AI server for cosmetics manufacturers: product information file, claims, labels, formulas

What an AI server does for a cosmetics manufacturer

An AI server for a cosmetics manufacturer runs language and document models on the company's own formulas, raw-material files and test reports, inside its own building: it assembles the data for the product information file (PIF) and the safety report, checks marketing claims against the evidence in the file, and drafts and translates labels for each market. The safety assessor still writes and signs the assessment, and the responsible person still releases the product. A server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards handles this for a small or mid-sized maker.

Private-label and contract manufacturers can keep each brand client's formulas apart on the same server, and the INCI list itself comes from the formula software.

This guide is written for the owner, the head of regulatory affairs or the R&D manager of a cosmetics maker or a contract manufacturer without a large IT team. It starts from the work the EU rules require, explains the technical terms the first time they appear and keeps the hardware in one section. Everything legal here is general information on EU law as of October 2026; your safety assessor and legal counsel confirm what applies to your products.

The product information file in plain words

Regulation (EC) No 1223/2009 sets the frame for every cosmetic product in the EU. A product may be placed on the market only if a responsible person is designated within the EU (Article 4). That person ensures that a safety assessment is carried out before the product is placed on the market and that a cosmetic product safety report is set up in accordance with Annex I (Article 10(1)). The assessment itself is carried out by a person with qualifications in pharmacy, toxicology, medicine or a similar discipline (Article 10(2)).

Article 11 then requires the product information file. It contains a description of the product, the safety report, a description of the manufacturing method with a statement on good manufacturing practice, proof of the effects claimed where justified, and data on any animal testing. The file is kept for ten years following the date on which the last batch was placed on the market, readily accessible at the address on the label, in a language the competent authorities easily understand.

The safety report has two parts. Part A, the safety information, collects ten headings of data: composition, physical and chemical characteristics and stability, microbiological quality, impurities and packaging, normal and foreseeable use, exposure to the product and to the substances, the toxicological profile of the substances, undesirable effects and information on the product. Part B is the assessment: the conclusion, labelled warnings and instructions, the reasoning, and the assessor's credentials and approval.

The split decides the role of the AI. Part A is data gathered from many documents; Part B is a professional judgement signed by a qualified person.

Regulation (EC) No 1223/2009, Articles 10 and 11 and Annex I (consolidated text of 1 May 2025 on EUR-Lex, read 25 October 2026); role of the AI is our assessment.

In practice the assistant reads the supplier documents and test reports that arrive as PDFs and scans, files each result under its Part A heading with a link to the source page, and produces a list of what is missing for the assessor before the assessment starts. A vision-language model, an AI model that reads page images as well as text, does the reading. It does not invent a value: an empty field stays empty, and the assessor sees where every number came from.

Cosmetic claims substantiation: the evidence behind each claim

Article 20 forbids text, names, pictures and other signs in the labelling, making available and advertising of a product that imply characteristics or functions it does not have. Commission Regulation (EU) No 655/2013 adds six common criteria that every claim must meet: legal compliance, truthfulness, evidential support, honesty, fairness and informed decision-making. Under its Article 2 the responsible person ensures that the wording of a claim complies with them and is in line with the documentation in the PIF. On evidence, the regulation says it "shall take into account state of the art practices".

This is a comparison task a language model does well. The assistant reads the marketing text, the packaging copy or a product page, lists each claim ("hydrates for 48 hours", "suitable for sensitive skin", "dermatologically tested") and shows the study or test in the file that supports it, with the study's conditions next to the wording. Claims without matching evidence are flagged for the regulatory team. The judgement whether the evidence is sufficient stays with your experts; the assistant's list puts every claim it found in front of them.

We build AI servers for this work and our Private AI/ML service sets up the document reading and the search over your files; send us the number of products, the languages you label in and the documents behind a typical PIF, and our team replies with a configuration.

INCI lists and labels in many languages

Article 19 lists what the packaging must show, among them the name and address of the responsible person, the nominal content, the date of minimum durability or the period after opening, precautions, the batch number, the function and the list of ingredients. The ingredient list follows strict rules: descending order of weight at the time of adding, ingredients under 1 % in any order after the others, perfume and aromatic compositions as "parfum" or "aroma", and substances that Annex III requires to be named listed in addition. The names come from the Commission's glossary of common ingredient names (Articles 19(6) and 33), which is the INCI-based nomenclature used on labels.

Two parts of that work need two different tools. The ingredient list is a sorting job on the formula, best done by the formula software from the exact shares, so a model never writes it. The other label texts, such as precautions and instructions, depend on language rules: under Article 19(5) the language of several particulars is set by the law of the Member State where the product reaches the end user. A translation model drafts them for every market; EuroLLM-9B-Instruct, published under Apache 2.0, lists all 24 official EU languages and takes 17.05 GiB, a fraction of one card. A native speaker checks the first labels per language. Our guide to packaging artwork checks with AI before print covers the last step, comparing the print PDF with the approved text.

The allergen relabelling wave. Commission Regulation (EU) 2023/1545 amended Annex III for fragrance allergens. Its recitals note that 24 fragrance allergens are currently named individually on labels, that the Commission's scientific committee identified 56 more, and that labelling applies above 0.001 % in leave-on and 0.01 % in rinse-off products. Products that do not meet the new rules may be placed on the market until 31 July 2026 and made available until 31 July 2028. For a maker with hundreds of products in many languages, every affected product needs the allergen data from the fragrance supplier, a new ingredient list and new artwork. The assistant keeps track of which products are affected, which supplier data are still missing and which language versions are updated; the fragrance supplier's side is described in our guide to an AI server for fragrance and flavour houses.

Private label and contract manufacturing: client formulas kept confidential

A contract manufacturer holds the formulas of its brand clients, often under confidentiality terms, and its own development work beside them. Article 4 assigns the responsible person: for a product made in the EU it is the manufacturer, unless it appoints another person by written mandate, and a distributor that places a product on the market under its own name or trademark becomes the responsible person itself. Who keeps the PIF and who receives which data therefore follows the contract, and so does the question whether a client's formula may be processed by an outside AI service.

A server in your own building answers that question technically. Each client's formulas, files and drafts sit in their own area, with access rights per client and per role, and the query log shows who asked what. Our article on trade secrets, NDAs and AI tools explains the contract side; your legal counsel confirms the terms with each client.

Server size for PIF, claims and label work

A cosmetics maker's AI work is mostly reading and drafting documents. Qwen3.8-27B is an open vision-language model under Apache 2.0; its FP8 checkpoint takes 28.75 GiB. Our users-per-card rule takes 90 per cent of the 95.6 GiB the driver reports on a 96 GB card, less 3 GiB, less the weights, which leaves 54.3 GiB on one RTX PRO 6000, Max-Q or Server Edition. With an FP8 cache a conversation of 32,768 tokens (a token is a word or part of a word) takes 1.14 GiB, so one card serves 47 such conversations at once. For document reading, our guide to vision-language models for document AI estimates about 0.24 pages a second per RTX PRO 6000 Server Edition with this model. As an example with values to replace by your own, 300 products with 40 pages of supplier and test documents each make 12,000 pages, about 14 hours of one card.

Qwen3.8-27B-FP8 30.9 GB = 28.75 GiB and EuroLLM-9B-Instruct 18.3 GB = 17.05 GiB (Hugging Face files pages, read 25 October 2026); 1.14 GiB FP8 cache per 32K conversation; page rate by our estimate from our document AI guide; power by our estimate (cards, processor or processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

Two RTX PRO 6000 Max-Q in a workstation. The NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition is a 300 W card with a blower fan; NVIDIA allows up to four per workstation. A workstation with two of them suits a regulatory team that starts with the PIF work and has a side room with a door.

Two RTX PRO 6000 Server Edition in a rack server. This is the entry configuration on our AI servers page: two Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. The cards are cooled by the server's fans and suited to continuous operation. It is the machine we would size first for a maker that wants the PIF work, claims checks and label translation for the whole company.

Four RTX PRO 6000 Server Edition. Four cards suit a maker that also wants an assistant for sales, quality and production on the same machine, or a contract manufacturer with many clients and a large backlog of documents. With the cards at 600 W it draws about 3.4 kW and needs a 32 A or three-phase feed, as our guide to GPU server power on 16 A, 32 A and three-phase feeds explains; capped at 450 W it draws about 2.7 kW. Formulation work on your lab data, such as predicting stability or viscosity, can run on the same server, as our guide to formulation R&D with AI on your own lab data describes.

Practically all the power becomes heat, so the room needs cooling that runs all year. A company without a server room can have the platform run by our Private AI/ML service on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU. A company without IT staff will find the division of tasks in our guide to an AI server without an IT department.

Notes for the IT team

1. Systems of record. The formula software and the document management system stay the source. The assistant reads them with read-only rights and writes drafts and gap lists into a review queue; vLLM's structured outputs let it answer in a JSON schema that an import can validate.
2. Separation per client. For contract manufacturers, index each client's documents separately and pass the user's identity to the search, so the model only sees documents that user may open.
3. Engine and network. Run a current vLLM release (0.32.0 of 18 October 2026 at the time of writing) behind a gateway with keys and a query log. Its --api-key protects only paths under /v1, /v2, /inference and /cohere, so keep the engine on an internal network.
4. Versions and records. Pin model revisions in safetensors format, record the model version in every draft, and keep the NVIDIA driver at a release listed as fixed in NVIDIA's bulletin of 30 September 2026.
5. Software on delivery. On request the server arrives with the operating system, the NVIDIA driver, CUDA and a container runtime installed.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build workstations with two RTX PRO 6000 Max-Q and AI servers with two or four RTX PRO 6000 Server Edition cards to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the document reading, the search over your files and the connection to your formula software.

FAQ