AI in drinking-water utilities – high-risk classification, safety testing and manual fallback
Guide for water utility operators on classifying safety-component AI, reviewing vendor evidence, isolated tests and documenting manual fallback.
Short answer: Annex III point 2 of EU Regulation 2024/1689 (AI Act) lists AI intended to act as a safety component in the management and operation of critical infrastructure, including water supply. Such systems may require high-risk classification under Article 6. An isolated administrative report writer is not automatically high-risk merely because a water utility uses it. Following the AI Omnibus amendment, the special Annex III high-risk rules apply from 2 December 2027. Existing water quality, cybersecurity and operational responsibilities remain separate.
When is water utility AI a safety component?
Assess the system's intended purpose, real safety function and role in operational decisions, rather than its marketing label. Identify provider, deployer, water utility operator and decision-maker. Examine Article 6 and any precisely applicable exceptions; a risk classification is not permission to connect AI to live water treatment.
Decision table: functions and classifications
| AI function | Practical classification question |
|---|---|
| AI recommends actions but an operator reviews them | Assess real dependency: can the advice drive safety-critical decisions? Calling it a recommendation engine is not automatically an exemption. |
| AI is an intended safety barrier for water supply | Annex III point 2 is directly relevant. Identify safety function, interfaces, possible failures, supplier role and Article 6 assessment. |
| AI prepares administrative reports | Not automatically a point-2 safety component. Still review accuracy, personal data and sectoral obligations. |
| AI prioritises maintenance without direct control | Check whether its practical decision chain creates a safety-critical function even without a control interface. |
| A manual fallback is proposed | A fallback is a useful control but not proof of legal compliance, CE assessment or safe operation. |
Which laws and dates apply?
Timeline: AI Act provisions apply on different dates; Annex III high-risk rules apply from 2 December 2027 following the 2026 amendments. Article 27: the specific fundamental-rights impact assessment excludes Annex III point 2 systems; this does not remove other risk or GDPR impact assessments. Water rules: Sweden's LIVSFS 2022:12 governs drinking-water quality within its scope, with some testing requirements from 1 January 2026. Cybersecurity: operators covered by Sweden's Cybersäkerhetslag 2025:1506 must address security and continuity. The Swedish Food Agency oversees the drinking-water sector under that law.
A safe pre-deployment review process
- Map the AI function, the decision chain, actual environment and available human stop authority. Separate critical systems from ordinary reporting.
- Assess Annex III point 2 and Article 6, documenting the basis for high-risk or non-high-risk treatment. Identify provider and deployer.
- Request real vendor operating instructions, model versions, verified testing evidence and change-management responsibilities.
- Assess water-quality, continuity, cybersecurity and possible GDPR duties separately. AI Act is not a replacement for them.
- Use only authorised synthetic alarm and data-error cases in an isolated test environment, never connected to real dosing, pumps or valves.
- Have qualified water/process/OT staff evaluate independent barriers, human stop authority and manual fallback, and record a STOP or REVIEW decision.
Two practical classification cases
Case A – alarm interpretation: AI prioritises critical alerts. Assess the intended use, safety dependencies, independent protections and who can interrupt operations. Examine Annex III point 2 instead of automatically classifying all AI at the utility.
Case B – weekly report: An isolated language tool summarises non-sensitive operating statistics without making safety decisions. It is not automatically a safety component under point 2. Review confidentiality, factual reliability and human verification.
Documenting manual fallback
Record the authorised person who may stop use, the key AI dependencies, independent verification points and formal sign-off. Operational fallback instructions and control setpoints must be prepared and tested only by the utility's qualified water and OT specialists in approved systems. A generic guide must not give pump values, chemical-dosing instructions or PLC code. Keep only pseudonymised test IDs, results and evidence references.
Frequently asked questions
Is every AI tool used at a waterworks high-risk?
No. Annex III point 2 targets intended safety components in water supply. Article 6 must be assessed for the specific use.
Does Article 27 always require a FRIA at a water utility?
No. Its specific FRIA duty excludes Annex III point 2. Other impact and operational risk assessments may remain applicable.
Can this documentation authorise live control?
No. Only properly qualified personnel can validate safety controls and approve operational use following separate assessments.
Primary legal sources and limitations
ai · art6 · art26 · art27 · dates · water · cyber · supervision
Editorial legal check 13 November 2026. Based on the consolidated AI Act after the Digital Omnibus, Commission guidance, Swedish drinking-water rules and cybersecurity law. Not individual legal advice, an external legal review, technical certification or live operating instructions.
Related document templates
Get the editable Word, PDF and Excel package: AI in Water Treatment Plants 2027 – High-Risk Classification, Security Testing, and Manual Backup Operations | Word/PDF/Excel. Check the product description for its precise contents.
