Key Takeaways
- A generic ecommerce recommendation or dynamic-pricing engine is not high-risk under the EU AI Act by default. Under Article 6 and Annex III, the only high-risk retail-adjacent categories are creditworthiness/credit scoring (excluding fraud detection) and pricing in life and health insurance — your product recommender crosses the line only if it profiles natural persons or evaluates that kind of pricing.
- EU AI Act Article 50 transparency duties apply from 2 August 2026, and the EU AI Office's Code of Practice on Transparency of AI-Generated Content has a signatory deadline of 22 July 2026, 18:00 CEST.
- Non-compliance with Article 50 carries administrative fines of up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)(g)), with lower thresholds for SMEs.
- Air Canada was held liable for its website chatbot's wrong answer in Moffatt v. Air Canada, 2024 BCCRT 149 — a shopping assistant's mistake is the merchant's mistake, not the vendor's.
- The EU currently offers general AI incident-readiness exhortation (the Commission's 7 July 2026 Action Plan), not a mandatory AI-incident procedure — so the practical move is to borrow GDPR Article 33/34 breach-notification discipline as your operational default.
Introduction
Picture your storefront on an average Tuesday: a recommendation engine ranks the homepage, a dynamic-pricing tool nudges margins by customer segment, and a support chatbot answers a shopper at 2 a.m. Three third-party AI systems, three vendors — and if a regulator asked tomorrow, possibly no one who could say which of them is high-risk, what data each touches, or who is liable when one gets an answer wrong. This is the practical, ecommerce-specific companion to our pillar on [what AI governance is](/resources/blog/what-is-ai-governance): less about definitions, more about the four decisions that actually land on an operator's desk — classifying your recommendation and pricing engines, vetting AI vendors, meeting EU AI Act Article 50 at checkout, and having an incident plan before you need one. It is informational content, not legal advice.
Is My Recommendation Engine High-Risk Under the EU AI Act?
In almost all cases, no. Under EU AI Act Article 6 and Annex III, a generic ecommerce recommendation engine is not a listed high-risk system. It becomes high-risk only if it performs profiling of natural persons — which strips away the narrow-task exclusion — or crosses into creditworthiness or insurance-pricing evaluation.
This is the single most confused question in ecommerce AI, and practitioner sources answer it both ways, so it is worth resolving against the primary text. Article 6 sets two, and only two, routes into the high-risk tier. The first: the AI system is a safety component of, or is itself, a product covered by the Annex I EU harmonisation legislation that already requires third-party conformity assessment (think medical devices, machinery). The second: the system falls under one of the use cases enumerated in Annex III. A homepage recommender is neither a regulated product nor an Annex III use case.
Annex III is a closed list, and it is specific about retail-adjacent AI. It expressly names AI used to evaluate creditworthiness or establish credit scores (explicitly excluding fraud-detection systems) and AI used to evaluate and classify pricing in life and health insurance. It does not name generic product-recommendation engines or dynamic/personalized retail pricing. The direct read: a consumer-facing recommender or pricing tool sits outside the enumerated high-risk list — unless it either performs profiling of natural persons or wanders into credit or insurance-pricing territory.
The word doing the heavy lifting is profiling. Article 6 lets an Annex III system escape high-risk status when it only performs a narrow procedural task, improves the result of a completed human activity, detects decision patterns without replacing human judgment, or does preparatory work — but that exclusion is expressly unavailable if the system profiles natural persons. So the honest stance for an operator is this: "not high-risk" is a conclusion you document, not a status you assume. Write down the Annex III cross-check and the profiling test for each system; our [EU AI Act compliance guide](/resources/blog/eu-ai-act-compliance-guide-how-privacyforge-helps) walks that classification in detail. The trade-off is a page of paperwork now versus an unanswerable question during an audit later.
How Do I Audit My Dynamic Pricing for Bias?
Start from GDPR, not the AI Act. A dynamic-pricing bias audit for ecommerce checks two things: whether the pricing decision is solely automated with significant effect (GDPR Article 22 territory), and whether the model's inputs proxy for protected characteristics. Most retail pricing is not high-risk under the AI Act, but it can still trigger automated-decision rights.
GDPR Article 22(1) gives every data subject "the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her." A price that materially changes what one identifiable customer pays, set with no human in the loop, is exactly the kind of decision that provision was written for. Where such automation rests on contract necessity or explicit consent, Article 22(3) still requires safeguards — at minimum the right to obtain human intervention, to express a point of view, and to contest the decision. And Article 22(4) bars automated decisions built on special-category data unless narrow conditions apply. Our companion piece on [automated decision-making in ecommerce](/resources/blog/duaa-automated-decision-making-ecommerce) covers the fraud-scoring parallel, which follows the same logic.
A workable bias audit runs in four steps:
- Inventory the inputs. List every feature the pricing model consumes. Flag anything that could proxy for a protected characteristic — postcode (a race/income proxy), device type, browsing language, referral source.
- Test for disparate outcomes. Segment recent pricing decisions and compare the distribution of prices across those proxy segments. A gap you cannot explain by cost or genuine demand is a finding, not a footnote.
- Confirm the human route works. For any solely-automated price with significant effect, verify the Article 22(3) intervention and contest path exists as a real workflow, not a promise on a policy page.
- Document lawful basis and disclosure. Record why the automation is permitted and where the customer is told. New York's Algorithmic Pricing Disclosure Act, which took effect 10 November 2025, requires businesses using personalized algorithmic pricing to show disclosure text near the price — a signal of where transparency expectations are heading on both sides of the Atlantic.
How Do I Run an AI Vendor Risk Assessment?
Use NIST's four functions as a ready-made vendor questionnaire. The NIST AI Risk Management Framework (AI 100-1), released 26 January 2023, organizes AI risk around Govern, Map, Measure, Manage — a structure that converts cleanly into the questions you should put to any chatbot, recommendation-SaaS, or pricing-tool vendor before you sign.
This is where most small merchants have the least leverage and the most exposure: you can govern a model you built, but the AI baked into a rented SaaS tool is governed only if your vendor review forces it to be. The European Commission's Action Plan on Cybersecurity and Artificial Intelligence, presented 7 July 2026, leans in the same direction — noting that the AI Act "requires advanced AI models to be evaluated and their risks to be assessed before they are placed on the EU market" and that the Commission will help "strengthen third-party assessment of AI capabilities and risks." It imposes no new obligations, but it signals where scrutiny is heading.
Map the four functions to concrete asks:
| NIST function | What to ask the AI vendor | Evidence to demand |
|---|---|---|
| Govern | Who is accountable for the model? Is there a data processing agreement, and does it name AI Act deployer duties? | Signed DPA, sub-processor list, named contact |
| Map | What data does the system ingest, where does it run, and is customer data used to train shared models? | Data-flow description, hosting region, training opt-out |
| Measure | How is the model tested for accuracy and bias, and how are hallucinations bounded? | Test methodology, error rates, red-team summary |
| Manage | What is the incident and breach process, and how fast will they tell you when it fails? | Notification SLA, rollback plan, changelog for model updates |
For chatbots and shopping agents specifically, the generative-AI risks — fabricated answers, prompt injection, data leakage — are the ones that bite. NIST released a Generative AI Profile (NIST.AI.600-1) on 26 July 2024 precisely to catalogue those risks; treat its topics as the checklist for the "Measure" row above. The strong recommendation: a vendor that cannot tell you where its model runs or whether your customers' data trains it is answering the question by not answering it. Score, don't just collect — a questionnaire nobody grades is theater. The trade-off is time; the alternative is inheriting a vendor's model failure as your own liability, which the next section shows is not hypothetical.
What Does EU AI Act Article 50 Require at Checkout?
Article 50 imposes transparency duties, not risk-tier obligations, and three of them touch a storefront directly: disclose AI interactions, mark AI-generated content, and label deepfakes. These apply from 2 August 2026, the AI Act's general application date under Article 113 — regardless of whether any of your AI is high-risk.
Concretely, three mechanics matter at and around checkout. First, Article 50(1) requires that natural persons be informed they are interacting with an AI system, unless that is "obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect" — so a support chatbot generally needs a plain disclosure. Second, Article 50(2) requires providers of systems generating synthetic audio, image, video, or text to mark outputs "in a machine-readable format and detectable as artificially generated or manipulated," using solutions that are effective and robust "as far as this is technically feasible" — relevant if you auto-generate product descriptions or imagery. Third, deployers who publish deepfakes must disclose the content is artificially generated, with a narrow exception for "evidently artistic, creative, satirical, fictional" work. Our [Article 50 ecommerce transparency guide](/resources/blog/eu-ai-act-article-50-ecommerce-transparency) breaks the UI mechanics down further.
Two dates sit close together right now. The EU AI Office's Code of Practice on Transparency of AI-Generated Content has a signatory deadline of 22 July 2026, 18:00 CEST — ahead of the 2 August general application date. Eligibility to sign is limited to providers and deployers subject to Article 50(2) and/or 50(4), and per the official FAQ, signatories "may rely on the code to demonstrate compliance" with their Article 50(2), (3) and (5) obligations, subject to a positive Commission assessment — signing is evidence of good faith, not an automatic pass. The reason to care about the mechanics rather than the label: under Article 99(4)(g), an Article 50 breach can draw fines of up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (with lower ceilings for SMEs). That is a disclosure line of text weighed against a percentage of turnover.
Who Is Liable When Your AI Chatbot Gets It Wrong?
You are. When a customer-facing AI assistant gives wrong information, the merchant — not the AI vendor and not "the chatbot" — carries the liability. The reference precedent, widely reported across legal commentary, is a Canadian tribunal decision that says exactly this in plain terms.
In Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia's Civil Resolution Tribunal held Air Canada liable for negligent misrepresentation after its website chatbot gave a customer incorrect information about bereavement-fare eligibility. The airline argued the chatbot was, in effect, a separate legal entity responsible for its own words. The tribunal rejected that, holding the airline accountable for all the information on its website "whether it came from a static page or a chatbot." (This case is cited from secondary legal summaries — the primary tribunal text could not be retrieved for this article — so it is offered as a widely-reported decision with its citation, not a primary quote.) The transferable lesson is not about airlines: if your shopping agent quotes a return window, a discount, or a delivery promise your policy does not support, a court is unlikely to accept "the AI said it" as a defence. The governance response is oversight — bounded answer sets for anything that states policy, a human-review route for edge cases, and a log of what the assistant told whom.
What Should an AI Incident-Response Plan Cover?
At minimum: detection, containment, assessment, notification, and remediation — the same spine as a data-breach plan, applied to AI failures like a hallucinating chatbot, a discriminatory pricing output, or a leaked prompt. The honest picture is that the EU tells you to be ready without telling you exactly how, so you supply the procedure.
Be clear-eyed about what the law currently provides. The Commission's 7 July 2026 Action Plan states organizations "should intensify cyber hygiene practices, risk management measures, and security by design principles" and use AI to "fix vulnerabilities faster" — a general exhortation, expressly "not new legislation," rather than a granular mandatory AI-incident procedure. There is no AI equivalent of a prescriptive breach-notification clock yet. So the practical recommendation is to borrow one that already works: model your AI-incident protocol on GDPR Article 33/34 breach-notification discipline — a defined internal trigger, a documented assessment of who was affected and how, and a decision on whether the event also constitutes a personal-data breach requiring regulator and data-subject notification.
A default protocol an operator can adopt now:
- Detect and log. Route AI failures (wrong answers, biased outputs, unexpected model changes) into one intake channel, the way you route security reports.
- Contain. Have a rollback or kill-switch for each AI surface — a chatbot you cannot switch off is a chatbot you do not control.
- Assess overlap. Decide whether the incident is also a GDPR personal-data breach; if it is, the GDPR clock and duties apply on top.
- Notify. Tell affected customers and, where required, the regulator — plainly, without spin.
- Remediate and record. Fix the root cause, update the vendor questionnaire, and keep the record. This is exactly the transparency evidence the EDPB's 2026 Coordinated Enforcement Framework — launched 19 March 2026 with 25 national data protection authorities — is now asking controllers to produce.
How PrivacyForge Helps
AI governance for ecommerce is mostly a visibility-and-evidence problem, and both are tooling-shaped. PrivacyForge gives an operator one place to inventory every AI system — recommender, pricing engine, chatbot, and the vendor AI hiding inside your CRM and helpdesk — classify each against its EU AI Act tier and its GDPR Article 22 status, attach the vendor assessment and bias-audit findings, and hold the incident log and Article 50 disclosure evidence in a form you can hand to a regulator or an enterprise buyer.
The point is not a dashboard for its own sake; it is turning four scattered judgment calls — high-risk classification, vendor risk, checkout transparency, incident readiness — into a maintained, auditable record. If you want to see how the inventory, risk classification, and monitoring fit together, our walkthrough of the [AI Governance dashboard](/resources/blog/ai-governance-dashboard-managing-eu-ai-act-compliance) shows each part in practice. Governance you can produce on demand beats governance you can only describe.
Frequently Asked Questions
Is my ecommerce recommendation engine high-risk under the EU AI Act?
In almost all cases, no. Under Article 6 and Annex III, a generic product-recommendation engine is not on the high-risk list. Annex III names creditworthiness/credit scoring and life and health insurance pricing — not retail recommendations. It becomes high-risk only if it profiles natural persons or evaluates that kind of pricing.
Do I need to audit my dynamic pricing algorithm for bias?
If your pricing is solely automated and materially affects what an identifiable customer pays, yes — GDPR Article 22 applies, and a bias audit is how you meet it. Inventory the model's inputs for proxies of protected characteristics, test for unexplained disparate outcomes, confirm a human-review route exists, and document your lawful basis and disclosure.
How do I vet an AI chatbot vendor for GDPR and AI Act compliance?
Use NIST's four functions — Govern, Map, Measure, Manage — as a vendor questionnaire. Demand a signed data processing agreement, a data-flow description with hosting region, evidence of accuracy and bias testing, and an incident-notification SLA. A vendor that cannot say where its model runs or whether your customer data trains it has answered the question.
Am I liable if my AI shopping assistant gives a customer wrong information?
Generally yes. In Moffatt v. Air Canada (2024 BCCRT 149), a tribunal held the company liable for its chatbot's incorrect answer, rejecting the argument that the chatbot was a separate entity. The practical takeaway: bound your assistant's answers on anything policy-related, provide a human-review route, and log what it tells customers.
When do EU AI Act Article 50 transparency rules apply to my store?
Article 50 transparency duties apply from 2 August 2026, the AI Act's general application date. They require disclosing AI interactions such as chatbots, marking AI-generated content in a machine-readable way, and labeling deepfakes. The related Code of Practice on Transparency of AI-Generated Content has a signatory deadline of 22 July 2026, 18:00 CEST.
Conclusion
AI governance for ecommerce is not a policy you write once — it is the standing ability to answer four operator-level questions on demand: is this system high-risk, is this vendor trustworthy, is my checkout transparent, and do I have a plan when something breaks. The sequence is clear. Classify your recommendation and pricing engines against the real Article 6 and Annex III test rather than a rumor about them. Run every AI vendor through NIST's four functions before you sign. Get your Article 50 disclosures live ahead of 2 August 2026. And write your AI-incident protocol on the bones of GDPR Article 33/34, because the EU has not written one for you. Start with the artifact every part of this depends on — a complete inventory of the AI you already run. [See how PrivacyForge turns that inventory into an auditable AI governance record.](/resources/blog/ai-governance-dashboard-managing-eu-ai-act-compliance)
Sources
- [EU AI Act — Article 6 (classification of high-risk AI systems)](https://artificialintelligenceact.eu/article/6/)
- [EU AI Act — Annex III (high-risk use cases)](https://artificialintelligenceact.eu/annex/3/)
- [EU AI Act — Article 50 (transparency obligations)](https://artificialintelligenceact.eu/article/50/)
- [EU AI Act — Article 99 (penalties)](https://artificialintelligenceact.eu/article/99/)
- [GDPR — Article 22 (automated individual decision-making)](https://gdpr-info.eu/art-22-gdpr/)
- [EU AI Office — FAQ on signing the Code of Practice on Transparency of AI-Generated Content](https://digital-strategy.ec.europa.eu/en/faqs/signing-code-practice-transparency-ai-generated-content)
- [European Commission — Action Plan on Cybersecurity and Artificial Intelligence (7 July 2026)](https://commission.europa.eu/news-and-media/news/new-eu-plan-address-risks-and-opportunities-advanced-ai-cybersecurity-2026-07-07_en)
- [EDPB — 2026 Coordinated Enforcement Framework on transparency and information obligations](https://www.edpb.europa.eu/news/news/2026/cef-2026-edpb-launches-coordinated-enforcement-action-transparency-and-information_en)
- [NIST — AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework)
- [New York Attorney General — algorithmic pricing law consumer notice](https://ag.ny.gov/press-release/2025/attorney-general-james-warns-new-yorkers-about-algorithmic-pricing-new-law-takes)