EU AI Act for Professional Services: What Actually Applies

Probably not high-risk. A consultancy drafting memos, an accounting firm summarising audit files, an agency generating campaign copy — none touches the high-risk tier of the EU AI Act, and treating it as if it does wastes budget on obligations that never apply.

Regulation (EU) 2024/1689 classifies AI by use case, not by sector. Most professional-services work sits at limited or minimal risk, so the real exposures come from four narrow places — the AI-literacy duty, confidentiality over client data, transparency when output reaches clients, and the trap where a firm that rebrands and resells an AI tool becomes a provider. This page covers those four, plus the one genuine high-risk line: candidate screening on the recruiting side.

---

Why most professional-services AI is not high-risk

High-risk status is defined by what a system does, not by who deploys it. Under Article 6, a system is high-risk only in two cases: it is a safety component of a product covered by Annex I (machinery, medical devices, vehicles), or it falls within one of the eight areas in Annex III. Nothing in the Regulation creates a "professional-services" category.

Run the ordinary advisory uses against the eight Annex III areas and they map to none of them. Drafting memos, summarising documents, desk research, checking a financial model, generating slides, marketing copy and internal knowledge search are none of biometrics, critical infrastructure, education, employment decisions about your own staff, essential services such as creditworthiness scoring, law enforcement, migration, or justice. All sit at limited or minimal risk.

The consequence is large. Because the classification does not attach, neither does the high-risk obligation stack in Articles 8 to 15 — risk management, data governance, technical documentation, human oversight, accuracy and robustness. There is no conformity assessment, no EU database registration under Article 49, and no Annex IV technical file for everyday AI use. That stack is built for providers of systems that screen job candidates or score loan applications, not for a consultant running a research query. An advisory firm's obligations instead come from four places, none depending on a high-risk label: Article 4 literacy, confidentiality over client data, Article 50 transparency, and the Article 25 provider trap.

---

Article 4: the AI-literacy duty that already applies to every firm

Article 4 requires providers and deployers to ensure a sufficient level of AI literacy among their staff and other persons operating AI systems on their behalf, proportionate to those people's knowledge, experience and the context of use. It reaches every firm regardless of risk class.

This duty has applied since 2 February 2025 — it is not deferred. A firm using only limited-risk tools still has to satisfy it; there is no minimal-risk carve-out.

For professional services the scope is concrete. Consultants, accountants, account managers and analysts who use AI on client work are "persons operating" those systems within the meaning of Article 4. They should understand what each tool can reliably do, where it fails, and what counts as appropriate use — an AI summary of a client's filings still needs a human check before it leaves the firm.

Article 4 has no penalty article dedicated solely to it, but it is a documented obligation authorities and clients can ask about in a due-diligence questionnaire, and a short literacy programme with a record of who completed it is the evidence. It is principle-based and proportionate, not a prescriptive curriculum.

---

Confidentiality and data governance over client data

Professional-services firms hold client confidential information, personal data, and regulated or contractual confidentiality obligations. Feeding that material into AI tools raises real governance questions, but those sit alongside the AI Act, not under it.

The AI Act does not create a new confidentiality regime for deployers of limited-risk tools. The governing rules are the General Data Protection Regulation (Regulation (EU) 2016/679), professional-secrecy duties under national law, and the confidentiality terms in client engagement letters — not the AI Act itself.

The Act's own data-governance rules are narrow. Article 10 — covering training, validation and testing data — binds providers of high-risk systems, and Article 78 protects confidentiality of information obtained by authorities under the Regulation within the high-risk regime. An ordinary advisory firm using third-party limited-risk tools is outside both.

The practical controls double as Article 4 and Article 50 evidence: a register of every AI vendor and tool and which client data each touches; contractual data-processing terms confirming the firm stays the controller; a rule on what client-identifying data may be entered into a tool; and a rule on when AI involvement must be flagged to a client or third party.

---

Article 50: transparency when AI output reaches clients or third parties

Article 50 imposes transparency duties, not a risk classification: it applies to specific interactions and outputs whether or not the system is high-risk, so a limited-risk firm is squarely in scope. Three triggers matter. Under Article 50(1), a chatbot that interacts with clients or prospects must tell them they are dealing with an AI system, unless it is obvious. Under Article 50(2), a provider of generative AI must mark its outputs as artificially generated in a machine-readable format. Under Article 50(4), deepfake and synthetic image, audio or video — and text published to inform the public on matters of public interest — must be disclosed as artificially generated or manipulated.

The content-marking obligations under Article 50 take effect on 2 December 2026. The Digital Omnibus was adopted in June 2026 — plan against 2 December 2027.

The practical line: set a house rule on labelling AI-assisted deliverables and any client-facing chatbot. The obligation falls on the firm running the interaction, not the client.

---

The Article 25 trap: when your firm becomes a provider

This is the most underrated exposure for agencies and consultancies that productise AI. Under Article 25(1), a deployer, distributor or importer is treated as the provider of a high-risk system — and takes on the full provider obligation stack — when it does any one of three things: puts its own name or trademark on the system, makes a substantial modification to it, or modifies the intended purpose so the system becomes high-risk.

The classic trigger is the first: a firm builds an internal AI tool, brands it, and licenses it to clients under its own name. It then steps into provider shoes — inheriting conformity assessment, Annex IV documentation and registration duties — but only if the productised use also falls within Annex III.

That conditional is the part firms miss, so be precise about the two-step test. Article 25 changes who the provider is; Article 6 and Annex III still decide whether the product is high-risk. Putting your trademark on a general document-drafting assistant makes you its provider, but that assistant is not in Annex III, so you are the provider of a non-high-risk product. It is the use case, not the badge, that pulls in Articles 8 to 15.

The decision rule is short. Building for internal advisory use makes you a deployer with limited-risk duties; rebranding and selling to clients means you re-run the Annex III test before launch. The one rebranded use that crosses into high-risk is screening.

---

The one clear high-risk line: recruiting and HR-services screening

The exception to the "mostly not high-risk" thesis is the staffing side. A recruiting, RPO or HR-services firm that runs AI to screen, filter, rank or evaluate candidates for clients is operating in Annex III point 4(a) — recruitment and selection — and that use is high-risk. Ordinary advisory consulting does not cross this line.

There is a prohibited overlay above the high-risk tier. Article 5(1)(f) prohibits emotion recognition in the workplace and education, and Article 5(1)(g) prohibits biometric categorisation that infers sensitive attributes such as race, political opinion or sexual orientation — both subject to narrow exceptions, in force since 2 February 2025. A screening vendor that bolts emotion analysis onto a video interview sits in the banned tier, not merely the high-risk one.

Operating a third-party screening tool makes the firm a deployer; building and branding its own can also make it the provider under Article 25, inheriting the full Articles 8 to 15 stack. The detailed mapping — provider versus deployer, conformity route, the GDPR Article 22 overlap — lives on the dedicated HR and recruitment guide.

---

Mapping common professional-services AI uses to risk tiers

Reading the table: use case decides the tier, not the firm

Only two rows leave limited or minimal territory: reselling a branded tool (which changes the role) and candidate screening (high-risk outright). "Limited risk" means transparency duties only — chiefly Article 50. "Minimal risk" means no specific AI Act obligation beyond the Article 4 literacy duty that applies to everyone.

What changes the tier (resale, screening, public-interest content)

*The content-marking duties under Article 50 take effect 2 December 2026; the stand-alone high-risk obligations under Article 6(2) apply from 2 December 2027 under the Digital Omnibus, adopted in June 2026. Three things move a firm out of the comfortable rows: reselling a branded tool, candidate screening, and publishing synthetic public-interest content. Absent those, advisory AI stays limited or minimal — the firm's whole footprint is a training record, a vendor register and a chatbot notice.

---

Worked example: a 140-person consultancy productises an AI tool

The advisory work: limited-risk, Article 4 and Article 50 only

Consider Meridian Advisory, a 140-person management-and-accounting consultancy in Dublin. Its consultants use AI for drafting, desk research and analysis on client engagements, and it runs a small support chatbot on its website.

The everyday work is unremarkable: drafting and research are limited or minimal risk, touching no Annex III area. Meridian's obligations are therefore Article 4 literacy training, a client-data input policy backed by a vendor register, and an Article 50(1) disclosure on the chatbot. No conformity assessment, no Annex IV file, no Article 49 registration.

The product launch: where Article 25 flips the role

Then Meridian pivots. It packages its internal model-review assistant, brands it "Meridian Assure", and licenses it to client finance teams. That branded product makes Meridian the provider under Article 25. Run the two-step test: a general finance-document assistant is not in Annex III, so Meridian is the provider of a non-high-risk product — Article 50 transparency may apply, but the Articles 8-15 stack does not. Change one fact: if "Meridian Assure" screened job candidates for client employers, it would land in Annex III point 4(a) and trigger the full high-risk provider stack.

The same firm carries trivial obligations for its advisory work and a materially heavier burden the moment it brands-and-resells into a high-risk use. The Article 25 plus Annex III two-step is the decision that matters.

---

Deadlines and penalties that apply to professional-services firms

What is in force now vs deferred

Three duties are already live: the Article 5 prohibitions and the Article 4 AI-literacy obligation since 2 February 2025, and GPAI model obligations under Articles 51 to 55 since 2 August 2025 (which bind the model provider, not its deployers).

Still coming: the stand-alone high-risk Annex III obligations under Article 6(2) apply from 2 December 2027 under the Digital Omnibus, adopted in June 2026, and Article 50 content-marking takes effect 2 December 2026. The Digital Omnibus, adopted in June 2026, settled these dates.

Penalty exposure by obligation

Fines scale with the breach:

• prohibited-practice infringements — including emotion recognition in hiring — up to €35 million or 7% of total worldwide annual turnover, whichever is higher (Article 99(3));
• most other provider, deployer and operator infringements up to €15 million or 3% (Article 99(4));
• supplying incorrect, incomplete or misleading information to authorities up to €7.5 million or 1% (Article 99(5)).

For SMEs and start-ups, Article 99(6) caps the fine at the lower of the fixed amount or the percentage. For an advisory firm the realistic exposure is the Article 5(1)(f)/(g) prohibition tier (only if it touches emotion recognition or sensitive categorisation in hiring) and the Article 99(5) information-accuracy tier — not the high-risk fines.

---

How Confir helps

Most professional-services firms need two things: an inventory of every AI tool in use, and an honest classification of each so effort lands where obligation sits. Confir's rule-based engine walks each tool through the Article 6 and Annex III analysis — asking about intended use, whether the firm is a deployer or a provider under Article 25, and whether the use is screening or public-interest content — then records the risk tier and the obligations that attach. The reasoning is deterministic and rule-based: no model inference, no hallucination. For a firm productising a tool, it flags when a rebrand triggers the Article 25 re-test. GPAI workflow support is partial and on the roadmap.

---

Frequently asked questions

Is AI used in consulting or accounting work high-risk under the EU AI Act?

Almost never. High-risk status depends on the use case, not the sector. Drafting, summarising, desk research and financial-analysis support map to none of the eight Annex III areas under Article 6, so they sit at limited or minimal risk with no conformity assessment, no technical file and no registration. The obligations that do apply come from Article 4 AI literacy, Article 50 transparency, the GDPR and client confidentiality, and the Article 25 provider trap if the firm resells a branded AI tool.

Does my firm have to train staff on AI even if our tools are low-risk?

Yes. Article 4 requires providers and deployers to ensure a sufficient level of AI literacy among the people operating AI systems on their behalf, and it has applied since 2 February 2025 regardless of risk class. For a professional-services firm, that means consultants, accountants and analysts using AI on client work should understand each tool's capabilities, limits and appropriate use. Article 4 is principle-based rather than a fixed curriculum, but a documented training programme is the evidence authorities and clients expect.

When does a professional-services firm become a "provider" under the AI Act?

Under Article 25(1), a deployer becomes a provider when it puts its own name or trademark on a system, makes a substantial modification, or changes the intended purpose so the system becomes high-risk. The classic trigger is building an internal AI tool, branding it, and licensing it to clients. Article 25 changes who the provider is; whether the product is high-risk still depends on Article 6 and Annex III, so re-run that test before launch. Naming alone does not make a non-high-risk tool high-risk — it is the use case that matters.

Do we have to tell clients when a deliverable was made with AI?

It depends on the format, not on a general rule. Article 50(1) requires you to inform people when they interact with a chatbot or conversational AI, unless it is obvious. Article 50(2) and 50(4) require machine-readable marking of generative-AI outputs and disclosure of synthetic image, audio or video and certain public-interest text. These content-marking duties take effect 2 December 2026. Plain AI-assisted text in an internal memo is not generally a labelled deliverable, but a consistent house labelling rule keeps you on the right side of the chatbot disclosure duty.

Is a recruiting or staffing firm's AI candidate screening high-risk?

Yes. AI used to screen, filter, rank or evaluate candidates falls within Annex III point 4(a) and is high-risk, triggering the Articles 8-15 obligation stack. Emotion recognition in hiring or the workplace is separately prohibited under Article 5(1)(f), and biometric categorisation inferring sensitive attributes under Article 5(1)(g). This is the clearest high-risk line in professional services, and ordinary advisory consulting does not cross it. Detailed provider-versus-deployer mapping lives on our HR and recruitment guide.

What are the penalties if a firm gets this wrong?

Fines scale with the breach. Prohibited-practice infringements — for example, emotion recognition in hiring — reach up to €35 million or 7% of total worldwide annual turnover, whichever is higher, under Article 99(3). Most other operator infringements reach up to €15 million or 3% under Article 99(4). Supplying incorrect, incomplete or misleading information to authorities reaches up to €7.5 million or 1% under Article 99(5). For SMEs and start-ups, Article 99(6) caps the fine at the lower of the two. For most advisory firms the realistic exposure is the information-accuracy tier.

Which AI Act deadlines actually affect professional-services firms right now?

Two duties are already live: the Article 5 prohibitions and the Article 4 AI-literacy obligation, both in force since 2 February 2025. GPAI obligations under Articles 51-55 applied from 2 August 2025. Article 50 content-marking takes effect 2 December 2026. Stand-alone high-risk obligations under Article 6(2) apply from 2 December 2027 under the Digital Omnibus, adopted in June 2026. The Digital Omnibus, adopted in June 2026, settled these dates.

---

Related guides

• HR and recruitment AI under the EU AI Act
• EU AI Act compliance for law firms
• Article 25: when you become a provider
• Article 50 transparency obligations
• Article 4 AI literacy
• what limited risk means