A customer asks a bank's assistant whether she qualifies for a loan. The assistant answers in fluent, confident language. Under the EU AI Act, two questions follow. Was she told she was talking to an AI system? And was the assistant only answering, or was it evaluating her creditworthiness?

The first question has been live since August 2, 2026. The second decides whether the system is high-risk, and the date for that changed this summer. This guide covers what Regulation (EU) 2024/1689 asks of teams that deploy or provide customer-facing generative AI and agents, and the evidence each duty implies. Dates and quotes reflect the consolidated text as of November 11, 2026.

Key takeaways
  • The Act assigns duties by role. A deployer uses an AI system under its authority, and a provider places one on the market under its own name. It also reaches firms outside the EU when the output is used in the Union.
  • Article 50 transparency duties have applied since August 2, 2026: people must be told they are interacting with an AI system unless it is obvious, and generated content must be marked in a machine-readable format.
  • A system is high-risk because of its intended use. Credit scoring and life and health insurance pricing are listed in Annex III, and those high-risk rules now apply from December 2, 2027 after the AI Omnibus.
  • High-risk systems need documented risk management, testing against prior defined metrics, logging, human oversight and post-market monitoring. Deployers keep logs for at least six months.
  • Fines reach EUR 35 000 000 or 7 % of worldwide annual turnover for prohibited practices, and EUR 15 000 000 or 3 % for breaches of provider, deployer and transparency obligations.

Who is in scope: providers, deployers and reach outside the EU

The Act assigns duties by role. A provider is a person or body "that develops an AI system or a general-purpose AI model," or has one developed, and places it on the market or puts it into service "under its own name or trademark" (Article 3(3)). A deployer is a person or body "using an AI system under its authority," outside personal non-professional activity (Article 3(4)).

A bank that runs a vendor's assistant under the bank's authority is a deployer. A bank that has an assistant built on a third-party model and puts it into service under its own name can be a provider of that system. That is our reading of the definitions, and a question for your counsel.

Location does not settle it. Article 2(1) covers providers placing systems on the Union market "irrespective of whether those providers are established or located within the Union or in a third country." It covers deployers "that have their place of establishment or are located within the Union." And it covers providers and deployers in a third country "where the output produced by the AI system is used in the Union."

One duty applies to both roles regardless of risk tier. Providers and deployers "shall take measures to support the development of AI literacy of their staff" and others who operate AI systems for them (Article 4). The AI Omnibus added that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual."

The risk tiers

The Commission describes four levels of risk.

  • Unacceptable risk. Practices listed in Article 5 are banned. The Commission describes them as "a clear threat to the safety, livelihoods and rights of people."
  • High risk. "AI use cases that can pose serious risks to health, safety or fundamental rights." These carry the documentation, testing, logging and oversight duties described below.
  • Transparency risk. Systems where people need to know AI is involved. This is where most chatbots and generated content sit, under Article 50.
  • Minimal or no risk. "The AI Act does not introduce rules for AI that is deemed minimal or no risk."

General-purpose AI models have their own chapter. It sits beside the tiers, not inside them.

Which customer-facing uses are high-risk

A system is high-risk because of its intended use, not its technology. Article 6(2) says "AI systems referred to in Annex III shall be considered to be high-risk." For financial services, insurance and healthcare, the relevant entries are in Annex III, point 5:

  • Credit. Systems "intended to be used to evaluate the creditworthiness of natural persons or establish their credit score, with the exception of AI systems used for the purpose of detecting financial fraud" (5(b)).
  • Life and health insurance. Systems "intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance" (5(c)).
  • Public benefits and healthcare services. Systems used by or for public authorities "to evaluate the eligibility of natural persons for essential public assistance benefits and services, including healthcare services" (5(a)).
  • Emergency triage. Systems that evaluate and classify emergency calls or dispatch first responders, and "emergency healthcare patient triage systems" (5(d)).

There is a narrow exit. An Annex III system is not high-risk "where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making" (Article 6(3)). The conditions include a system that performs "a narrow procedural task" or "a preparatory task to an assessment." The exit closes if the system "performs profiling of natural persons." A provider that relies on it "shall document its assessment" before placing the system on the market (Article 6(4)).

For generative AI, the line is easy to cross without noticing. An assistant that explains loan products is not evaluating creditworthiness. An agent that reads an application and recommends approve or decline may be. Our inference: the written intended purpose, and tests that show the system stays inside it, are the first evidence to hold.

Transparency duties for chatbots and generated content

Article 50 applies to systems in any tier. It has applied since August 2, 2026 (Commission timeline).

Chatbots and agents. Providers must ensure that systems "intended to interact directly with natural persons" are built so that those persons "are informed that they are interacting with an AI system, unless this is obvious" to a reasonably well-informed person (Article 50(1)).

Generated content. Providers of systems "generating synthetic audio, image, video or text content" must ensure outputs "are marked in a machine-readable format" and are "detectable as artificially generated or manipulated" (Article 50(2)). The duty does not apply where the system performs "an assistive function for standard editing." Systems placed on the market before August 2, 2026 have until December 2, 2026 to comply (Article 111(4)).

Deployers. Deployers of a system that produces a deep fake "shall disclose that the content has been artificially generated or manipulated." Deployers that publish AI-generated text "with the purpose of informing the public on matters of public interest" must disclose it, unless the text "has undergone a process of human review or editorial control" (Article 50(4)).

Timing. The information must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" (Article 50(5)).

Two supporting documents are at different stages. The Commission lists a final code of practice on AI-generated content, published June 10, 2026. It is voluntary. The Commission's guidelines on Article 50 were published as a draft on May 8, 2026. As of the page's October 19, 2026 update, no final version is listed.

General-purpose AI models and what flows to deployers

Most customer-facing generative AI runs on a general-purpose AI model from a third party. The model provider's duties have applied since August 2, 2025 (Article 113(b)). The Commission's enforcement powers over those providers began on August 2, 2026, and models placed on the market before August 2, 2025 must comply by August 2, 2027 (Commission guidelines page).

Under Article 53(1), model providers must keep technical documentation "including its training and testing process and the results of its evaluation." They must also give documentation to "providers of AI systems who intend to integrate the general-purpose AI model into their AI systems," enough to "have a good understanding of the capabilities and limitations" of the model. The General-Purpose AI Code of Practice, published July 10, 2025, is "a voluntary tool" with three chapters: Transparency, Copyright, and Safety and Security.

Three points matter downstream.

  • The documentation goes to system providers. Article 53 does not name deployers. A deployer that buys a finished assistant gets model information through its vendor and its contract.
  • Changing the purpose can change your role. A deployer or other third party is treated as a provider of a high-risk system if it modifies "the intended purpose of an AI system, including a general-purpose AI system," so that it "becomes a high-risk AI system" (Article 25(1)(c)). Pointing a general assistant at credit decisions is the obvious example.
  • Suppliers owe cooperation in writing. The provider of a high-risk system and a third party that supplies "an AI system, AI model, tools, services, components, or processes" must "by written agreement, specify the necessary information, capabilities, technical access and other assistance" (Article 25(4)).

The records and evidence the Act expects

The requirements below attach to high-risk systems. They are also the clearest statement in the Act of what good evidence looks like.

Risk management. "A risk management system shall be established, implemented, documented and maintained." It is "a continuous iterative process planned and run throughout the entire lifecycle" (Article 9(1) and (2)).

Testing. Testing must happen "prior to their being placed on the market or put into service," and "against prior defined metrics and probabilistic thresholds that are appropriate to the intended purpose" (Article 9(8)).

Logging. High-risk systems "shall technically allow for the automatic recording of events (logs) over the lifetime of the system" (Article 12(1)). Deployers keep those logs for "at least six months" (Article 26(6)).

Human oversight. Systems must be built so "they can be effectively overseen by natural persons." The people assigned must be able to "disregard, override or reverse the output" and stay aware of "automation bias" (Article 14). Deployers assign oversight to people with "the necessary competence, training and authority" (Article 26(2)).

Monitoring after launch. Providers "shall establish and document a post-market monitoring system" that collects data "on the performance of high-risk AI systems throughout their lifetime" (Article 72). Deployers "shall monitor the operation of the high-risk AI system on the basis of the instructions for use" and report risks and serious incidents (Article 26(5)).

Retention. Providers keep technical documentation and quality management records "for a period ending 10 years" after the system is placed on the market (Article 18(1)).

Deployers of credit and insurance systems. Deployers of systems under Annex III points 5(b) and 5(c) "shall perform an assessment of the impact on fundamental rights" before first use (Article 27(1)). Deployers of Annex III systems that make or assist decisions about people "shall inform the natural persons that they are subject to the use of the high-risk AI system" (Article 26(11)).

Financial institutions. The Act connects to existing rules in several places. For deployers that are financial institutions subject to internal governance requirements under Union financial services law, the monitoring duty "shall be deemed to be fulfilled by complying with the rules on internal governance arrangements, processes and mechanisms" (Article 26(5)). Logs and technical documentation are kept as part of the documentation already required under that law (Articles 26(6) and 18(3)).

Penalties

Article 99 sets three ceilings. Prohibited practices carry fines "of up to EUR 35 000 000" or "up to 7 % of its total worldwide annual turnover for the preceding financial year, whichever is higher." Breaches of provider obligations (Article 16), deployer obligations (Article 26) and transparency obligations (Article 50) carry up to EUR 15 000 000 or 3 %. Supplying "incorrect, incomplete or misleading information" to authorities carries up to EUR 7 500 000 or 1 %. For SMEs, the lower of the amount and the percentage applies (Article 99(6)).

The timeline as of November 11, 2026

The dates changed in 2026. The AI Omnibus, Regulation (EU) 2026/1744, was published in the Official Journal on July 24, 2026 and entered into force on July 27, 2026. It is adopted law, not a proposal. The table reflects Article 113 and Article 111 as amended.

DateWhat appliesStatus
February 2, 2025 Chapters I and II: definitions, AI literacy and prohibited practices (Article 113(a)). In effect
August 2, 2025 Obligations for general-purpose AI model providers, and governance (Article 113(b)). In effect
August 2, 2026 General application: "It shall apply from 2 August 2026." Includes Article 50 transparency duties and the start of enforcement (Commission timeline). In effect
December 2, 2026 New prohibitions added by the AI Omnibus (Article 113(a)). Article 50(2) marking for systems placed on the market before August 2, 2026 (Article 111(4)). Upcoming
August 2, 2027 General-purpose AI models placed on the market before August 2, 2025 must comply (Article 111(3)). Upcoming
December 2, 2027 High-risk requirements and obligations for systems "classified as high-risk pursuant to Article 6(2) and Annex III" (Article 113(c)(i)). Moved by the AI Omnibus. Upcoming
August 2, 2028 The same for systems "classified as high-risk pursuant to Article 6(1) and Annex I" (Article 113(c)(ii)). Moved by the AI Omnibus. Upcoming

One transitional rule matters for systems already live. The Act applies to operators of high-risk systems placed on the market before the application date "only if, as from that date, those systems are subject to significant changes in their designs" (Article 111(2)). Whether a model swap or a prompt rewrite counts is a question for your counsel.

How GenGuardX helps organize the evidence

The later date for Annex III systems is time to build a record, not a reason to wait. GenGuardX is where business teams test a GenAI application, risk teams approve it against evidence, and both monitor it after launch. The table shows how that evidence can be mapped to the Act's articles. The mapping supports your program. It does not make a system compliant, and conformity assessment and legal classification stay with you and your counsel.

What the Act asks forGenAI evidenceGenGuardX capability
Documented, continuous risk management (Article 9) Every prompt, retrieval source, model and agent, with an owner, an intended purpose and the risks identified for it. Inventory management and lineage tracking, with findings tracked from test to fix.
Testing against prior defined metrics and thresholds (Article 9(8)) Versioned tests for groundedness, policy adherence, PII leakage and safe escalation, plus scenarios run by business users. Standardized evals, Human Integrated Testing and annotation queues.
Human oversight by people with competence and authority (Articles 14 and 26(2)) Who reviewed the system, who approved it, on what evidence, and when. Approval workflows and role-based access in the Governed Workbench.
Monitoring in operation (Articles 26(5) and 72) Evidence the system still behaves as approved after vendor model updates and policy changes. Monitoring for drift from the approved version, with alerts on live traffic and re-evaluation.
Staying inside the intended purpose (Articles 6 and 25(1)(c)) Tests showing an assistant declines or escalates requests that would turn it into a credit or pricing decision. Risk dashboards & curated reports and the Trust Score, tracked version over version.

Event logs under Article 12 are produced by the AI system itself and retained by its provider and deployer. They sit beside this record, not inside it.

What are the trade-offs of building EU AI Act evidence now?

Building the record before the high-risk dates costs documentation, testing and legal review time now, and in return one body of evidence can serve several articles of the Act at once.

What you gainWhat it costs
One set of test results can support risk management (Article 9), human oversight (Article 14) and monitoring (Article 26(5)).The risk management system must be documented and maintained "throughout the entire lifecycle," which is ongoing work, not a one-time filing.
Tests against metrics and thresholds defined in advance, as Article 9(8) asks.Defining metrics and thresholds for each intended purpose takes expert time before any testing starts.
A clear record of intended purpose, which helps show when a change would make you a provider under Article 25(1)(c).Changes to the model, prompts or purpose need retesting and review, which slows releases.
Records ready when an authority asks.Deployer logs kept for at least six months and provider documentation kept for 10 years carry storage and retention costs.
Time to fix gaps before December 2, 2027.The dates and rules already changed once in 2026, so some work may need revising.
Evidence kept in one place, if you use a platform such as GenGuardX.A platform adds license, setup and reviewer training costs, and it does not replace conformity assessment or legal classification.

To keep the costs down, classify each system first and put full effort only where a system is likely to be high-risk; financial institutions can also build on the internal governance records Union financial services law already requires (Article 26(5)).

FAQ

Does the EU AI Act apply to companies outside the EU?

Yes, in three cases. It covers providers that place AI systems or general-purpose AI models on the Union market, "irrespective of whether those providers are established or located within the Union or in a third country." It covers deployers established or located in the Union. And it covers providers and deployers in a third country "where the output produced by the AI system is used in the Union" (Article 2(1)).

Is a customer service chatbot high-risk under the EU AI Act?

Not because it is a chatbot. A system is high-risk when its intended use falls in an Annex III area, such as evaluating the creditworthiness of natural persons or risk assessment and pricing for life and health insurance. A chatbot outside those uses still falls under Article 50: people must be informed that they are interacting with an AI system, unless this is obvious.

When do the high-risk obligations apply?

As of November 11, 2026, Article 113 sets December 2, 2027 for systems that are high-risk under Annex III, and August 2, 2028 for systems that are high-risk under Article 6(1) and Annex I. The AI Omnibus, in force since July 27, 2026, set these dates.

What are the fines under the EU AI Act?

Article 99 sets three ceilings. Prohibited practices: up to EUR 35 000 000 or 7 % of total worldwide annual turnover. Breaches of provider, deployer and transparency obligations: up to EUR 15 000 000 or 3 %. Incorrect, incomplete or misleading information to authorities: up to EUR 7 500 000 or 1 %. For SMEs, the lower of the amount and the percentage applies.

Last reviewed: November 11, 2026. This article summarizes public law and guidance and is not legal advice.

See how GenGuardX organizes GenAI testing and approval evidence against the EU AI Act.