AI server for testing labs: report drafts under ISO/IEC 17025 and confidential client data

What an AI server does for a testing laboratory

An AI server for a testing or calibration laboratory runs language models on the laboratory's own hardware to draft the text around its results: the narrative parts of test reports, answers from released methods and SOPs, readings of instrument printouts and replies to client enquiries. The numbers never come from the model: results, units, uncertainties and conformity decisions are taken from the LIMS or the instrument export, a script checks that every number in the draft matches them, and a person approves and releases the report. For most contract laboratories, a server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards carries this work, and client data stays on hardware the laboratory controls.

We wrote this for heads of testing, calibration and analytical laboratories and their quality managers, and for the IT team that runs the systems. The laboratory side comes first, the server and the notes for IT after it. Research and development departments that search their own test reports have different needs, covered in our guide to an AI server for product development lab reports; GxP work in pharma is in our guide to GPU servers for pharma and life sciences.

What ISO/IEC 17025 asks of a report

ISO/IEC 17025:2017 is the third edition of the standard for the competence of testing and calibration laboratories. ISO's page states that it sets out requirements "for the competence, impartiality, and consistent operation of laboratories" and that it was last reviewed and confirmed in 2023. Clause 7.8 covers reporting, and several of its requirements decide where a language model may help.

ISO/IEC 17025:2017 as quoted in the European co-operation for Accreditation's FAQ on requirements for test reports and in PJLA's webinar slides of 14 August 2024, and as summarised in NIST's crosswalk for clause 7.8 (February 2018), all read 25 October 2026; the last column is our recommendation.

The European co-operation for Accreditation quotes clause 7.8.2.1 f), which asks every report to identify "the method used" unless there are valid reasons not to, and 7.8.1.2, which asks for results provided "accurately, clearly, unambiguously and objectively". NIST's crosswalk summarises the rest of 7.8: reports are reviewed and released before issue; the laboratory is responsible for all information in the report except data supplied by the customer, which is identified as such; when a report states conformity with a specification, the laboratory documents and applies a decision rule; and an amendment after issue is marked as such and refers to the original.

These are the reasons for a clear split. Everything that is a result or a decision is produced by the systems your accreditation already covers. The model writes what is text: the summary of what was tested and why, the description of a deviation that the analyst has noted, the cover letter, a translation into the client's language, and, after a change, the reason for the amendment. Opinions and interpretations, where your scope includes them, are written or approved by the person responsible for them.

Numbers from the LIMS, words from the model

The design we recommend keeps the model away from the numbers in four steps.

1. Export. The LIMS exports the data of one report as a structured file: sample and item data, method and version, results with units and uncertainties, the conformity decision and the analyst's notes.
2. Template. The report template fills every fixed field, every table of results and every conformity statement directly from that file. The model is not involved.
3. Text. The model receives the file and writes the narrative parts. Where the text needs a value, it writes a placeholder that names the field, and the template inserts the value from the export.
4. Check. A script extracts every number from the finished draft and compares it with the export. A number that does not appear there, or a unit that does not match, stops the draft before it reaches the person who approves it.

Step 4 is the systematic check that clause 7.11.6 asks for, applied to the model's output. It also catches a figure the model has rounded or invented, because such a figure cannot pass. The person who approves the report then reads text whose numbers are known to be right, and spends the review on whether the text says the right thing.

For the IT team, the same split keeps the integration simple: the AI server reads from an export or a read-only interface and writes drafts into a staging area. The LIMS stays the record, and the release happens there. We supply the server and set up the model, the templates and the check through our Private AI/ML service; send us one anonymised report, your LIMS and the number of reports a day, and our team replies with a configuration.

The AI step inside your information management

Clause 7.11 of the standard deals with the control of data and information management. As quoted in PJLA's slides, it asks that the laboratory information management system is "validated for functionality, including the proper functioning of interfaces" before it is introduced, and that changes are documented and validated before they are implemented. It asks for protection against access without permission, safeguards against tampering and loss, and records of system failures with the actions taken. Where such a system is "managed and maintained off-site or through an external provider", the laboratory makes sure the provider meets the requirements too.

By our reading, a step that writes report text from LIMS data is part of that information management, and it is safer to validate it as such than to argue later that it is not. In practice that means four things:

• A pinned stack. The model file by its exact revision, the engine version and the prompt templates are recorded, and none of them changes without a change record.
• A regression set. Fifty to a hundred past reports with their approved texts form the test set. A new model or template runs against them before it goes live, as our guide to LLM evaluation before model upgrades describes.
• Logs. Every draft keeps the export it was built from, the model revision and the result of the number check, so an assessor can follow how a report came about.
• Access. Only staff with rights to a client's data can ask the model about it.

NIST's summary for laboratory directors notes that validation of laboratory spreadsheets takes significant time and must be redone when systems and software are upgraded. The same holds for a model, which is why a validated stack should change on your schedule, not on a cloud provider's. Your quality manager and your accreditation body confirm how they treat the AI step; this is general information.

Methods, SOPs and client enquiries

The second job is knowledge. Staff ask which method version applies to a sample type, what the SOP says about a deviation, or how a similar case was handled last year. An assistant with retrieval over your released methods, SOPs and validation reports answers with the document and revision it used, as our guide to RAG on company data explains. Only released versions go into the index, so your document control stays the record. The text of ISO standards and published test standards is licensed material: ISO's terms restrict loading it into AI systems without a separate licence, as our guide to standards in a company AI sets out, so the index holds your own procedures and only those standards your licence allows.

Client enquiries are the third job. A client asks which test fits a product, how much sample material to send, or what a result means. The model drafts the reply from your service catalogue and methods, and staff send it. Prices and turnaround commitments come from your own systems and people, never from the model.

Instrument printouts and client documents

Most instrument data should reach the LIMS by direct export. Printouts remain for older instruments, for subcontracted tests and for documents clients send: specifications, earlier certificates, drawings. These go through OCR, the conversion of page images into text, or a vision-language model that reads images directly, as our guide to vision-language models for document AI explains. By the method of that guide, one RTX PRO 6000 Server Edition converts about 0.69 pages a second with a 5B OCR model, so 2,000 pages take about 50 minutes of one card. Values read this way are customer-supplied or legacy data in the sense of clause 7.8.2.2, so the draft marks them, and a person checks them against the source before they enter the LIMS.

Client confidentiality and accreditation

A contract laboratory often works for companies that compete with each other, and its files hold their formulations, failure data and unreleased products. ISO/IEC 17025 treats confidentiality as a requirement of its own in clause 4.2. By our reading, a public AI service is an external provider in the sense of clause 7.11.4 once it handles report data, and many client contracts restrict passing data to third parties. A server in your own server room keeps the samples' data, the prompts and the logs under the access rules your quality system already defines, and separates clients by access rights inside the index. Health data from medical testing are a special category under the GDPR; your data-protection officer confirms what applies.

Accreditation in the EU is organised under Regulation (EC) No 765/2008, which says that "each Member State shall appoint a single national accreditation body". That body assesses your laboratory against the standard, so it is the one to ask how it views an AI step in reporting. Our article on AI questions in customer audits shows the evidence clients and assessors ask for.

Server options

The models need memory on the graphics card, the GPU, and each request in progress needs extra memory for its context, the cache. Our users-per-card rule takes 90 per cent of the 95.6 GiB that the driver reports for a 96 GB card, less 3 GiB per card, less the weights: 83.0 GiB per card.

Reference configurations on our AI servers page; capacities by our users-per-card rule with gpt-oss-120b (60.8 GiB, Hugging Face file list read 25 October 2026) and 1.125 GiB of 16-bit cache per 32K request; power by our estimate with the cards at 600 W, two processors and 15 per cent for the rest of the server, confirmed by a measurement on the delivered system.

gpt-oss-120b is published under Apache 2.0, with 117B parameters of which 5.1B are active per token; its files take about 65 GB, 60.8 GiB, which leaves 22.2 GiB of cache on one card. Its config.json gives 18 full-attention layers with 8 key/value heads of dimension 64, 36 KiB of 16-bit cache per token by our arithmetic, so a request of 32K tokens takes 1.125 GiB and one card answers 19 at the same moment. A request counts only while it is being answered, so far more staff can use the assistant than that number suggests.

The daily volume is small for this hardware. Take 300 reports a day with 1,500 tokens of model text each as an example, 450,000 tokens:

MLPerf® Inference: Datacenter v6.1, closed division, available category, gpt-oss-120b, offline scenario, entry 6.1-0050, HPE ProLiant Compute DL380a Gen12 with eight RTX PRO 6000 Blackwell Server Edition, 15,256 tokens/s, retrieved from mlcommons.org on 24 October 2026. Result verified by MLCommons Association. Per card is our division of the system result, not an MLPerf metric. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information. OCR rate by the method of our document AI guide.

The offline figure of 15,256 tokens per second comes from a benchmark that keeps eight cards busy with many requests at once; 1,907 tokens per second per card is our division of it, and a queue of report drafts reaches less, which is why the table also gives the time at half that rate. Our guide to batch LLM inference on-premise explains the queue.

Two cards. The entry server has two RTX PRO 6000 Server Edition, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives. Card 1 runs gpt-oss-120b for report text, the SOP assistant and enquiries; card 2 carries the OCR model, a vision-language model for printouts and the embedding model for the search index, split by MIG into isolated instances as our article on MIG profiles of the RTX PRO 6000 describes.

Four cards. The inference node with 768 GB and four NVMe drives runs gpt-oss-120b as a tensor-parallel pair, which leaves 2 × 83.0 − 60.8 = 105.3 GiB for the cache, 93 requests of 32K, for a laboratory with several sites or departments. The fourth card holds a copy of the pinned stack, so a new model version can run against the regression set while the validated one keeps serving. To start small, describe a report type you would automate first, and we propose a pilot and the server for it.

The RTX PRO 6000 Max-Q is the workstation card, a 300 W blower card; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition.

Notes for the IT team

1. Network. Put the AI server in its own segment. It reads from a LIMS export or a read-only interface and writes drafts to a staging area; it has no write access to the LIMS.
2. Engine and API. vLLM serves gpt-oss-120b through an OpenAI-compatible API behind an LLM gateway with keys and a query log. The model card states that gpt-oss "should only be used with the harmony format", so use an engine release that supports it.
3. Versions. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere, while routes such as /invocations and /metrics answer without the key (vLLM 0.32.0 security guide), so keep the engines on the internal segment. Run a current vLLM release (0.32.0 of 18 October 2026 at the time of writing), the NVIDIA Container Toolkit 1.17.8 or later and a driver release that NVIDIA lists as fixed in its bulletin of 30 September 2026; record each version in the change log of the validated stack.
4. Placeholders and check. Ask the model for structured output with named placeholders instead of values, and run the number check as a separate program that the model cannot influence.
5. Backups. The validation record is only useful if the model files, images and templates it names can be restored; our guide to backing up an AI server covers what to copy.
6. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime; our guide to RTX PRO Blackwell software support lists the engine versions for these cards.

Power and room

By our estimate, the two-card server draws about 1.8 kW at the wall with two EPYC 9135 processors and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. The four-card server draws about 3.4 kW with the cards at 600 W and needs 32 A or three-phase, or about 2.7 kW with the cards capped at 450 W. A server belongs in a server room, not in the laboratory beside the instruments, as our guide to GPU server power on 16 A, 32 A and three-phase feeds explains. Without a server room of your own, our Private AI/ML service runs the same platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU; by our reading that makes it an external provider in the sense of clause 7.11.4, which your quality manager takes into account.

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 the servers to order with two, four or eight RTX PRO 6000 Server Edition cards, 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 model, the report templates, the number check and the index over your methods. Our guide to buying a GPU server for LLMs lists every configuration.

FAQ