Claude at the distillery: faster safety data sheet reviews at Győri Szeszgyár

Project story · approx. 3 min read

With the help of Attention CRM Consulting, Győri Szeszgyár és Finomító Zrt. has introduced Anthropic’s Claude. The first use case is a safety data sheet (SDS) review: Claude analyses an SDS, cross-checks it against the relevant legal documents and gives an estimate of its compliance status. The team now saves 6–8 hours every week that it used to spend checking data sheets against the regulations by hand.

The client and the challenge

Győri Szeszgyár produces high-purity industrial alcohol, which pharmaceutical manufacturers, chemical companies and cosmetics producers use in their own products. It also sells bottled chemical products. These products are subject to strict regulations, so every product comes with a safety data sheet. Safety data sheets are tightly regulated documents, and mistakes here have real consequences.

Compliance checks are event-driven, and each one is carried out by comparing the safety data sheet with the relevant legislation in detail, section by section. The team was already using generative AI (ChatGPT), but it had no environment that worked from verified sources and stated the source behind every answer.

Foundation first

The project began with setting up the Claude Team organisation properly, because this is where the company decides who can use which Claude features, and under what controls. The aim is to move all eight users over from ChatGPT, each with named access and no shared accounts. Under Anthropic’s commercial terms, the client’s inputs and outputs are not used to train the models.

The redesigned review process

The second part is a Claude project built around a highly specialised skill. The project’s knowledge base does not contain product documents. It holds the regulatory baseline that every SDS is checked against, and only the knowledge base owner and the quality and regulatory owner can edit it. The product-specific documents are uploaded separately for each analysis, so every review is tied to a single product.

The skill packages a quality assurance expert’s methodology into an 18-step workflow. It audits all 16 sections of the data sheet, checks how they relate to one another (for example composition against classification, or classification against the label), and ranks its findings by severity. The core of the methodology is caution:

• Every regulatory conclusion cites the article or point of the legal act it relies on.
• Every piece of data is recorded with where it came from and how well it has been verified. The existing SDS is treated as evidence to be checked, never as the truth.
• Missing or unverifiable facts are flagged for expert input, not filled in.
• Weakening a hazard classification requires a verified source, while a stricter classification is flagged as a proposal.
• Every output is returned as a “draft pending review”.

The process works in both Hungarian and English.

Results

• 6–8 hours saved per week on reviewing and cross-checking safety data sheets
• 8 users being moved from ChatGPT to Claude Team, each with named access
• 2 working languages: Hungarian and English