What Leal actually said
Leal works across retail, food and automotive, and she has also advised the OECD. Her argument, as reported by Quantum Zeitgeist, is that trust and governance hold back sustainable AI adoption in retail. Shops and chains switch tools on before they have IT governance, internal controls and vetted data quality in place.
None of that is exotic. A stock-forecasting model fed by a product master file with duplicate SKUs and stale supplier lead times will produce confident nonsense. The model is rarely the failure point. The data pipeline and the absence of anyone accountable for it usually are.
Her second thread is cybersecurity. Leal warns that the combination of retail AI and advancing quantum computing threatens the encryption and password security that retail systems depend on today, through intrusions that stay quiet in a network for a long time. Her third is people: she puts the refresh cycle for workforce skills at 15 to 18 months.
One caveat before the rest. These are adviser remarks in an interview format, not a published study with a methodology section. Treat them as the judgement of someone with 15 years inside Brussels advisory work, not as measured results.
Where retail AI meets the AI Act
The useful part for a European retailer is that most of what Leal describes already has a legal home. The classification below follows the regulation's own annexes and articles; the dates reflect the current state of enforcement, including the mid-2026 omnibus that moved Annex III high-risk duties to 2 December 2027.
| Retail use case | AI Act treatment | Applies from |
|---|---|---|
| Customer-facing chatbot or voice assistant | Article 50(1) duty to tell users they are talking to a machine | 2 August 2026 |
| Synthetic product images, virtual try-on clips, AI-written ad copy | Article 50(2) machine-readable marking; Article 50(4) deepfake disclosure | 2 August 2026 |
| Emotion recognition aimed at customers in store | Annex III point 1(c), high-risk; plus Article 50(3) notice to exposed persons | Notice now; high-risk duties 2 December 2027 |
| Emotion recognition on shop-floor staff | Prohibited under Article 5(1)(f) | Since 2 February 2025 |
| Credit scoring for buy-now-pay-later | Annex III point 5(b), high-risk | 2 December 2027 |
| Demand forecasting, planogram optimisation, plain recommenders | Not listed as high-risk; GDPR and consumer law still apply | GDPR already in force |
Two things in that table catch retailers out. First, the Article 50 duties are not future work. They are live, and national market surveillance authorities can act on them now. Second, the deferral to December 2027 buys time to build a quality management system, technical documentation and human oversight for high-risk systems, not time to keep deploying them unclassified.
Penalties follow the usual ladder. Prohibited practices sit at the top with fines up to €35 million or 7% of worldwide annual turnover, while breaches of other obligations, transparency included, go up to €15 million or 3%. The Commission's AI Act pages carry the current text and timeline.
The 15 to 18 month skills estimate
Leal's figure is the one most European operations managers should sit with. If a skill refresh cycle runs 15 to 18 months, an annual training plan misses by half a cycle. That matters more now that agentic workflows are moving from demo to production: a merchandiser who learned to prompt a chat window in early 2026 is dealing with a different job in late 2026, one where the assistant executes multi-step tasks against a product database and needs checking rather than phrasing.
She frames this as change management and continuous reskilling rather than one-off courses. From a cost side, it is cheaper than the alternative. A team that cannot audit what an agent did last Tuesday is a governance problem in Leal's sense, not a training problem.
Quantum computing and store encryption
The quantum warning is often told badly, so here is the mechanism without the drama. No publicly known quantum computer can break current public-key encryption. The risk model is retrospective: traffic and database dumps captured today can be stored and decrypted later, once the hardware exists. That is why the migration starts before the threat arrives.
NIST published the first post-quantum standards in August 2024: FIPS 203, FIPS 204 and FIPS 205, covering key encapsulation and digital signatures. Retail relevance is narrower than the headlines suggest but real. Customer records, payment tokens, supplier APIs and loyalty databases are the assets with a long confidentiality life. Loyalty data collected in 2026 may still be sensitive in 2036.
The practical first step is dull: a cryptographic inventory. Where does data travel, which libraries terminate TLS, which internal services still use RSA-2048, which third-party payment provider has published a migration roadmap? You cannot migrate what you have not listed.
The arithmetic of a fine against inference
Model spend is not where the money goes. Running an open-weight European model such as Mistral Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens, roughly €0.65 and €1.94 at about $1.08 per euro. A mid-sized retailer pushing 200 million input and 50 million output tokens a day pays about $245 daily, near €227, or some €6,800 a month.
Set that against the €15 million ceiling for transparency breaches. It equals roughly 2,200 months of that inference bill, or about 180 years. The point is not that fines are typical, they are not, and authorities set actual amounts case by case. The point is that the governance layer is the expensive part of a retail AI programme, and the token bill is the cheap part. For retailers who want to keep inference in-house and in the EU, open weights make that possible, and we benchmark European open models on local hardware in AI Arena.
What a retailer can check this quarter
- Write down every AI system in production, including the ones inside SaaS tools, with an owner name next to each.
- Classify each against Annex III. Most will fall outside it, and the ones that do not need documentation now, not in 2027.
- Audit Article 50 disclosures: chatbot notices, synthetic image marking, any deepfake-style creative in paid media.
- Check emotion recognition anywhere in the estate, including pilot cameras in staff areas, against Article 5.
- Start the cryptographic inventory and ask payment and loyalty vendors for their post-quantum roadmap in writing.
- Move reskilling from an annual event to a rolling cycle shorter than 18 months.
Leal's guidance arrives without published data behind the 15 to 18 month figure, and the interview format leaves the reskilling and quantum claims at the level of expert judgement. The regulatory dates, by contrast, are written down, and the earliest of them has already passed.
Does the AI Act apply to a small independent retailer?
It depends on the role. If you only buy and use an AI tool, you are a deployer, and the transparency duties in Article 50 still reach you: customers must know when they are talking to a machine, and synthetic content must be marked. The heavier high-risk obligations apply when the system falls into an Annex III category, such as credit scoring for a payment plan.
Is the December 2027 date a delay for everything?
No. It covers the high-risk obligations for Annex III systems. Prohibited practices have been enforceable since February 2025, general-purpose AI duties since August 2025, and Article 50 transparency, including emotion recognition notice, since 2 August 2026.
Should a retailer buy quantum-safe encryption right now?
Buying is premature for most. Knowing where your encryption lives is not. An inventory of algorithms, libraries and third-party data flows is the step that makes a later migration possible, and it is the same inventory most GDPR and security audits already ask for.