".

None of the publications discussed below requires a firm to hold a document called an AI governance framework. The duty to govern AI comes from rules that already apply. The SRA's warning notice "Misuse of AI" (17 August 2026) says firms must have effective governance structures, systems and controls in place to manage risks, including those arising from the use of AI. The FCA's "AI in financial services: good practice and areas for improvement" (19 October 2026) opens with the heading "Existing rules and regulatory framework apply to firms' use of AI", and says firms "are responsible for AI systems they develop, procure and deploy". UK data protection law applies whenever personal data goes into an AI system. ICAEW's Code of Ethics asks members to judge whether reliance on the output of technology is reasonable.

A firm therefore already has AI governance obligations. A framework is how it meets them deliberately and can show that it has.

The Law Society's note also says "there is no single universal framework or standard for AI governance". It points to two recognised ones: the US National Institute of Standards and Technology's AI Risk Management Framework, and ISO/IEC 42001:2023, which treats AI governance as a management system. Neither is specific to the UK or to regulated professional firms. The framework on this page is narrower. It is written for a mid-sized UK professional or regulated firm that mostly buys AI tools and does not build them.

What are the eight components?

  1. Ownership: a named senior owner and a place where decisions are made.
  2. An AI register: every tool and every use, in one list.
  3. Risk tiers: a simple way to sort uses by how much harm they could do.
  4. Approval before use: an assessment and a sign-off sized to the tier.
  5. An acceptable use policy: the rules for staff.
  6. Human oversight: designed for each use, with a named reviewer.
  7. Training: by role, and kept current.
  8. Monitoring and review: incidents, checks on outcomes, and a fixed review date.

The FCA's publication says governance should cover the whole life of an AI system, which it lists as initial scoping, design and development, testing and validation, deployment, ongoing monitoring, and decommissioning. Components 2 to 4 cover the stages before a tool goes live. Components 5 to 8 cover the stages after.

1. Ownership

Someone senior is accountable for the firm's use of AI, and there is a known place where AI decisions are taken.

Minimum to have in place:

  • One named senior owner, recorded in the board or partnership minutes.
  • A decision forum. It can be an existing committee, such as the risk committee or the management board, with AI as a standing agenda item.
  • A statement of how much risk the firm will accept, in a paragraph or two. It should say which uses the firm wants and which it will not allow.

The FCA's publication describes "assigning clear SMF ownership for AI risks" as good practice, and says it does not expect senior managers to be technical specialists. It expects them to take reasonable steps to understand the risks in their area and to provide oversight and challenge. For smaller firms it describes "named accountable individuals, with clear decision ownership and sign-off". The Law Society's note says accountability for an organisation's approach to AI should sit with its leadership, who should set the overall risk appetite and allocate enough resource to governance, training, and oversight.

Owner: the board or the partners appoint; the named senior owner holds it.

2. An AI register

The firm cannot govern what it has not listed. The register is a single list of the AI tools in use and what each is used for.

Minimum to have in place, per entry:

  • The tool, the supplier, and the version or plan the firm pays for.
  • Each approved use.
  • The internal owner of the tool.
  • The kinds of information that go into it, and whether any of it is personal data or client confidential.
  • The risk tier (component 3), the date of approval, and the next review date.

Include AI features inside software the firm already licenses. The Law Society's note observes that where AI is embedded in existing software a firm may not have clear visibility of where it operates or what data it can reach, and suggests asking suppliers for regular updates. The note describes an AI register as recording approved systems, approved use cases, the owner of each system, and dates of approval and review. The FCA's publication describes firms keeping inventories of AI use cases and the models they are built on, and, for more autonomous systems, registering AI agents in a central inventory showing who owns them and what systems and data they can access.

Start the register by asking staff what they already use, including tools the firm did not buy.

Owner: the named senior owner, with the IT or security lead keeping it up to date.

3. Risk tiers

Not every use of AI needs the same scrutiny. Tiers let the firm spend its attention where the harm would be greatest.

We suggest three tiers for a firm of this size:

  • Lower: internal drafting, summarising, and research support with no client confidential information and no personal data, where a person checks the result before it is used.
  • Medium: uses that process client confidential information or personal data in an approved tool, or that produce work sent to clients after review.
  • Higher: uses that influence a decision about a person, produce documents for a court or a regulator, affect customers directly, or act with a degree of autonomy.

Two questions sort most uses: what goes in, and what happens to the output. The FCA's publication describes smaller firms using "simpler risk tiers" inside their existing governance to increase testing, guardrails, and monitoring for higher-risk uses, and escalating only the higher risks.

Owner: the decision forum sets the scheme; the named senior owner assigns the tier for each use.

4. Approval before use

No tool is used on the firm's work, and no approved tool is put to a new use, until someone with authority has assessed it and signed it off.

Minimum to have in place:

  • A one-page request: the purpose, the information involved, the supplier, and the proposed tier.
  • For medium and higher tiers, supplier due diligence: where data is processed, whether it is used for training, how long it is kept, who the sub-processors are, and what notice the supplier gives of changes to its models.
  • A decision on whether a data protection impact assessment is needed, taken by the Data Protection Officer or data protection lead.
  • For the higher tier, sign-off by the decision forum, with the reasons recorded.

The SRA's notice says client information should be entered into AI systems only where appropriate contractual, technical and organisational safeguards are in place, which a firm cannot know without asking the supplier. The FCA's publication lists "retrospective governance" as an area for improvement: firms that deployed AI and then tried to fit it into existing frameworks found it harder to show who was accountable. It also warns against overreliance on untested vendor assurances, and says "firms can outsource AI services, but remain responsible for them". The ICO's guidance says you should carry out a data protection impact assessment for any solely automated decision with a significant effect on a person.

Our article on AI procurement due diligence sets out the supplier questions in full.

Owner: the named senior owner approves lower and medium tiers; the decision forum approves the higher tier; the Data Protection Officer decides on the impact assessment.

5. An acceptable use policy

The policy turns the firm's decisions into rules staff can follow: which tools, what may be entered, how output is checked, and who to ask.

Minimum to have in place: a policy of two or three pages, issued to everyone, with the list of approved tools attached. Our AI acceptable use policy template gives fourteen clauses to copy and adapt.

The Law Society's note says a firm needs to communicate what is acceptable even where it has not formally adopted AI. The template page sets out why.

Owner: the named senior owner; the board or the partners approve it.

6. Human oversight

For each use above the lower tier, the firm decides in advance who reviews the output, what they check, and what authority they have to change or stop it.

Minimum to have in place:

  • A named reviewer or reviewing role for each medium and higher tier use.
  • A written description of what the review covers. "Check the output" is too vague to supervise.
  • A record, on the file, that the review happened and what the reviewer did.
  • A route for the reviewer to escalate, and a threshold for when they must.

The FCA's publication says effective oversight "is not simply about reviewing AI outputs, but about ensuring people have the right authority, information, capability and escalation routes to challenge or intervene where needed". It describes human review becoming a "box-ticking" exercise as an area for improvement. The SRA's notice says appropriate human oversight, informed professional judgement, and a proportionate, risk-based approach are essential, and that supervisors remain accountable for work carried out through others.

Where a use produces a significant decision about a person, data protection law adds a specific test. The UK GDPR's automated decision-making rules (Articles 22A to 22D, which replaced Article 22 under the Data (Use and Access) Act 2025) apply where such a decision is taken with no meaningful human involvement. The ICO's guidance says a decision that includes meaningful human involvement is not solely automated. Where the rules do apply, the firm must provide safeguards: information about the decision, and the ability to make representations, obtain human intervention, and contest it.

The guidance says spot checking does not meet the human intervention safeguard. It was still a draft under consultation in November 2026, so check the current version.

Owner: the head of each practice area or function designs the review for its uses; supervisors carry it out.

7. Training

People are trained for the role they play, before they are given access, and again when the tools or the rules change.

Minimum to have in place:

  • Baseline training for every user: the policy, what must not be entered, and how to check output.
  • Additional training for supervisors and reviewers on how to review AI-supported work and when to override it.
  • A briefing for the leadership on the firm's obligations, its risk appetite, and the decisions they will be asked to make.
  • A record of who has completed what.

The SRA's Codes require firms to ensure their managers and employees are competent and keep their knowledge up to date. The FCA's publication describes role-specific training as good practice and names three groups: senior stakeholders, technical teams, and human-in-the-loop reviewers. It lists training that is "generic or limited to awareness" in oversight functions as an area for improvement. ICAEW's Code says maintaining professional competence requires "a continuing awareness and understanding of technical, professional, business and technology-related developments relevant to the professional activities undertaken".

Owner: HR or the learning lead keeps the record; the named senior owner decides what each group needs.

8. Monitoring and review

Approval does not end the firm's responsibility for a tool. The firm keeps checking that each tool still does what it was approved to do, hears about problems quickly, and revisits the framework on a fixed date.

Minimum to have in place:

  • An incident route: who staff tell, and who decides whether a client, an insurer, or a regulator must be told.
  • A periodic check on each medium and higher tier use. The Law Society's note suggests reviewing which systems are in use, who is using them, what information is uploaded, recurring errors, compliance incidents, and client complaints.
  • A watch on supplier changes: new features, new terms, and changes to the underlying model.
  • A review of the whole framework every six months, and whenever a regulator publishes something new.
  • A way to retire a tool: access removed, data deleted or returned, the register updated.

The FCA's publication lists treating governance "as a point-in-time exercise" as an area for improvement, and describes firms that monitored technical measures and not outcomes as having less chance of meaningful oversight.

Owner: the named senior owner reports to the decision forum; the IT or security lead watches supplier changes.

Who owns what?

  • The board or the partners appoint the senior owner, set the risk appetite, approve the policy, and receive a report at least twice a year.
  • The named senior owner runs the framework: the register, tiering, approvals for lower and medium tier uses, and the review cycle. In an FCA-regulated firm this person should be, or report to, the senior manager responsible for the relevant business area. In a law firm the COLP should be involved by name, because the SRA's notice says the COLP must take all reasonable steps to ensure compliance with the firm's regulatory arrangements.
  • The decision forum approves higher tier uses and hears escalations.
  • The Data Protection Officer or data protection lead decides on impact assessments and on any use that touches the automated decision-making rules.
  • The IT or security lead checks tool settings, access, and retention, and keeps the register current.
  • Heads of practice areas and functions own the uses in their area and design the human review for each.
  • Supervisors carry out the review and are accountable for the work that goes out.
  • Every member of staff follows the policy and reports incidents.

In a smaller firm one person may hold several of these roles. The FCA's publication says smaller firms may have simpler governance structures, "but they still need to have clear responsibilities, understand the risks, and make informed decisions".

How does the framework relate to what each regulator has published?

For SRA-regulated firms, the warning notice is the document to read against components 1, 4, 6, and 7. It deals with governance and controls, confidentiality safeguards for client information, supervision, and competence. The SRA says it will have regard to the notice when exercising its regulatory functions.

For FCA-regulated firms, the good practice publication speaks to most of the components, and to ownership (component 1) and monitoring (component 8) in most detail. Its examples are illustrative and, in the FCA's words, "do not create new requirements". The requirements sit in the existing rules it lists, including the Principles for Businesses, the Consumer Duty, the Senior Managers and Certification Regime, and the operational resilience and outsourcing rules.

For any firm that handles personal data, the ICO's guidance bears on components 4 and 6. The ICO has said its guidance on automated decision-making will inform a statutory code of practice on AI and automated decision-making that it will develop.

For ICAEW members, the Code of Ethics provisions on technology bear on components 4, 6, and 7. The helpsheet lists the factors a member weighs before relying on the output of technology: the firm's oversight of the technology, the controls over its use, and the appropriateness of the inputs and of the decisions made by the people using it. ICAEW says the helpsheet is no substitute for reading the Code in full.

Where a firm answers to more than one of these bodies, we suggest one framework, applying the stricter requirement wherever two differ.

What does a proportionate start look like?

A firm with nothing in place does not need to build all eight components at once. Work in this order:

  1. Appoint the senior owner and choose the decision forum.
  2. Build the first register by asking every team what it uses.
  3. Stop the uses that are plainly unsafe, such as client information in personal accounts.
  4. Agree the three tiers and assign one to each entry in the register.
  5. Issue the acceptable use policy with the list of approved tools.
  6. Train everyone on the policy, and train supervisors on review.
  7. Put the first six-month review in the diary.

The FCA's publication describes smaller firms starting with narrower uses and taking an incremental, risk-based approach, which let them test governance and build capability before scaling. We suggest the same for a professional firm: govern the uses that already exist before approving new ones.

Learn AI's AI governance briefing for regulated firms works through these components with a firm's compliance, legal, and data protection leads in a half-day session. The free AI Readiness Assessment shows where a firm stands today.

What does this framework not cover?

It is written for UK obligations. A firm that offers AI-enabled services in the European Union, or serves clients there, may also have duties under the EU AI Act, which this page does not address.

It assumes the firm buys AI tools and does not build them. A firm that develops or trains its own models needs additional controls on data, testing, and model risk.

It does not replace the sector rules themselves. The sector articles on law firms, financial services, finance teams, and hiring cover those in more detail.

Sources

Each source was read at the publisher's own site on 12 November 2026.