The regulatory map every firm now has to read
A partner at a mid-size accounting firm uses an AI tool to draft a client's transfer pricingtransfer pricingThe prices charged when units of the same company trade goods, services or IP across borders, set to satisfy tax authorities and reflect market value.View full definition → memo. The tool is built on OpenAI's GPT models and hosted on Microsoft infrastructure. The client is incorporated in Ireland with operations in California, and a UK-qualified associate reviews the memo. Five different rulebooks may apply to that one memo, and none of them agree on what "AI use" even means. This is not an edge case in 2026. It is Tuesday.
Professional services firms (law, audit, consulting, tax, financial advisory) now sit where four regulatory logics overlap: horizontal AI law, sector-specific financial and legal regulation, professional conduct rules, and private contract law. A firm that reads only one of these maps will get lost.
Layer 1: horizontal AI regulation (the EU AI Act)
The EU AI Act (Regulation (EU) 2024/1689, official text here) is the first comprehensive, cross-sector AI law. It classifies AI systems by risk tier, not by industry:
- Unacceptable risk: banned outright (e.g. social scoring). These bans have applied since February 2025.
- High-risk: heavy obligations, including risk management, human oversight, logging and conformity assessment. This tier covers AI used in employment decisions, credit scoring and some biometric uses. Most day-to-day generative AI drafting tools used by lawyers or consultants do not automatically fall here.
- Limited risk: transparency duties only (e.g. disclose that a chatbot is AI).
- Minimal risk: no specific obligation.
The timeline matters. Obligations for general-purpose AI (GPAI) models have applied since August 2025. High-risk obligations were originally due in August 2026. In November 2025 the European Commission proposed a "Digital Omnibus" package that would delay them, tying the start date to the availability of harmonised standards, with backstop dates in December 2027 and August 2028. Check the current legislative status before telling a client a deadline.
For professional services, the sharpest edge is indirect. Suppose a client uses your firm's AI-assisted output to make a high-risk decision, such as an AI-scored creditworthiness assessment feeding a bank's lending decision. The classification travels with the use case, not the vendor. A consulting firm that builds that scoring tool can become a "provider" under the Act with full high-risk obligations, even if it thinks of itself as a service firm and not a software company.
GPAI models such as GPT-5, Claude or Gemini carry their own transparency and systemic-risk obligations under the Act. Those sit on the model providers, on top of whatever your firm does with the models.
Layer 2: Sector regulators don't wait for horizontal law
Financial and legal regulators were writing AI-adjacent rules before the AI Act existed, and they keep moving independently.
- The SEC (U.S. Securities and Exchange Commission) has brought "AI washing" enforcement actions against firms that overstated their AI capabilities to investors. Its 2023 proposal on conflicts of interest from predictive data analytics at broker-dealers and advisers was withdrawn in June 2025. Existing antifraud and fiduciary rules still apply to AI-driven advice.
- The UK Financial Conduct Authority applies existing frameworks to AI use, especially the Consumer Duty and the Senior Managers and Certification Regime (SMCR). The UK has no separate "AI Act". The FCA's position (see FCA's AI pages) is that principles-based rules already cover most AI harms. Accountability sits with a named senior manager, whichever model made the recommendation.
- Bar associations and law societies have issued their own guidance. The American Bar Association's Formal Opinion 512 (2024) and various state bar opinions address generative AI competence, confidentiality when client data goes into third-party tools, and billing (firms cannot bill client hours saved by AI as if nothing changed).
- Audit oversight bodies, including the PCAOB (US Public Company Accounting Oversight Board) and European equivalents, are scrutinizing AI in audit sampling and analytics. Their focus is whether AI tools undermine auditor independence or professional skepticism.
None of these bodies use the EU AI Act's risk tiers. A tool that is "minimal risk" under the Act can still trigger an SEC enforcement action if the marketing claims about it are false. The same tool can draw a bar complaint if it leaks privileged client data.
Layer 3: Professional conduct rules travel with the license, not the tool
Non-technical readers often miss this layer. Professional obligations (competence, confidentiality, independence, professional skepticism) attach to the licensed individual, not the software vendor.
Take a lawyer whose AI research tool hallucinates a fake case citation. That lawyer is still personally sanctionable, as multiple US court sanctions since 2023 show (*Mata v. Avianca* remains the reference point, and courts have issued hundreds of similar rulings since). The AI Act's conformity assessment does nothing to protect that lawyer. Professional discipline runs on a separate, parallel track.
For consultants and auditors, the equivalent risk is over-reliance. They use AI-generated analysis and skip the professional judgment step that the client is paying for and that regulators expect.
Layer 4: Client contracts add a fourth, private layer of law
Client engagement letters and master service agreements increasingly specify:
- Which AI tools may or may not touch client data.
- Whether client data can be used to train or fine-tune models. Most sophisticated clients now explicitly say no.
- Disclosure obligations, such as whether the firm must tell the client when AI materially contributed to a deliverable.
- Liability allocation if AI-assisted work causes loss.
These clauses are not harmonized across jurisdictions, or even across clients of the same firm. A firm serving 200 clients can face 200 slightly different private AI policies on top of public law. This is why many firms now keep an internal AI use register: a log of which tools are approved, for which task types, under which client contract terms.
Example: minimal AI use register schema
tool_name, vendor, data_residency, client_data_allowed (Y/N),
approved_use_cases[], risk_tier_eu_ai_act, accountable_person,
human_review_required (Y/N)This register connects all four layers in practice. It is the document a regulator, a bar investigator or a client auditor will ask to see.
Knowledge check
1. Under the EU AI Act, how is the risk tier of an AI system primarily determined?
2. A consulting firm builds a scoring tool that a bank later uses to make lending decisions. What does this scenario illustrate about the AI Act's high-risk classification?
3. Why does the opening example of the transfer pricing memo (multiple jurisdictions, vendors, and reviewers) matter conceptually for professional services firms?
4. Select ALL correct answers about the EU AI Act's risk-tier system as described in the lesson.
Select all the correct answers.
5. Select ALL correct answers about why professional services firms face a complex regulatory landscape for AI use.
Select all the correct answers.
What this means for cross-border advisory work
A firm advising a client with EU, US and UK touchpoints cannot pick "the strictest rule" and apply it everywhere. The rules ask different kinds of questions, so they cannot be ranked from loosest to strictest:
- The EU AI Act asks what risk tier the system falls into.
- The SEC asks whether you misrepresented it.
- The FCA asks which senior manager is accountable.
- The bar asks whether you maintained competence and confidentiality.
- The contract asks whether you disclosed AI use and got consent.
A practical governance response, in order:
- MapMapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → the use case, not the tool. The same GPT-based drafting assistant can be low-risk in one workflow (internal summarization) and high-risk in another (feeding a client-facing credit decision).
- Assign a named accountable person to each AI use case, following the FCA's SMCR logic even outside UK-regulated entities. Diffuse accountability is the most common finding in AI governance failures.
- Build a human review checkpoint wherever professional judgment is the product being sold. Conduct rules require it, and so does client trust.
- Reconcile contract clauses with the AI use register before deployment, not after a client complaint.
- Document reasoning, not just output. Auditors and regulators increasingly ask why a conclusion was reached, not only what the AI produced.
- Track the dates. EU AI Act deadlines are still moving, and the SEC's agenda changed in 2025. Review the register at least quarterly.
The EU AI Act Explained
The OECD AI Policy Observatory tracks national AI regulatory developments close to real time. Use it to check how regulators are coordinating, or failing to.
Key Takeaways
- Four distinct regulatory layers apply to AI use in professional services: horizontal law (EU AI Act risk tiers), sector regulators (SEC, FCA, PCAOB), professional conduct rules (bar associations, audit standards) and private client contracts. They classify the same use case differently and do not defer to one another.
- EU AI Act timing is in flux. GPAI obligations have applied since August 2025, and the Commission has proposed delaying high-risk obligations beyond August 2026. The SEC withdrew its predictive analytics proposal in 2025, but AI washing enforcement continues.
- Risk classification follows the use case, not the vendor or tool. The same model can be minimal-risk in one workflow and high-risk, or professionally sanctionable, in another.
- Professional discipline is a personal, license-level risk. It exists whether or not the firm complies with the EU AI Act.
- An AI use register that maps tool, use case, data handling and accountable person lets a firm answer to all four layers at once. Reconcile the regimes rather than looking for one "strictest rule".