The EU AI Act and your chatbot: who is liable when it gets it wrong

1.Key regulatory milestones

The AI Act, Regulation (EU) 2024/1689, applies through a phased implementation timetable (Article 113). The Digital Omnibus on AI, Regulation (EU) 2026/1744, was adopted on 8 July 2026, published on 24 July, and entered into force on 27 July 2026: it adjusted the high-risk application dates and clarified the organizational AI literacy requirement. Every date listed below represents enacted European law, not pending proposals.

1. 1 August 2024In force

   Entry into force of the EU AI Act

   Regulation (EU) 2024/1689 takes legal effect across the European Union, initiating phased transition periods.

2. 2 February 2025In force

   Prohibited AI practices and literacy duties

   Universal bans on manipulative, deceptive, and exploitative AI practices (Article 5) and baseline AI literacy mandates (Article 4) become enforceable.

3. 2 August 2025In force

   General-purpose AI governance and penalty framework

   Statutory rules for providers of general-purpose AI models (Chapter V), European AI Office governance structures, and administrative sanction caps (Article 99) take effect.

4. 27 July 2026In force

   Digital Omnibus on AI takes effect

   Regulation (EU) 2026/1744 reframes organizational AI literacy under Article 4 and extends transitional deadlines for high-risk systems.

5. 2 August 2026In force

   Chatbot transparency and general enforcement

   Article 50 applies in full: natural persons must be informed that they are conversing with an AI system unless immediately obvious. National market surveillance authorities assume active enforcement powers.

6. 2 December 2026Upcoming

   Marking transition for generative systems and specialized bans

   Generative systems commercialized prior to 2 August 2026 must ensure synthetic content is marked in machine-readable format (Article 50(2) and Article 111(4)). Specific prohibitions on non-consensual synthetic intimate imagery and child exploitation content apply.

7. 9 December 2026Upcoming

   Revised Product Liability Directive transposed

   Member States must transpose Directive (EU) 2024/2853 into national law. Strict product liability, discovery orders, and presumptions of defect apply to software and AI systems put into service after this date.

8. 2 December 2027Upcoming

   Annex III high-risk standalone AI applications

   Stringent provider and deployer obligations become binding for Annex III domains, including credit scoring, automated recruitment, and health and life insurance risk assessment. Postponed from 2 August 2026 by the Digital Omnibus.

9. 2 August 2028Upcoming

   Annex I high-risk regulated product components

   Obligations apply to AI systems integrated as safety components in products covered by EU harmonization legislation, such as medical devices, aviation, and machinery. Postponed from 2 August 2027.

2.What the AI Act demands of an enterprise chatbot

The vast majority of commercial customer-facing systems—such as website help desks, WhatsApp booking assistants, and automated customer support agents—are not classified as high-risk systems. For these standard commercial deployments, three immediate sets of obligations apply today, alongside a fourth if the system makes automated decisions regarding individuals.

• Provide clear, timely notice of AI interaction. Since 2 August 2026, AI systems engineered to interact directly with natural persons must be designed to inform users of their artificial nature, unless this is self-evident to a reasonably well-informed, observant observer (Article 50(1)). Transparency notices must be clear, easily distinguishable, and delivered no later than the initial point of interaction (Article 50(5)). The European Commission's formal guidelines of 20 July 2026 emphasize that customer support bots whose responses could reasonably be mistaken for human dialogue can never rely on the “obvious” exemption.
• Maintain demonstratable AI literacy across operational teams. Both providers and deployers must institute concrete internal measures to foster AI literacy among staff responsible for designing, deploying, and overseeing AI systems (Article 4, as amended by Regulation (EU) 2026/1744). While the law does not prescribe rigid standardized exam scores for individual employees, it requires organizations to document ongoing, role-appropriate training initiatives.
• Strictly respect prohibited AI practices. Article 5 strictly outlaws systems employing subliminal, deceptive, or manipulative techniques that impair autonomous decision-making and cause significant harm, as well as systems exploiting vulnerabilities related to age, physical or mental disability, or specific socio-economic conditions. These statutory prohibitions have been active law since 2 February 2025.
• Verify whether automated decisions trigger Annex III high-risk status. Using AI to evaluate creditworthiness, establish credit scores, or assess underwriting risk and premium pricing in life and health insurance constitutes high-risk AI (Annex III, point 5). Deployers of high-risk systems must operate them strictly in accordance with technical instructions, delegate oversight to trained personnel, verify the quality of input data, maintain continuous monitoring, retain system event logs for at least six months, and notify affected individuals (Article 26). These high-risk duties become enforceable on 2 December 2027.

3.Allocating responsibility: provider, deployer, and your organization

The EU AI Act allocates duties according to defined functional roles rather than software authorship alone. A single enterprise can hold multiple roles simultaneously. Civil liability toward your customers, however, is an independent question governed by civil and consumer law.

• Your company, toward its customers

  Typical entity: The commercial enterprise providing the conversational interface on its website, application, or phone line.

  Legal scope of responsibility: Civil accountability for all statements, pricing, and commitments communicated by the assistant, under standard contract and consumer protection law, regardless of regulatory categorization under the AI Act.

• Provider (AI Act, Article 3(3))

  Typical entity: The entity that develops an AI system (or commissions its development) and places it on the market or into service under its own name or trademark.

  Legal scope of responsibility: Ensuring the technical architecture discloses AI identity (Article 50(1)) and, if the system is classified as high-risk, satisfying every provider duty, logging standard, and risk management obligation.

• Deployer (AI Act, Article 3(4))

  Typical entity: Any natural or legal person using an AI system under its professional authority, such as an enterprise running a third-party commercial assistant.

  Legal scope of responsibility: Fulfilling workforce AI literacy mandates (Article 4), delivering user-facing disclosures under Article 50(3) and 50(4) when applicable, and executing operational duties under Article 26 for high-risk systems.

• Foundation model vendor

  Typical entity: The provider of the underlying general-purpose AI model (e.g., Anthropic, OpenAI, Google, Microsoft).

  Legal scope of responsibility: Its own duties as a provider of general-purpose AI models (Articles 53 to 55), including technical documentation and systemic risk mitigation. Upstream model providers do not assume your civil liabilities toward end consumers.

• Integrator or agency

  Typical entity: The external development consultancy or technical partner that assembled or fine-tuned the assistant for your company.

  Legal scope of responsibility: Contractual deliverables defined in your master services agreement. If the agency distributes the resulting product under its own brand, it assumes the legal role of provider.

When an end consumer receives false or damaging advice, legal analysis begins with basic agency and contract law. In the landmark case Moffatt v. Air Canada (2024 BCCRT 149, 14 February 2024), an airline's automated chatbot erroneously informed a customer that bereavement rates could be applied retroactively after completing travel, directly contradicting the airline's published terms. Air Canada defended itself by arguing that its chatbot was an independent legal entity responsible for its own actions.

The Civil Resolution Tribunal firmly rejected this defense, stating that “it should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.” The tribunal ordered the airline to pay damages, interest, and costs. While rendered by a Canadian tribunal, this ruling mirrors the exact legal rationale European civil courts and consumer protection authorities apply.

4.Civil liability when an assistant delivers erroneous output

The EU AI Act is predominantly administrative public law: market regulators conduct audits and levy fines. Compensation for an injured, misled, or financially harmed customer stems from civil codes and consumer protection directives that are fully operational today.

• Consumer protection and contract law apply today. European law classifies commercial practices as unfair and misleading if they contain inaccurate information or deceive an average consumer regarding product characteristics, pricing, payment terms, or legal rights (Directive 2005/29/EC, Article 6). In Spain, this principle is codified in the Unfair Competition Act (Ley 3/1991, Articles 5 and 7). Spain's Consumer Protection Act (TRLGDCU) explicitly requires truthful, complete pre-contractual disclosures (Article 60) and establishes that pre-contractual representations and promotional statements bind the merchant even if omitted from the final signed contract (Article 61.2). When a chatbot promises a commercial condition, businesses cannot simply disclaim legal intent.
• The Revised Product Liability Directive takes effect on 9 December 2026. Directive (EU) 2024/2853 explicitly brings software and AI systems under the legal definition of a product (Article 4(1)), treating AI providers as statutory manufacturers (recital 13). The regime imposes strict, no-fault liability for wrongful death, personal injury (including medically recognized psychological harm), property destruction, and the loss or corruption of personal data (Article 6). Crucially, civil courts are authorized to compel defendants to disclose internal technical evidence (Article 9) and must presume product defectiveness whenever technical complexity renders proof disproportionately burdensome for the claimant (Article 10). Substantially modifying an existing AI system shifts statutory manufacturer liability onto the modifying party (Article 8(2)).
• Withdrawal of the separate AI Liability Directive proposal. The European Commission formally withdrew its 2022 proposal for a standalone AI Liability Directive (AILD); this formal withdrawal was recorded in the Official Journal on 6 October 2025 (C/2025/5423). As a result, civil claims arising from AI malfunctions proceed directly under national civil and commercial codes, consumer legislation, and the modernized Product Liability Directive.
• Administrative fines under the AI Act operate on an independent track. Violating transparency obligations under Article 50 or deployer obligations under Article 26 carries administrative penalties of up to €15 million or 3% of worldwide annual turnover, whichever is higher; prohibited practices trigger penalties of up to €35 million or 7% (Article 99). For SMEs and start-ups, the statutory penalty is capped at whichever figure is lower (Article 99(6)), a protective ceiling extended to small mid-caps since July 2026 (Article 99(6a)). Each EU Member State establishes designated enforcement authorities and procedural sanction structures.

5.Where customer-facing assistants fail in practice

Empirical audits across enterprise conversational systems reveal recurrent failure patterns, each implicating distinct statutory liabilities.

• Quoting obsolete pricing or fee structures

  An assistant quotes expired discount tiers or last year's rate schedule as currently valid terms.

  Applicable legislation: Misleading commercial practices (Directive 2005/29/EC, Article 6; Ley 3/1991, Article 5); binding pre-contractual offers and disclosures (TRLGDCU, Articles 60 and 61).

• Fabricating coverage terms or policy exceptions

  An insurance assistant claims a surgical procedure carries no waiting period, or an airline bot promises retroactive fare adjustments, as in Moffatt.

  Applicable legislation: Breach of contract, detrimental reliance, and statutory consumer protection: the consumer reasonably relied on official company channels.

• Misrepresenting mandatory consumer rights

  An automated agent tells a purchaser that goods cannot be returned, or misstates statutory withdrawal windows.

  Applicable legislation: Deceptive commercial practices regarding consumer statutory rights (Directive 2005/29/EC, Article 6(1)(g)).

• Concealing artificial nature from users

  A customer chat widget featuring a real employee photograph and human persona operates without stating it is automated software.

  Applicable legislation: Infringement of EU AI Act transparency mandates (Article 50(1) and (5)), enforceable since 2 August 2026.

• Dispensing harmful technical or medical guidance

  An automated assistant supplies erroneous pharmaceutical dosage guidelines, flawed equipment repair sequences, or unsafe component tolerances.

  Applicable legislation: Tort liability for personal injury and property destruction; governed by strict liability under Directive (EU) 2024/2853 from 9 December 2026.

• Executing unauthorized operational actions

  An autonomous agent mistakenly cancels an active booking, processes an unearned refund, or alters contract parameters.

  Applicable legislation: Direct contractual liability for acts executed through authorized systems; indemnification against vendors depends entirely on negotiated contract terms.

6.Spain: AESIA governance and national legislative landscape

The EU AI Act is a European regulation: it is directly applicable and legally binding throughout Spain without requiring transposition legislation (Article 113).

Spain established the Spanish Agency for the Supervision of Artificial Intelligence (AESIA), headquartered in A Coruña, under statutes formally enacted by Royal Decree 729/2023. AESIA regularly issues technical and operational guidance on applying the AI Act.

The national legislative proposal intended to harmonize Spanish procedural law and establish domestic sanctions (Proyecto de Ley Orgánica para el buen uso y la gobernanza de la inteligencia artificial, parliamentary file 121/000096) was introduced in Congress on 28 May 2026. The bill officially lapsed on 6 October 2026, when Royal Decree 806/2026 dissolved the Cortes Generales and called a general election for 29 November 2026. Under Spanish constitutional procedure, lapsed legislative proposals must be formally reintroduced by the incoming government.

Consequently, as of 23 October 2026, there is no specialized Spanish domestic AI statute in force. The EU AI Act's substantive provisions apply in Spain by direct effect, with domestic sanctioning mechanics to be formalized following the constitution of the new legislature.

7.A practical diligence checklist for AI deployers

For every automated conversational assistant, chatbot, or agent facing customers, partners, or internal staff:

1. 1.Maintain a comprehensive asset registry. Catalog business owner, underlying model vendor, operational channels, action capabilities (read-only versus transactional), and all connected knowledge sources.
2. 2.Formally classify your statutory role. Determine whether your business qualifies as a Provider or Deployer under the AI Act, and verify whether any workflow touches Annex III classifications.
3. 3.Enforce clear, upfront AI disclosure. Deploy unambiguous transparency notices upon initial contact: a prominent header banner, an introductory conversational prompt, or an automated voice disclosure (Article 50).
4. 4.Document rigorous operational boundaries. Define permitted query topics, mandatory refusal rules, and automated escalation triggers leading directly to human personnel.
5. 5.Run benchmark evaluations against official documentation. Test the system with realistic questions spanning pricing, coverage, deadlines, and statutory rights, with expected ground truth verified by relevant policy owners. Archive all test runs.
6. 6.Execute continuous regression testing across changes. Re-evaluate performance after updating models, system prompts, retrieval corpora, or commercial terms to ensure no silent behavioral drift.
7. 7.Continuously monitor real-world production runs. Inspect statistical samples of live user interactions, retain structured logs, and systematically track customer disputes and hallucinations.
8. 8.Designate accountable oversight personnel. Assign human operators with documented competence and clear executive authority to correct, pause, or decommission the assistant.
9. 9.Audit upstream vendor contracts. Ensure contractual guarantees regarding technical documentation, audit logs, security incident notifications, bug correction windows, and liability allocation.
10. 10.Preserve an unbroken evidentiary audit trail. Maintain dated evaluation benchmarks, retrieval snapshots, error analyses, and corrective measures ready for presentation to regulators, judicial courts, or commercial insurers.

8.Establishing credible evidence of reasonable care

No European statute provides a simple checkbox that guarantees conversational correctness. What courts, arbitrators, and regulators evaluate is whether an organization took reasonable care: measuring error rates against ground truth, correcting detected failures, and maintaining active oversight. That evidentiary benchmark decided Moffatt v. Air Canada, and defines the discovery records courts can compel under the revised Product Liability Directive.

• Systematic empirical testing: Benchmark sets built from official internal documentation, pairing every test question with an authoritative source excerpt and run before launch and after every iteration.
• Clear scope documentation: Precise written definitions of the assistant's intended purpose, operational constraints, and version-controlled knowledge bases.
• Longitudinal quality monitoring: Documented tracking of accuracy over time, ensuring model revisions or content ingestions do not degrade performance.
• Demonstrated human oversight: Named individuals with authority to review failures, implement fixes, and pause the system when anomalous behavior emerges.
• Comprehensive audit logging: Immutable records of inputs, retrieved context, and generated outputs, enabling rapid forensic reconstruction of disputed interactions.

A Caudals evaluation provides this exact evidentiary record. We extract test questions directly from your verified documentation and real customer queries, collaborate with domain specialists to validate ground truth, evaluate every response semantically against cited source text, and deliver scored audit reports detailing failure types and remediation roadmaps. Monthly evaluation runs introduce new edge cases and track longitudinal regressions. Every result maintains full lineage back to its question, source passage, human reviewer, test execution, and score. Review our evaluation methodology, inspect a sample evaluation report, run a free website evaluation, or contact our evaluation team.

What we do not do: a Caudals evaluation report does not certify your AI system, does not constitute a conformity assessment, and does not make an organization compliant with the EU AI Act. It delivers verifiable empirical evidence that legal counsel and executive leadership can rely on. We never conduct adversarial testing, including prompt injection or jailbreaks, without explicit written authorization.

9.Frequently asked questions

10.Authoritative sources

Official legislative and judicial texts, verified as of the date indicated at the top of this guide.

1. 1.Regulation (EU) 2024/1689 (Artificial Intelligence Act)Official Journal of the European Union, 12 July 2024. Articles 3, 4, 5, 25, 26, 50, 99, 113 and Annex III.
2. 2.Regulation (EU) 2026/1744 (Digital Omnibus on AI)Of 8 July 2026, Official Journal 24 July 2026, in force 27 July 2026. New Article 4 and amendments to Articles 99, 111 and 113.
3. 3.Guidelines on the transparency obligations under Article 50 of the AI ActEuropean Commission, C(2026) 5054, 20 July 2026.
4. 4.Commission publishes guidelines on transparency obligations (IP/26/1653)European Commission press release, 20 July 2026: most of the AI Act, including the enforcement powers, applies from 2 August 2026.
5. 5.Directive (EU) 2024/2853 on liability for defective productsOf 23 October 2024, Official Journal 18 November 2024. Articles 2, 4, 6 to 10, 21 and 22; recitals 13 and 24.
6. 6.Withdrawal of Commission proposals (C/2025/5423)Official Journal of the European Union, 6 October 2025, including the AI Liability Directive proposal, COM(2022) 496.
7. 7.Directive 2005/29/EC on unfair commercial practicesArticles 6 and 7 on misleading actions and omissions.
8. 8.Moffatt v. Air Canada, 2024 BCCRT 149Civil Resolution Tribunal of British Columbia, 14 February 2024.
9. 9.Real Decreto Legislativo 1/2007 (TRLGDCU), Spanish consumer lawBoletín Oficial del Estado, consolidated text. Articles 60 and 61.
10. 10.Ley 3/1991 de Competencia Desleal (Unfair Competition Act)Boletín Oficial del Estado, consolidated text. Articles 5 and 7.
11. 11.Real Decreto 729/2023, statute of AESIABoletín Oficial del Estado, 2 September 2023.
12. 12.AESIA guides on the AI ActAgencia Española de Supervisión de la Inteligencia Artificial.
13. 13.Proyecto de Ley Orgánica para el buen uso y la gobernanza de la inteligencia artificial (121/000096)Congreso de los Diputados: presented 28 May 2026, lapsed 6 October 2026.
14. 14.Real Decreto 806/2026, dissolution of the Cortes and call of electionsBoletín Oficial del Estado, 6 October 2026.