Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

Dynamic pricing and recommenders: how the EU AI Act treats e-commerce algorithms

Plain product recommenders stay minimal risk under the EU AI Act. The real exposure sits in Article 5 manipulation, Article 50 transparency and Annex III credit scoring.

Renomira Team · 11 min read
Pharmacy shelf showing brand and generic products side by side with price tags
Photo: Daniel Case / CC BY-SA 3.0 via Wikimedia Commons

Dynamic pricing engines and product recommenders sit at the heart of modern online retail, which is why the AI Act for ecommerce personalisation has attracted anxious speculation. The AI Act dynamic pricing algorithm rules are narrower than the headlines: demand-based repricing is not banned, and a plain “customers also bought” rail is minimal risk with no duties of its own. Real exposure concentrates in three places — manipulative techniques caught by Article 5, disclosure duties under the transparency article, and the Annex III cases where a retailer becomes a credit scorer or a workplace monitor. This article maps each line and shows where personalisation stops and regulated AI begins.

The AI Act dynamic pricing algorithm rules, in plain English

The AI Act works on four rungs: unacceptable risk (banned outright under Article 5), high risk (permitted but heavily regulated), limited risk (transparency duties only) and minimal risk (no obligations at all). Almost everything e-commerce teams call “AI” sits on the bottom two rungs.

A repricing engine that adjusts prices from demand, stock levels and competitor feeds decides about goods, not about people. It appears in neither Annex III nor the transparency article, and a search-ranking model or warehouse-slotting optimiser sits alongside it. Treat these as minimal risk and document the rationale — the test an auditor applies is whether the system affects commercial decisions or natural persons’ rights.

Profiling shoppers does process personal data, so GDPR duties (lawful basis, transparency, the right to object to profiling) apply in full. That is a data-protection conversation rather than an AI Act one. The Act pulls a pricing or recommendation system upward only when it manipulates, conceals AI interaction or AI-generated content, or strays into Annex III.

Is a recommender system high risk under the AI Act?

Short answer: no, not on its own. The product recommendation AI Act question turns on what the system decides, not what it is called. The common “recommender system high risk ai act” worry assumes that any model ranking products for shoppers falls under Annex III — which covers eight domains (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice), and selling more jumpers is none of them.

The same technology earns a different classification when it points at people rather than products:

  • Recommending to shoppers — minimal risk. Ranking products, content or offers for a consumer has no Annex III hook.
  • Screening job applicants — high risk (Annex III point 4(a)). Valmara’s storefront and its careers site classify differently.
  • Monitoring warehouse pickers — high risk (Annex III point 4(b)), which covers AI that monitors and evaluates the performance and behaviour of workers.
  • Scoring buy-now-pay-later credit — high risk (Annex III point 5(b)), covering AI that evaluates the creditworthiness of natural persons or establishes credit scores, except where it only detects financial fraud.

Classify the use, not the algorithm. Two teams can deploy the same ranking library and land in different tiers.

Manipulative AI pricing and Article 5: where the ban actually bites

Article 5 is where pricing becomes unlawful outright, carrying the top penalty tier — up to €35 million or 7% of global turnover. The bar is deliberately high: it catches AI systems employing subliminal techniques beyond conscious control, or purposefully manipulative or deceptive techniques, that materially distort behaviour causing or likely to cause significant harm. A second limb covers systems exploiting vulnerabilities linked to age, disability or a specific social or economic situation.

The element most teams miss is significant harm. A checkout that pushes shoppers towards the pricier option, or a countdown timer pressuring a purchase, does not reach that threshold — however poor the tactic. The manipulation ai pricing article 5 concern becomes concrete where pricing exploits a shopper’s addiction, financial desperation or cognitive impairment: escalating high-cost credit offers aimed at people in problem debt, or nudging compulsive buyers into ruinous spend.

Loyalty pricing, dark patterns and the nudge manipulation ban

Loyalty pricing algorithm compliance eu retailers ask about comes down to one question: does showing returning customers a higher price break the AI Act? On its own, no. Article 5 targets technique and harm together, and a segment-based price is neither subliminal nor hidden.

The ai dark patterns nudge manipulation ban has teeth where interface design meets AI. If your system learns which users are most susceptible and works against them — fees revealed only at the final step, artificial scarcity on the fly, pre-ticked consent, a cancellation flow engineered to exhaust the user — that is purposefully manipulative territory. It becomes an Article 5 breach when significant harm follows; consumer law reaches it regardless.

A practical test for Valmara’s trading team: if a customer would describe the technique as a trick rather than an offer, treat it as one — first with consumer regulators and, if AI tailored the trick to that individual’s vulnerabilities, under the AI Act too.

Personalised pricing transparency: what EU law expects you to disclose

The personalised pricing transparency eu law conversation has two strands, and conflating them is the most common error in mid-market reviews.

The AI Act strand (Article 50). Two duties touch retail. AI systems intended to interact directly with people must make clear that the customer is talking to an AI — a shopping assistant, an AI sizing adviser, a generated product Q&A — unless it is obvious. And content produced or manipulated by AI, such as synthetic product photos, AI voiceovers or generated reviews, must be marked in a machine-readable format and detectable as artificial, with deepfakes shown to shoppers (an AI “customer testimonial”, say) disclosed as such. These duties are enforceable from 2 August 2026.

The consumer-law strand. EU consumer law imposes no general duty to tell a shopper their price was personalised. The silence is criticised, and the Commission has signalled a Digital Fairness Act to update the rules on dark patterns, personalisation and loyalty schemes. Until then, omitting personalisation is lawful — but risky, because the moment a customer cannot reproduce your price, fairness and transparency expectations under consumer law engage instead.

E-commerce AI use cases at a glance

The fastest classification exercise a trading team can run — one row per tool in the stack, then done:

Use case AI Act tier Why First action
Demand-based repricing Minimal risk Decides about goods, not people; no Annex III hook Record the rationale in the AI register
Product recommender Minimal risk Ranking products for a consumer is not an Annex III domain Register row; GDPR profiling notice
BNPL / credit scoring at checkout High risk Evaluates creditworthiness of natural persons — Annex III point 5(b) Provider conformity route; deployer oversight and notice
Review moderation / generated content Minimal to limited risk Disclosure duties attach where content is AI-generated or the shopper interacts with AI Machine-readable marking; Art. 50 copy
Workplace performance monitoring High risk Monitors and evaluates workers — Annex III point 4(b) Full deployer duty set; worker information
Careers-site applicant screening High risk Evaluates people for employment — Annex III point 4(a) Classify separately from the storefront
Manipulative personalisation exploiting vulnerabilities Prohibited Article 5 — subliminal or purposeful manipulation causing significant harm Stop the practice; document the stop decision

Classify the use, not the algorithm. The same ranking library lands in different tiers depending on whose behaviour it shapes, which is why one row per use case — not one row per model — is the unit of classification.

Beyond the AI Act: the neighbours your legal team already tracks

Three adjacent regimes catch e-commerce AI that the AI Act leaves alone, and conflating them is how mid-market reviews go wrong:

  • Unfair commercial practices and consumer law reaches deceptive design and hidden pricing personalisation regardless of AI — no AI Act hook needed, no AI Act defence either.
  • The Digital Services Act obliges platforms to explain how their recommender systems work in their terms — transparency about ranking, not classification.
  • The Digital Fairness Act package is expected to update rules on dark patterns, personalisation and loyalty schemes; treat Commission signals as a watch item, not a deadline.

The practical stance: run the AI Act classification for its own tiers, and keep the consumer/DSA checklist beside it — an e-commerce stack can be perfectly classified under Annex III and still lose a consumer-law argument.

AI Act for ecommerce personalisation: what to document now

For a mid-market retailer, the 2026 worklist is short and mostly clerical:

  1. Register rows with classification rationales for every pricing, recommendation and profiling model — one paragraph each, naming the tier and article considered.
  2. Article 50 disclosure copy for AI assistants and generated content: the interaction notice, machine-readable marking, and where each appears in the funnel — stored as evidence.
  3. The Article 5 screen for personalisation logic: does any technique exploit vulnerabilities, and would a shopper call the mechanic a trick? Keep the dated answer either way.
  4. Annex III escalation checks wherever the same data science serves HR, credit or worker monitoring — the three doors into the high-risk tier for retailers.
  5. Vendor questionnaires covering the engine’s provider status, documentation, and what happens to your register entry when the vendor ships a model update.

None of this requires a compliance department. It requires one owner — usually the Head of Legal or DPO — and a register that trading, engineering and legal all update rather than negotiate.

How to check your own position in three minutes

The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves mapped to the dates that apply. Confirm your email and the full report unlocks.

Frequently asked questions

Is dynamic pricing itself banned under the EU AI Act?

No — demand-based repricing decides about goods, not people: minimal risk, no duties of its own. Article 5 only catches pricing using subliminal or purposeful manipulation materially distorting behaviour with significant harm, such as targeting addiction or financial desperation. Ordinary segment pricing stays in consumer-law territory.

Are product recommenders high-risk under the AI Act?

Not for recommending products — selling more jumpers is not an Annex III domain. The same technology becomes high-risk when it points at people: screening job applicants (Annex III point 4(a)), monitoring workers (point 4(b)), or scoring buy-now-pay-later credit (point 5(b)). Classify the use case, not the algorithm.

What about AI-generated reviews and product photos?

Content produced or manipulated by AI — synthetic product images, AI-written review summaries, generated testimonials — must be marked in a machine-readable format and disclosed to shoppers where the deepfake rules apply. These duties are enforceable from August 2026, which makes the marking pipeline a 2026 worklist item rather than a 2027 one.

Where does our BNPL credit scoring land?

Annex III point 5(b): high-risk, because it evaluates the creditworthiness of natural persons — with the narrow exception for systems that only detect financial fraud. That flips on the provider duty set (conformity, documentation, registration) and, for the retailer deploying it, Article 26 deployer duties including human oversight, applicant notice and log retention.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

ai act dynamic pricing algorithm rulesrecommender system high risk ai actpersonalised pricing transparency eu lawai act for ecommerce personalisationproduct recommendation ai act obligationsmanipulative ai pricing article 5ai dark patterns nudge manipulation banloyalty pricing algorithm compliance eu

Keep reading