The EU AI Act for back-office AI: what owners must do, and by when

The highest penalty in the EU AI Act is EUR 35 000 000 or 7 % of total worldwide annual turnover, whichever is higher.¹ It applies to a short list of banned practices, and most back-office automation is nowhere near it. The part of the Act that gives an operations team real work is the second tier: fines of up to EUR 15 000 000 or 3 % for companies that breach the duties of those who use AI systems built by someone else.¹

The timetable changed this summer. Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026, moved the obligations for high-risk systems listed in Annex III from 2 August 2026 to 2 December 2027.² Some executives will read that as a reprieve. It is a partial one. The prohibitions have applied since 2 February 2025, the AI literacy duty in Article 4 since the same day, and the transparency rules in Article 50 since 2 August 2026.¹ This article maps those obligations for companies that run AI in back-office workflows such as hiring, credit decisions, document handling and customer messaging.

What the Act bans outright

Article 5 lists the practices the Act prohibits, and they have applied since 2 February 2025.¹ Most sit far from a finance or operations team, but several can touch a workflow. Among the prohibited practices:

• subliminal, manipulative or deceptive techniques that cause significant harm;
• exploiting vulnerabilities linked to age, disability or social or economic situation;
• social scoring of people;
• inferring the emotions of people in the workplace or in education institutions, except for medical or safety reasons;
• biometric categorisation used to infer race, political opinions, trade union membership, beliefs, sex life or sexual orientation.

The workplace emotion rule deserves attention from operations leaders. A tool that infers staff emotions from voice or video is inside the ban unless its purpose is medical or safety-related.¹ The amending regulation also adds two prohibitions on certain sexually explicit synthetic content, which apply from 2 December 2026.¹ A breach of any Article 5 prohibition carries the top fine ceiling.

When a back-office tool becomes high-risk

High-risk status is defined by use. Annex III names the uses, and four of them matter to a typical employer or lender:¹

• point 4(a): systems for recruitment or selection, including targeted job ads, analysing and filtering applications, and evaluating candidates;
• point 4(b): systems that make or influence decisions on the terms of work relationships, promotion or termination, that allocate tasks based on behaviour or personal traits, or that monitor and evaluate performance and behaviour;
• point 5(b): systems that evaluate the creditworthiness of natural persons or establish their credit score, except those used to detect financial fraud;
• point 5(c): risk assessment and pricing for life and health insurance.

The obligations for these systems now apply from 2 December 2027, and for systems covered by the product-safety legislation in Annex I from 2 August 2028.² The original date for both was 2 August 2026.²

What is missing from Annex III matters as much as what is on it. As we read the list, tools that extract data from invoices, route support tickets or draft routine emails are not named. A single feature can change that: a CV-ranking function added to a general-purpose assistant falls under point 4(a) because of what it does, whatever the product is called. The delay also gives no relief to systems bought now. A screening or credit tool adopted this year will still be running in December 2027, and it is cheaper to ask the vendor the right questions before signing than after.

What a deployer has to do

A company that uses a high-risk system supplied by a vendor is a deployer, and Article 26 sets its duties.¹ In summary:

• use the system according to the provider’s instructions for use;
• assign human oversight to people who have the competence, training, authority and support to do it;
• where you control the input data, make sure it is relevant and representative for the intended purpose;
• monitor the system, inform the provider or the authority and suspend use when it presents a risk, and report serious incidents;
• keep the automatically generated logs for at least six months;
• as an employer, inform workers’ representatives and the affected workers before putting a high-risk system into use at the workplace;
• where Annex III decisions about natural persons are made or assisted by the system, inform those persons.

Article 26 also tells deployers to use the provider’s information to carry out a data protection impact assessment, and to cooperate with the authorities.¹ A separate fundamental rights impact assessment under Article 27 applies to public bodies and to some private deployers, including those using credit scoring and life and health insurance systems; we did not verify its full scope, so ask counsel whether it reaches you.

Several of these duties depend on documents only the vendor can provide. You cannot follow instructions for use you never received, and you cannot keep logs a system does not produce. That is our reading and no clause of the Act says it, yet it is a practical reason to write both into the contract.

Literacy and transparency already apply

Article 4 changed shape this summer. Recital 8 of the amending regulation describes the earlier text as an obligation to ensure AI literacy.² The replacement asks providers and deployers to take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account technical knowledge, experience, education, training and context.² The new text sets a lower bar:

…does not require providers or deployers to guarantee any specific level of AI literacy of any individual.

The text does not say how to show that you took measures. A dated record of who was trained on what, adjusted to their role, is a sensible way to do it; that is a suggestion and the Act does not require it.

Article 50 has applied since 2 August 2026.¹ Providers of AI systems that interact with people must ensure those people are told they are dealing with an AI system, unless that is obvious. Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them. Deployers of deep fakes must disclose them. Deployers of AI-generated text published to inform the public on matters of public interest must disclose it, unless a person has reviewed or edited the text and holds editorial responsibility.¹ The information must be given clearly, at the latest at first interaction or exposure.¹ If a vendor built your customer chatbot, the first duty falls on the provider, so ask how the notice is shown; if you built it yourself, who counts as the provider is a question for counsel.

One transition applies to generative systems. Providers whose systems were placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking duty in Article 50(2).²

What the fines look like

Article 99 sets maximum amounts, each stated as a fixed sum or a share of worldwide annual turnover for the preceding financial year, whichever is higher. A breach of an Article 5 prohibition: up to EUR 35 000 000 or 7 %. A breach of the deployer duties in Article 26 or the transparency duties in Article 50: up to EUR 15 000 000 or 3 %. Supplying incorrect, incomplete or misleading information to authorities: up to EUR 7 500 000 or 1 %.¹ For small and medium-sized enterprises the lower of the two amounts applies, and the consolidated text extends the same rule to small mid-caps for the EUR 15 000 000 and EUR 7 500 000 tiers.¹

A ceiling is not a forecast, and nothing in the text tells you what regulators will actually impose. The more likely cost for most companies, in our judgement, is rework: a hiring or credit workflow built without logs, named reviewers or a worker notice may have to be rebuilt before December 2027. The rules are also still moving: the July 2026 amendment softened the literacy duty and pushed the Annex III deadline back by sixteen months, so a plan built on today’s text may need adjusting.

What to do next

None of the steps below needs a legal team to start. They need a list of what you run and the name of someone who owns it.

1. List every AI system in use, including features inside software you already license, and note what each one decides or influences.
2. Check each system against Article 5 and Annex III, and write down the reason for each answer.
3. For any HR or credit use, ask the vendor for the instructions for use and for access to logs, and name the people who will oversee it before 2 December 2027.
4. Put a dated AI literacy measure in place by role, and keep the record.
5. Review customer-facing chatbots and published AI-generated text against Article 50, and check the 2 December 2026 marking date with each generative-AI vendor.

If you do not yet know where AI sits in your workflows, the free pre-audit is a short questionnaire that returns a first estimate of where AI could cut cost. It is not a compliance check.