A language model like ChatGPT produces the first draft of a risk assessment in minutes: a structure by hazard class, the relevant statutory provisions, and an initial set of protective measures for the business described. That draft is audit-proof against the trade association only once it has been compared with an on-site walkthrough of the actual workplaces, revised for the specific business, and signed off by an appointed occupational safety specialist. This article explains what a good prompt looks like, which five review steps follow, and what AI structurally cannot do in a risk assessment.
This article is a detail page to AI in occupational safety, which covers the general use of language models in the workplace.
What the trade association checks in a risk assessment
The trade association does not check whether a document looks formally complete during inspections. It checks whether the risk assessment describes the specific business: the actual activities, the actual rooms, the actual equipment in use. A well-formatted generic AI text that uses only "office business with screen workstations" as context regularly fails this comparison.
The risk assessment is the systematic identification and evaluation of all work-related hazards at a workplace required by the employer under § 5 ArbSchG, with the aim of establishing suitable protective measures. It is mandatory from the first employee. It must be documented in writing under § 6 ArbSchG as soon as more than ten employees work in the business.
Inspectors from the trade association have the right to enter business premises and inspect documents under § 18 ArbSchG. Anyone who cannot produce a risk assessment during an inspection risks a written enforcement notice. Anyone who produces a generic template without recognisable business-specific content risks the same consequence.
| Inspection point | What is expected | Typical error in AI drafts |
|---|---|---|
| Business specificity | Risk assessment describes actual activities and rooms | Only standard hazard classes without business context |
| Currency | Date after last material change to workplaces or equipment | No date or outdated creation date |
| Measures | Specific measures with responsible persons and deadlines | General recommendations without assignment |
| Documentation | Written under § 6 ArbSchG, sign-off traceable | No evidence of specialist sign-off |
| Effectiveness review | Evidence that measures have been implemented and reviewed | Missing from almost every AI draft |
Where a business has multiple sites, a separate risk assessment or at least a site-specific supplement is required for each. A generically produced draft does not cover this.
The prompt: how an AI draft gets close to the actual business
The quality of an AI draft depends directly on the prompt. Four inputs are mandatory for the model to produce a business-specific draft: business type and sector, headcount, a list of main activities, and an explicit request for statutory references from ArbSchG, ArbStättV and relevant DGUV rules.
A weak prompt produces a generic draft with no practical value:
"Create a risk assessment for an office."
A usable prompt starts with the specific business:
"Create a risk assessment for a tax advisory firm with 18 employees, two floors of office space, a reception area, an archive room and two meeting rooms. Main activities: screen work, client meetings and archive work. Name the relevant paragraphs from ArbSchG, ArbStättV and the DGUV rules for each hazard area, and assign specific protective measures to each."
Check the resulting draft immediately on three points: Are the cited paragraphs real and current? Are there activities the business actually has that are missing? Does the draft contain hazard classes that do not apply to this business, such as handling hazardous substances in a pure office?
What language models reliably do in a risk assessment:
- Structure by hazard class (physical, ergonomic, psychological, fire, electrical, biological)
- Formulations that reflect statutory logic and can be cited
- Checklists for a structured walkthrough
- Protective measure proposals from standard regulatory frameworks
What language models do not do in a risk assessment:
- Take noise measurements, illuminance measurements and room temperature readings
- Assess whether an escape route is actually clear and wide enough
- Estimate psychosocial stress specific to the actual workplace
- Confirm that electrical equipment has been tested under DGUV Regulation 3
Five review steps after the AI draft
Between the AI draft and a legally sound risk assessment lie at least five steps. Anyone who skips them has a text, but not a risk assessment within the meaning of § 5 ArbSchG.
- Plausibility check. Look up every cited paragraph on gesetze-im-internet.de and check whether it applies to this business type. Language models occasionally invent regulations or cite outdated versions. This step takes ten minutes and prevents the most serious mistake: a risk assessment that cites provisions that do not exist.
- Walkthrough. Compare the revised draft room by room with the actual workplaces. Check: screen workstations under ASR A3.5 and ASR A3.4, escape and rescue routes under ASR A2.3, first-aid equipment under DGUV Regulation 1 and the date of the last DGUV V3 inspection. Add every hazard missing from the draft. Remove every hazard in the draft that does not exist.
- Measurements. Illuminance (under ASR A3.4 at least 500 lux at screen workstations) and room temperature (under ASR A3.5 a maximum of 26 degrees when outdoor temperatures exceed 26 degrees) can be taken with simple instruments. Without measurements, the assessment of physical hazards remains an estimate with no evidentiary weight.
- Review by the specialist. The appointed occupational safety specialist reviews the revised draft, supplements it from their technical expertise under § 7 ASiG, and signs it off in writing. Without this sign-off, there is no evidence of professional review.
- Documentation under § 6 ArbSchG. The final risk assessment is dated, signed by management or an authorised person, and filed so it can be produced immediately during inspections. Electronic storage is permitted provided the time of sign-off is traceable.
The office occupational safety checklist tool helps cover all areas of an office risk assessment systematically and miss no inspection point.
What AI cannot replace in a risk assessment
A language model has no body, no senses and no liability. These three factors confine its role in a risk assessment to drafting and formulation.
Local knowledge is not substitutable. A risk assessment under § 5 ArbSchG must assess the working conditions of the specific business, not of typical office businesses. Whether an escape route on the third floor is actually wide enough and not blocked by boxes, whether an archive room is adequately ventilated, and whether screens are positioned according to DIN EN ISO 9241-303 can only be known by someone who was there.
Specialist expertise is not substitutable. Assessing whether a measure meets the state of the art requires the technical expertise under § 7 ASiG. A language model can produce plausible formulations but cannot assess whether the measure described is actually effective in this specific business.
Accountability is not substitutable. Liability for the risk assessment rests with the employer under § 5 ArbSchG. Under § 13 ArbSchG, duties may be delegated to reliable and qualified persons, not to software. An error in an AI draft adopted without review is legally the company's fault.
AI literacy as a statutory duty. The EU AI Act has required operators of AI systems to take measures to ensure an adequate level of AI literacy among their staff since 2 February 2025 under Art. 4. Businesses using AI for the risk assessment should define internally who uses it for what and how results are systematically reviewed.
Audit-proof risk assessments with Safety Club
Safety Club provides the external occupational safety specialist who creates the risk assessment on-site, reviews it, embeds it in the business-specific documentation and keeps it current after every safety inspection. AI tools can speed up the preliminary work and structure walkthrough preparation; they do not replace the appointment and expertise of the specialist. 300+ companies supported in the partner network.
From first contact to signature: five working days at most. Start of on-site support: within ten working days.
