How One AI Model Can Be a Shop Assistant or a Banned Workplace Tool

A route-based method for checking prohibited practices, high-risk systems, transparency duties, deadlines and…

The short version

EU AI Act classification requires checking prohibited uses, two high-risk routes, transparency duties and general-purpose AI rules.
Annex I and Annex III follow different tests, documentation requirements and high-risk deadlines in 2027 and 2028.
Route-based records reveal overlapping obligations and preserve evidence better than assigning each system one permanent risk label.

That tidy EU AI Act risk pyramid is wrecking sensible compliance work. Founders give each system one permanent sticker, file the slide and later discover the product triggered several duties at once.

I treat EU AI Act risk levels like airport security, except one officer asks why the chatbot can access biometric data. First, I screen the intended use for prohibited practices. Then I check both routes into high-risk status, transparency duties and separate rules for general-purpose AI. A system clearing those branches may have no risk-tier-specific obligations, although European law has hardly gone on holiday.

The calendar makes the pyramid even less useful. Prohibitions, definitions and AI-literacy provisions began applying on 2 February 2025; before then, they had not taken effect.

Article 50 transparency requirements and enforcement started on 2 August 2026, before the main high-risk obligations scheduled for later years. One label cannot show which obligation applies or when.

Misclassify a system and everything downstream goes sideways. The company collects irrelevant evidence, asks vendors useless questions and writes a human-oversight plan for a deployment that may be prohibited. Europe’s framework is defensible. The cereal-box diagram deserves retirement.

The first gate can end the conversation

I start with the intended purpose in painfully ordinary language. What does the system do? Which decision does it influence? Who lives with the result? Then I compare that description with prohibited practices, including every condition and exception. A plausible match ends normal classification and starts legal review. Better logs and a heroic human-oversight PDF cannot rehabilitate a prohibited use. Only systems clearing this gate reach high-risk analysis.

The same model can land in radically different places. I can use a language model to answer shipping questions on a furniture website, then connect it to biometric sensors to infer whether warehouse workers look anxious. The shop assistant may carry a transparency duty. The warehouse deployment can be prohibited because workplace emotion inference is banned, subject to medical and safety exceptions. Brussels cares what I do to people, even when the demo resembles a deleted Black Mirror scene.

That distinction prevents compliance cosplay. Emotion recognition under the Act concerns biometric data within its legal definition. A tool classifying customer-email sentiment may create privacy or employment problems, but a sad-face dashboard icon does not automatically make it prohibited. Legal definitions remain immune to product-marketing vibes.

Clearing the prohibited-practices gate is not an EU safety blessing. GDPR, employment law and product-safety rules may still apply. The Act also requires providers and deployers to take proportionate measures supporting AI literacy. I use “minimal risk” as conversational shorthand, never proof that Brussels inspected the product and awarded it a tiny gold star.

Transparency is another branch. Under Article 50, systems designed to interact directly with people must disclose the AI interaction unless it is already obvious to a reasonably informed person in context. Specified synthetic content needs marking, while covered deployers have their own disclosure duties. These can overlap with high-risk obligations: a recruitment system with a candidate-facing chatbot may activate both branches. Transparency rules have applied since the August enforcement date, while principal high-risk deadlines remain further away. Waiting for those later deadlines means being late already.

High-risk AI has two legal doors

The first high-risk route runs through regulated products in Annex I. The AI must be a regulated product or safety component, and the relevant product must require third-party conformity assessment under Union law. Article 6 requires both. An AI feature does not qualify merely because it shares a casing with regulated hardware; its product role and conformity route determine classification. An assistant controlling convenience features therefore differs legally from software whose failure could endanger health or safety. Model size barely matters.

The second route starts with intended uses in Annex III, including biometrics, education, employment, essential services, law enforcement and migration. I match the actual function to the relevant entry, not the category a vendor typed into Salesforce. A modest system influencing credit access can qualify because of the decision it supports. A far more capable general chatbot may fall outside this route if its intended use never touches an Annex III function. Capability matters for engineering; legal use determines high-risk classification.

Annex III offers a narrow exit beside a pile of paperwork. A system can fall outside high-risk status if it poses no significant risk of harm and performs one of the exception’s limited functions, including not materially influencing decision-making. Profiling natural persons closes that exit. A provider invoking the exception must document its assessment before market placement or use, then register the system. “A human clicks approve” would not convince me alone. Humans have an almost spiritual ability to accept whatever recommendation uses the largest font.

Here is the causal chain. Annex I requires a regulated product or safety component plus third-party conformity assessment. Annex III starts with a listed intended use, followed by the significant-harm test and limited exceptions. Profiling keeps the system high-risk. Claiming an exception still creates documentation and registration duties. Once high-risk, requirements arrive according to the legal route and implementation date. A published harmonised standard supports conformity only for requirements within its scope. That limitation matters because the standards machinery remains half-built.

I once added one “AI Act risk level” column to a vendor spreadsheet and felt organised for ten minutes. Then I found a system combining one company’s model, another’s interface and our workflow logic. One cell could not explain the intended use, our legal role or overlapping duties. I replaced it with a plain-language deployment description, legal route and role. The sheet became uglier and much more useful.

The deadlines moved because the machinery is unfinished

Standalone Annex III systems now reach their main high-risk obligations on 2 December 2027, replacing the original August deadline from the previous year. Providers gained time, but every old compliance deck became suspicious on sight.

AI embedded in regulated Annex I products follows on 2 August 2028, one year later than originally planned. Because the routes use different conformity machinery, their schedules diverge. Before trusting an old deck, identify the route. Yesterday’s seafood risotto lasts longer than some AI Act slides.

Standards explain part of the problem. As of October this year, only the harmonised standard covering quality-management systems had been published in the Official Journal. Providers using it can receive a presumption of conformity for Article 17 requirements within its scope, not the whole Act. The remaining high-risk standards have no confirmed publication dates or reliable completion timetable. A consultant can still make a beautiful Gantt chart. The chart is decorative.

The Annex III exception has another open flank: consistency. The legal test, documentation duty and registration requirement exist, but available sources contain no published cross-market measurement of how providers classify comparable tools. Similar systems could receive different treatment when one provider is especially eager to avoid the high-risk chapter. Enforcement and later court decisions will shape that boundary. Until then, I would preserve the intended-purpose record and all evidence supporting the decision. A final label tells the next reviewer almost nothing.

AI agents create a stranger unresolved question. Barry Scannell, a partner at William Fry, argues that development and testing before market placement fall outside the Act, including when an incident later affects the EU market. His reading deserves a fair hearing because market placement matters throughout the legislation. The competing view focuses on what happens after an agent escapes testing and affects Europe. No published Commission decision or court ruling has settled whether the Act applies, much less identified a definitive legal basis. I support assertive enforcement against foreign labs selling into Europe, but regulators should explain how the law reaches a testing incident before stretching unsettled provisions behind closed doors. Digital sovereignty needs credible institutions, built through public reasoning companies can challenge in court.

At the International AI Summit in Dublin on 14 October 2026, after enforcement began that August, European Commission Executive Vice-President Henna Virkkunen said the AI Office “has started enforcing these rules,” according to coverage by Veritome and Linklaters. Good. Europe needs an AI Office willing to act, and show its work.

Digital sovereignty lives in the evidence layer

Hosting location does not change classification. A recruitment scorer in Frankfurt remains a recruitment scorer, and an American model does not become European through server osmosis. I assess intended use, then the legal route. Infrastructure comes next because it determines whether my company can produce evidence, retain access after a contract ends and understand material model changes. European cloud providers may offer local support and contractual jurisdiction while reducing dependence on a few American platforms. Those advantages matter, but cannot wash a high-risk use into a lower category.

Procurement starts with roles and control. I identify the provider and deployer for the deployment, then ask who supplies documentation and controls logs. I check whether output markings survive our editing pipeline and demand exportable evidence before signing, because learning your audit trail belongs to a vendor is special founder cardio. Rebranding or substantially modifying a high-risk system, or changing its intended purpose, can shift provider duties under the Act’s downstream-change rules. A general tool repurposed for CV screening is the classic trap: procurement bought software; integration created a regulated use.

My pro-European bias gets practical here. I want European companies able to inspect systems and move evidence between suppliers. I want EU institutions coordinating infrastructure and standards at continental scale, because fragmented national strategies leave us renting critical technology from California or China. Europe has the market, engineering talent and regulatory leverage to build its own AI champions, yet dilutes those advantages through national procurement schemes with incompatible priorities. Eurosceptics call coordination a loss of control. In cloud and AI, fragmentation gives that control to somebody else.

Buying European still demands hard questions. Twelve gold stars on a sales deck are branding. Portable documentation gives customers leverage.

Open, machine-readable evidence could help European cloud providers compete across borders. A useful format would capture intended purpose and legal route, with the supporting assessment attached for another supplier or regulator to inspect. That reduces lock-in without pretending a generic compliance badge settles the law. Research proposals are moving this way, although positive vibes do not turn a preprint into an EU-approved method. The unpublished standards will decide how much portability survives contact with procurement departments.

By the end of 2028, I expect companies with route-based records to clear procurement faster than competitors still asking vendors for one “risk level.” The pyramid crowd will spend months governing harmless chatbots while consequential systems vanish into hiring and credit workflows.

Europe has put up the red lights. The winners will be the companies that keep the receipts.

Frequently asked questions

How do EU AI Act risk levels work?

The EU AI Act does not work as four permanent labels. Classification starts with prohibited practices, then checks the Annex I and Annex III high-risk routes, Article 50 transparency duties, and separate general-purpose AI rules. Several obligations can overlap, while systems outside a risk tier may still face other European laws.

What makes an AI system high-risk under the EU AI Act?

An AI system can be high-risk through two routes. Annex I requires it to be a regulated product or safety component whose product undergoes third-party conformity assessment. Annex III covers listed intended uses, subject to a narrow significant-harm exception; profiling closes that exception, and invoking it still requires documentation and registration.

Does hosting with a European cloud provider reduce an AI system’s risk level?

Hosting location does not alter EU AI Act classification. A system’s intended use and legal route determine its status. European cloud providers can improve local support, contractual jurisdiction, evidence access and portability, but those infrastructure advantages cannot turn a prohibited or high-risk deployment into a lower-risk one.