Using AI Safely in Regulated Professions

Professional regulation exists because clients, courts, markets and the public place trust in the people and organisations whose work can directly affect rights, money, freedom, property, employment and health. AI does not inherit that trust. It does not accept a professional duty. It does not carry professional indemnity insurance. It cannot appear before a court to explain itself.

That does not mean regulated professionals should avoid AI. It means they should use it in a way that keeps professional responsibility where it legally and professionally belongs — with the qualified person who has accepted it.

This guide uses solicitors in England and Wales as the clearest detailed example, because their regulatory framework, court obligations and professional duties are well documented and publicly stated. Similar principles apply, in ways specific to each sector's rules, to accountants, financial advisers, insurance professionals, surveyors, architects, HR professionals, healthcare organisations, regulated consultants, public bodies and compliance teams. This guide does not apply any profession's rules identically to another.

Solicitors in England and Wales — current position

Solicitors in England and Wales are not generally prohibited from using generative AI or other legal technology. The Solicitors Regulation Authority does not ban the use of particular AI tools by name. What the SRA's regulatory framework requires is that solicitors meet professional obligations regardless of the tools they use.

Current SRA regulatory expectations relevant to AI include: maintaining competence, including understanding the technology used; protecting client confidentiality; maintaining legal professional privilege where it applies; complying with data protection law; ensuring that suppliers used to assist with client work are appropriate and that risks are managed; supervising work carried out under the firm's authority; providing accurate and non-misleading advice and court material; considering equality and bias where AI is used in decisions; keeping appropriate records; and maintaining adequate insurance.

The Law Society has issued practice notes and guidance on AI, covering areas including legal professional privilege, confidentiality and human oversight. The position is developing as AI capabilities, case law and regulatory guidance evolve. Verify the current SRA and Law Society positions using their official sources before acting.

A solicitor remains responsible for work performed in their name, including work assisted by junior staff, contractors, outsourced providers, software or AI systems. Delegation does not transfer professional responsibility.

Other regulated professions

Similar principles apply across regulated sectors, though the specific rules, regulators and requirements differ. Accountants and auditors operate under ICAEW, ACCA and FRC standards and must maintain professional scepticism, audit quality and client confidentiality. Financial advisers and wealth managers operate under FCA rules requiring advice to be suitable, accurate and demonstrably in the client's interest. Insurance professionals have duties of accuracy and fair dealing. Surveyors, architects and engineers carry professional responsibility for advice, calculations and signed documents. HR professionals and employers face obligations around fairness, discrimination and employment law. Healthcare organisations are subject to clinical governance, patient confidentiality, CQC requirements and NHS data standards.

In each case, the question is not simply whether AI may be used. It is whether AI is being used in a way that is consistent with the specific obligations of the regulated sector, the professional's duty to clients or patients, data protection law and the standards the regulator enforces.

Court documents and AI

The judiciary in England and Wales has issued guidance for legal professionals using AI for court work. The Courts and Tribunals Judiciary issued interim guidance in 2023 and has consulted further on AI in the courts. Verify the current status of any guidance, consultation or formal procedural rule against official judiciary sources before acting.

It is important to distinguish between the categories of material that may affect practice in this area:

• Current professional duty: the existing responsibility to provide accurate, non-misleading material to the court — which applies regardless of whether AI was used
• Guidance: advice issued by regulators or the judiciary about how duties apply to AI-assisted work
• Consultation: a proposal being considered — not yet a binding requirement
• Formal rule: a binding procedural or regulatory requirement once adopted

Cases in multiple jurisdictions have resulted in professional consequences where AI-generated legal citations that did not exist were submitted to courts without verification. In England and Wales, the existing duty of candour to the court applies fully to AI-assisted work. Submitting fabricated or unverified material to a court, regardless of how it was generated, is a serious professional matter. Check current SRA and judiciary sources for any specific guidance or requirements that have been adopted since this guide was last reviewed.

Legal research and hallucinated authorities

Generative AI systems may produce plausible but entirely invented legal citations. This is not a rare edge case — it has been observed across multiple platforms and has resulted in disciplinary and court proceedings when fabricated citations were submitted without verification.

AI may invent cases, citations, judges, quotations, legislation, procedural rules, dates, links and legal propositions. Each may appear entirely credible. The output may contain accurate references alongside invented ones, making selective checking insufficient.

For every legal authority used in professional work, verify through an authoritative legal database or primary source: that the case exists; that the neutral citation is correct; that the court is correctly identified; that any quoted passage appears in the actual judgment; that the legal proposition it supports is genuinely what the case decides; that the decision has not been overruled, distinguished or reversed; that subsequent treatment has been considered; and that the jurisdiction is relevant to the matter.

AI-generated legal summaries are a starting point for research, not a substitute for it. Professional reliance on AI output without independent verification is a professional responsibility question, not a technology question.

Confidentiality and privilege

Professional information may include client identity, case facts, legal advice, litigation strategy, privileged communications, medical information, financial information, employee records, commercial negotiations, witness statements, settlement positions and security information. The appropriate AI tool, account type and contractual arrangements depend on the specific nature of that information.

Consumer AI tools operate under general consumer terms with limited organisational control. Business or enterprise AI services are contracted services with additional administration, security and data commitments that vary by supplier and agreement. Private or controlled deployments operate within a defined technical and contractual environment. These are meaningfully different categories — the question is not simply which product name appears on the login screen.

Enterprise branding alone does not make every use lawful or professionally appropriate. The relevant questions are: what data enters the system; under what contractual terms; with what retention, training and subprocessor arrangements; and is the organisation authorised to disclose and process that information through that service given its professional obligations.

Personal data

UK data protection law applies to personal data — information relating to identifiable individuals — which includes most client, patient, counterparty and employee data that professional-services organisations handle. Entering personal data into an AI service is a disclosure and a processing activity. It requires the same legal analysis as any other disclosure.

Data protection obligations relevant to AI use include: identifying the correct controller and processor roles for each arrangement; establishing a lawful basis for the processing; considering transparency obligations to data subjects; applying data minimisation — providing only the information the AI task actually requires; ensuring appropriate security; understanding data retention and deletion arrangements; considering international transfers where data is processed outside the UK; addressing individual rights including subject access; assessing special-category data where relevant; and conducting data protection impact assessments where the processing involves high risk.

Consent is not automatically required — other lawful bases may apply depending on the use case. But the correct approach depends on the use case, the data, the people affected, the platform, the contract, the risk, the available safeguards and any applicable professional requirements. Removing a name does not automatically anonymise a document. The ICO has published detailed guidance on AI and data protection that should be reviewed when assessing any AI deployment involving personal data.

Professional privilege

Legal professional privilege protects confidential communications between a lawyer and client for the purpose of legal advice, and communications prepared for the dominant purpose of litigation. Privilege belongs to the client. Its protection and potential loss are matters of legal complexity.

Entering privileged or confidential information into a third-party AI service may create confidentiality risk, contractual risk, privilege concerns, disclosure risk, access by subprocessors, retention issues, data-location issues and evidential complications. This guide does not give a universal conclusion that privilege is automatically lost by using any particular service — the answer depends on specific facts, service terms and applicable case law.

Where sensitive or privileged information is involved, obtain appropriate specialist advice before proceeding. The Law Society and specialist commentators have published analysis on AI and privilege that should be consulted and verified against current authority.

First-draft use and the safe drafting sequence

AI may appropriately assist with document outlines, issue lists, chronology preparation, first-draft correspondence, document comparison, approved document summaries, internal checklists, question preparation, non-confidential research planning and formatting. These tasks involve preparation and organisation rather than the application of professional judgement.

AI should not be treated as a substitute for legal analysis, professional advice, evidential judgement, strategic decisions, final court drafting, client-specific recommendations, conflict checks or ethical decisions.

A safer drafting sequence: define the permitted task and the approved information to be used; ask for an outline and review it before proceeding; require uncertainty to be identified explicitly; produce the first draft from approved information; verify every source and authority against primary sources; apply professional judgement; obtain final human approval before the document leaves the firm.

Deadlines and calculations

AI may assist with extracting dates from documents, identifying relevant events, preparing a draft chronology, proposing reminders and highlighting missing information. These are useful preparation tasks.

Critical deadline calculations may depend on court rules, working days, bank holidays, service methods, time zones, jurisdiction, extensions, contractual wording, event timing, exceptions and rule changes. Each of these can affect whether a deadline is correct. A draft date produced by an AI system is a suggestion, not a verified calculation.

Independent verification through authoritative rules — court practice directions, limitation statutes, contractual provisions — and approved case management systems is required for every critical date. Professional diary management systems exist precisely because the consequences of a missed deadline can be catastrophic for clients and serious for the professional.

Prompt injection and hostile documents

Documents and external material may contain instructions intended to influence an AI system — a risk described as prompt injection. Instructions can be concealed in white text within a document, in comments or metadata, in hidden spreadsheet cells, in PDFs, in email content, in websites, in document management systems or in embedded content.

Prompt injection in a professional context could cause an AI system to produce misleading output, take unexpected actions, reveal information from other parts of the context, generate content serving the interests of an adversary rather than the professional's client, or trigger tool use without human authorisation.

Controls include: treating documents as untrusted data and separating source content from system instructions; limiting the tools available to the AI to those actually needed for the task; reviewing all output rather than accepting it automatically; using secure document viewers and content inspection where appropriate; prohibiting automatic external action based solely on document content; logging AI interactions with external documents; and testing workflows against adversarial document examples.

Automated external access

AI or browser automation may be used to access court portals, regulator portals, client systems, land registries, financial platforms, government services and document portals. Each carries obligations beyond technical capability.

Before automating access to any external system, consider: whether the organisation holds authorised access; what the system's terms of service permit; whether an official API exists and should be used instead; who owns the credentials used; what audit logs are maintained; what rate limits apply; how service availability affects workflow reliability; what data protection obligations apply to data collected; what professional obligations apply; and what happens if the automation fails.

Client communication and transparency

Whether and when to disclose AI use depends on professional rules, contractual commitments, client instructions, the materiality of the AI's contribution, data handling, service descriptions, automated decisions affecting clients and sector-specific requirements.

Situations where transparency may be particularly important include where AI materially shapes advice; where personal data enters an external service; where automated decisions affect a person; where the engagement letter restricts outsourcing or data sharing; where client instructions expressly prohibit AI use; where a regulator requires disclosure; or where the service description would otherwise be misleading.

Supplier due diligence

A professional AI supplier assessment should address: the legal entity providing the service; the specific product purpose and what the organisation intends to use it for; the account type and available organisational controls; the contract and its key terms; data ownership and what happens to inputs and outputs; whether inputs are used to train models; retention and deletion of data; the data locations used; subprocessor arrangements; encryption in transit and at rest; access controls and authentication requirements including MFA and single sign-on; audit logging capability; administrative and user-management controls; incident notification obligations and timescales; service availability and SLA arrangements; data export and portability; exit arrangements; any professional-sector-specific claims; stated accuracy limitations; model-change notification practices; insurance and liability limits; and the level of support provided.

Human review levels

Not every AI-assisted task requires the same level of professional review. A proportionate model for regulated work:

Level 1 — Low-risk administration: formatting, meeting notes, non-sensitive categorisation. Routine review is appropriate. The work does not leave the firm and does not create professional records.

Level 2 — Internal professional support: issue identification, chronology drafts, internal summaries, research plans. Qualified review before any reliance is placed on the output.

Level 3 — Client-facing professional work: advice drafts, reports, client correspondence, calculations. Detailed professional review and approval before the output is sent to the client.

Level 4 — Court, regulatory or rights-affecting work: court documents, regulatory submissions, expert reports, financial recommendations, disciplinary decisions, regulated advice. Explicit qualified approval and independent verification of critical facts, law, sources and calculations.

Audit trail

Regulated organisations may need records demonstrating that AI-assisted work was properly controlled. Records may need to show: the approved tool used; the user; the purpose; the source material; the model or service version; the instructions provided; the output; corrections made; which sources were checked; who carried out the review; who gave approval; the date; the final version; and any incident or exception.

Proportionality applies. Records should not retain unnecessary personal data merely to create an audit trail — retention periods and data minimisation requirements apply to records as they do to other processing. The obligation is to be able to demonstrate appropriate professional accountability, not to create comprehensive surveillance of every AI interaction.

Staff competence

Professional competence now includes understanding the AI tools in use: what they can and cannot do, where they are known to fail, what information may enter them, how confidentiality and privilege apply, what security controls are required, how data protection obligations apply, where bias may arise, how to verify output, when to escalate a concern, which tools are approved and which uses are prohibited.

Effective training for regulated work should include realistic exercises rather than theoretical descriptions. Exercises should involve invented legal sources that look plausible, missing information that should be noticed, confidential documents that should not enter the tool, prompt injection examples, proposed deadlines that require independent checking, ambiguous instructions, and persuasive but incorrect output. The objective is professional judgement about when not to trust the AI, not just technical skill in using it.

AI use policy for regulated work

A professional firm's AI policy should address: which tools are approved and for which uses; which account types are permitted; what information is prohibited from any AI tool; what uses are permitted; which uses are high-risk and require senior review; what level of human review is required for different task types; how sources must be verified; how client instructions about AI use are recorded and followed; what approval is required before court or regulator submissions; which AI supplier arrangements have been approved; what browser automation is permitted; what records must be kept; how incidents are reported; what training is required; what disciplinary or contractor consequences apply; and who owns policy review.

Formal policies for regulated organisations should receive appropriate legal and regulatory review before adoption. This guide does not provide employment-law or contractual wording.

Regulated AI use checklist

Before using AI for regulated or professional work, work through these questions:

Purpose: Is the task suitable for AI assistance? Is an accountable professional assigned? Is the risk level defined?

Tool: Is the tool approved? Is the account appropriate for this information? Have supplier terms been reviewed? Are security controls — MFA, SSO, audit logging — enabled?

Information: Is personal data involved? Is confidential or privileged information involved? Can the information be minimised? Are the client's restrictions on data sharing known?

Output: Could it affect rights, money, liberty or legal obligations? Are facts independently checked? Are all legal authorities opened and verified against primary sources? Are calculations checked separately?

Action: Will the output be sent externally? Will it be filed with a court or regulator? Will it trigger an automated action? Is explicit approval by a qualified person required?

Record: Is the final version retained? Is the reviewer recorded? Are the sources checked recorded? Are incidents and exceptions documented?

Warning signs

Stop or escalate where any of these apply:

• Public consumer tools receive client files
• The user cannot explain the supplier terms
• AI-generated legal authorities are not checked against primary sources
• A court or limitation deadline relies solely on AI output
• A document may contain hidden instructions
• A workflow can file or send without human approval
• Client restrictions on AI use are ignored
• The output affects a person automatically without human review
• Professional privilege has not been considered before entering information
• Personal data is excessive given the task
• AI output is treated as legal research evidence rather than a starting point
• Audit records do not exist for significant AI-assisted work
• Staff use personal AI accounts for professional work
• The business has no approved-tool list
• Senior professionals delegate review entirely to junior staff
• Polished wording is mistaken for professional competence
• The organisation cannot withdraw AI access quickly if needed

Practical business implications

• AI use is not generally prohibited — but professional duties still apply fully
• The professional remains responsible — software does not sign away accountability
• Confidentiality requires platform assessment — consumer and controlled business systems are not equivalent
• Legal authorities must be opened — a citation is not verified because it looks convincing
• Deadlines require independent checking — calendar errors can cause serious client harm
• Prompt injection is a document risk — files may contain hostile instructions for AI systems
• Automation needs authorisation — external systems and portals have access rules
• Client expectations matter — AI use should not make service descriptions misleading
• Audit trails support accountability — regulated work should remain explainable
• Human review must be meaningful — proofreading is not enough where judgement is required

The IT Club view

Regulated professionals should not reject useful AI merely because it creates risk. They should also not treat professional status as protection from poor technology decisions.

AI can assist with preparation, extraction, organisation, first drafts, comparison and controlled research support. It cannot accept professional responsibility.

IT Club recommends: using only approved tools for client-related work; establishing clear data boundaries that reflect confidentiality and privilege; requiring qualified human review proportionate to the consequence of error; verifying every legal authority and calculation independently; controlling automation and requiring approval before external action; assessing AI suppliers as carefully as any other professional service provider; training staff to recognise hallucinations, prompt injection and when to stop; maintaining audit trails that support professional accountability; reporting incidents and learning from them; and reviewing AI use through an Operational Heartbeat as regulatory guidance develops.

The right goal is not to keep AI out of professional work. It is to make every material use controlled, reviewable and accountable.