EU AI Act Compliance for the Chemicals Industry: Process Control, R&D, and the Lines That Actually Matter

A chemicals company runs AI in two very different places — on the plant floor and in the lab — and the EU AI Act treats them very differently. Plant process-control AI can land in scope as a safety component of critical infrastructure or under sectoral product-safety law. R&D molecule-discovery AI is usually minimal-risk, and where it is purely research it may not be covered by the Act at all. The mistake is to treat "we use AI" as a single compliance question. It is at least three.

The verdict up front: most chemicals AI is minimal-risk; a defined slice of plant-safety AI is high-risk; research AI is often exempt or minimal; and — as in every industry — the workforce-facing AI quietly carries the heaviest predictable duties. Classify each system by its use, under Regulation (EU) 2024/1689.

Plant process-control AI: two ways it can become high-risk

Process-control AI — distributed control systems with AI-driven setpoint optimisation, advanced process control, batch-quality prediction, anomaly detection on reactors — is operational by default. It becomes high-risk only through one of two specific routes.

Route 1 — Annex III point 2: safety component of critical infrastructure. Annex III point 2 captures AI used as a safety component in the management and operation of critical infrastructure, including the supply of water, gas, heating, and electricity. A chemicals site can fall here when its AI manages a safety function tied to such supply — for example AI governing a combined-heat-and-power plant that feeds a district network, or a facility designated as critical infrastructure whose AI controls emergency containment or shutdown. The trigger is the safety-management function, not the hazard of the chemistry alone. Optimising yield or energy use is operational; commanding an emergency shutdown to prevent a release is a safety function.

Route 2 — Article 6(1) + Annex I: sectoral product/safety law. Where AI is a safety component of a product covered by EU harmonised legislation listed in Annex I — pressure equipment, ATEX equipment for explosive atmospheres, machinery under the Machinery Regulation (EU) 2023/1230, and similar — and that product requires third-party conformity assessment, the AI is high-risk via the Article 6(1) product route. For pressure and ATEX equipment (Annex I Section A), the AI Act requirements are integrated into the existing sectoral conformity assessment. Machinery has sat in Annex I Section B since the Digital Omnibus, Regulation (EU) 2026/1744, so the AI requirements reach it through delegated acts under the Machinery Regulation instead. Either way the date is 2 August 2028, not the Annex III date. The same Omnibus added Article 6(1a): AI used solely for performance optimisation, automation or quality control is not a safety component unless its failure would endanger health and safety.

So a single reactor-safety AI can reach high-risk by either path depending on the facts: through Annex III point 2 if the site is critical infrastructure, or through Article 6(1) if the AI is a safety component of regulated machinery or pressure equipment. Most reactor-optimisation and quality-prediction AI is neither — it is minimal-risk operational tooling.

R&D molecule-discovery AI: often minimal, sometimes exempt

AI that proposes candidate molecules, predicts reaction yields, screens compound libraries, or models toxicity in silico is, in almost every case, minimal-risk. It does not appear in Annex III, it is not a prohibited practice, and it is not a transparency case. There is no mandatory compliance stack for a generative chemistry model used to suggest synthesis routes.

There is also a research carve-out worth knowing. Article 2 of the Regulation excludes AI systems and models specifically developed and put into service for the sole purpose of scientific research and development, and excludes research, testing, and development activity prior to a system being placed on the market or put into service. A molecule-discovery model used purely in the lab, never commercialised as an AI product, can sit outside the Act's scope on this basis. The carve-out is narrow — the moment the model is placed on the market as a product, or used operationally rather than for R&D, the exemption falls away — but for genuine in-house research it is real relief. Note that the AI Act exemption does not displace other law: GDPR, REACH data obligations, and lab-safety rules still apply on their own terms.

The one that always bites: workforce AI (Annex III point 4)

Across every chemicals company, the most predictable high-risk exposure is not the reactor — it is HR. AI that screens applicants for process-operator or chemist roles, ranks candidates, drives promotion or dismissal decisions, or allocates shifts by automated scoring is high-risk under Annex III point 4. It is a deployer duty (Article 26) when bought from a vendor, a provider duty (Article 16) when built in-house, and it tracks the 2 December 2027 deadline (originally 2 August 2026).

And the prohibited line: AI inferring the emotional or stress state of plant workers — stress, frustration or engagement read from faces or voices — is prohibited emotion recognition in the workplace under Article 5(1)(f), in force since 2 February 2025, unless the system is intended for medical or safety reasons. That exception is narrow; the Commission's guidelines name dangerous chemicals as a setting where it can apply, so document the specific danger if you rely on it. Fatigue is a different question: recital 18 excludes physical states such as fatigue from emotion recognition, so a drowsiness alert for operators is not prohibited, though it can be high-risk under Annex III point 4(b) where it is used to monitor workers. A breach of the prohibition sits in the highest fine tier: €35,000,000 or 7% of total worldwide annual turnover, whichever is higher (Article 99(3)).

Classification table for a chemicals business

The high-risk dates are law: the Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and has been in force since 27 July 2026. Stand-alone Annex III high-risk systems apply from 2 December 2027 (originally 2 August 2026). Annex I product-route systems track 2 August 2028 (originally 2 August 2027).

Worked example: Rheinpolymer Chemie GmbH, ~430 employees, ~€150M turnover

Rheinpolymer makes specialty polymers at a single site near Ludwigshafen and runs a small computational-chemistry research group. Its compliance lead works through the estate.

• Advanced process control optimising reactor temperature and feed rates for yield, and batch-quality prediction flagging off-spec product: operational, minimal-risk. No obligations.
• The reactor emergency-shutdown AI that can trigger containment on a runaway-reaction signal: Rheinpolymer's site is not designated critical infrastructure and does not supply a public utility, so Annex III point 2 does not bite. But the shutdown function is a safety component of pressure equipment regulated under sectoral law (Annex I Section A) — so it reaches high-risk via the Article 6(1) + Annex I product route, with conformity integrated into the equipment's sectoral assessment and the 2 August 2028 deadline. The compliance lead documents this carefully, because the route determines the deadline.
• The in-house molecule-discovery model the research group uses to propose monomer candidates: used solely for R&D, never commercialised — covered by the Article 2 scientific-R&D exclusion. Out of scope under the AI Act (REACH and lab-safety duties still apply).
• The applicant-screening tool ranking chemist and operator candidates: high-risk, Annex III point 4. As deployer, Rheinpolymer owes the Article 26 duties; the vendor holds the technical documentation.
• A proposed control-room "alertness" camera that also inferred operator stress from facial expressions: the stress function was flagged as emotion recognition under Article 5(1)(f) with no documented safety case, so it was dropped before procurement. A drowsiness-only alert would not have been prohibited.

Verdict: one Annex I product-route high-risk system (2028 deadline), one Annex III point 4 HR system (2027), one prohibited pilot stopped, one research model out of scope, and the rest minimal. The deadlines differ by route — which is exactly why the route has to be pinned down per system.

How Confir helps

The hard part for a chemicals company is not knowing the obligations — it is correctly sorting each AI system into minimal, R&D-exempt, Annex III high-risk, or Annex I product-route high-risk, and then proving why. Confir registers every AI system, classifies each under Articles 5 and 6 with Annex III logic through plain-English checklists, applies the Article 2 R&D filter where relevant, and derives the role — provider (Article 16) or deployer (Article 26). The engine is deterministic and rule-based: the same intake yields the same classification and surfaces the rule that fired and the route it implies, so the difference between a 2 December 2027 Annex III system and a 2 August 2028 Annex I system is recorded, not guessed. For the high-risk systems, Confir structures the Article 11 / Annex IV technical documentation and the Article 47 Declaration of Conformity. No consultants, no multi-month implementation.

Related guides

• Article 6 high-risk classification
• Critical infrastructure safety AI under Annex III
• Article 2 scope and exclusions
• Recruitment and screening AI as high-risk
• Manufacturing under the EU AI Act
• Energy and utilities under the EU AI Act
• Pharma under the EU AI Act