AI requirements template in business terms: what to write and what it means for the server

AI project requirements: what to write down before you ask for a server

An AI requirements document in business terms answers twelve questions about your company in your own words: who will use the AI, for which tasks, with which documents and languages, how confidential the material is, when the busy hours are, where the system may stand and who will look after it. A supplier turns those answers into the graphics cards, memory, storage, power and software, so you need no technical knowledge to fill in the template. Two pages of estimates in your own words are more useful to a supplier than a list of technical terms copied from a brochure.

This template is for owners, managing directors and heads of department in companies without a large IT team. It is the step before a technical request. Once it is filled in, your supplier or your IT writes the technical version from it, as described in our guide on buying an AI server: what to specify; a public body that must tender uses our GPU server tender specification.

Where you do not know a figure, write an estimate and mark it as one. A supplier can work with "about 40 people, perhaps 80 next year" far better than with an empty field.

The template: a checklist of twelve questions in plain language

Copy the twelve headings into a document and answer each in a few lines. Examples in brackets show the level of detail that helps.

1. Who will use it. Departments and the number of people in each, now and in two years (sales 12, quality 6, engineering 20, office 8).
2. Tasks, with one real example each. What people will ask or hand over (draft a reply to a customer complaint; find the test limit in a customer specification; summarise a 40-page supplier contract; translate a manual into Polish).
3. Documents. Which ones the AI should answer from, roughly how many and where they live (about 30,000 files on the file server, the quality manual in the document system, customer specifications as PDF in email).
4. Who may see what. Whether every user may see every document, or whether departments are separated (engineering drawings only for engineering and sales).
5. Languages. In which languages staff write and documents exist (German and English; supplier documents in Italian; staff on the shop floor writing in Romanian).
6. Confidentiality. Which material must never leave the company, which contains personal data, and whether health or other special data is involved (recipes and drawings; HR files; sick notes).
7. Busy hours and batch work. When most people will use it, and what can run overnight (mornings 8 to 10; translate the full manual set once a quarter).
8. Systems to connect. Where answers come from or go to (ERP for order status, the email system, the document system).
9. Where it may stand. A server room with a rack, an office corner, or nowhere on site (rack with 16 A power, no air conditioning; or an EU data centre).
10. Downtime. What happens if the AI is unavailable for a day (inconvenient; or the order desk stops).
11. Who looks after it. Your own IT, an external IT service, or a managed service from the supplier.
12. How you will judge success. One or two measurable results (time to answer a complaint falls from 30 to 10 minutes; 9 of 10 test questions answered correctly with the source).

The order matters less than completeness. Questions 1, 2, 3 and 9 decide the size of the hardware; questions 4, 6 and 11 decide how it is set up and run.

From your answers to the server: the translation table

The table shows what a supplier derives from each answer. The terms in the middle column are explained below it.

Our summary of how our sizing guides use each input; figures for each row in the sections below.

A request in flight is a question the AI is answering at this very moment. A conversation on the card is the memory one such request needs. The context length is how much text one request may carry: the question, the retrieved document passages and the answer. The search index is the prepared copy of your documents the AI searches before it answers.

People and busy hours: how many cards

Head count is not the figure that sizes a server. Most employees with an assistant open in the browser are reading or typing, and only some requests are being answered at any moment. Our company-size sizing guide uses example values: 40 per cent of staff active in the busiest hour, six requests per person per hour, 30 seconds per request and a peak twice the hourly average. That gives about four requests in flight per 100 staff. These are example values that a pilot replaces with yours, which is why question 7 asks for your busy hours.

Our users-per-card rule takes 90 per cent of the 95.6 GiB the driver reports on a 96 GB RTX PRO 6000, subtracts 3 GiB per card and the model's weights, and divides the rest by the memory of one conversation. With gpt-oss-120b at a 32K context and a 16-bit cache, one card holds about 19 conversations, as derived in our guide on how many users one RTX PRO 6000 serves. With the example values, that covers the busy-hour peak of about 475 staff on one card.

Batch work changes the picture. A quarterly translation of all manuals can run at night, when the assistant is idle, and needs no extra card if you say so in question 7. Our guide to batch LLM inference on-premise covers such night jobs.

Tasks and documents: which models, how much storage

Each kind of task adds a model. Chat and drafting use a language model such as gpt-oss-120b. Answering from your documents adds an embedding model, which turns passages into numbers the search can compare, and usually a reranker that sorts the hits; Qwen3 Embedding 0.6B, under Apache 2.0, produces vectors of up to 1,024 numbers and its card names more than 100 languages. Meeting transcripts add a speech-recognition model, and scanned forms add a model that reads images. The small search models fit beside the main model: one 96 GB card can be split into up to four isolated parts with NVIDIA's MIG feature, which is why our entry server has two cards, one for the assistant and one split for search and speech.

The length of a typical document sets the context length, and with it the number of conversations per card. By the same rule, gpt-oss-120b holds 79 conversations at 8K tokens, 19 at 32K and 4 at 128K on one card. An answer from a few short work instructions needs far less context than a summary of a whole contract, so the example in question 2 is the most useful line in the template.

The amount of documents sets the index. At four bytes per number, one million passages with 1,024 numbers each take about 4.1 GB before index structures and the text itself, by the arithmetic in our company-size guide. 30,000 files of ten passages each make 300,000 passages and about 1.2 GB of vectors by our estimate; the source files, every model you keep and the logs need more storage than that. Before the index is built, someone has to decide which documents go in and which version is current, as our article on preparing company documents for an AI assistant explains. We size servers from exactly these answers; send us the filled-in template, estimates included, and our team replies with a configuration.

Languages and confidentiality

Language is a model question. The Llama 3.3 model card, read on 25 October 2026, lists eight supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. A team writing in Romanian or Polish needs a different model, tested on its own texts, and our comparison of open LLMs for European languages shows candidates.

Confidentiality decides the set-up more than the size. Material that must not leave the company, such as recipes, drawings and price calculations, runs on the own server only. Personal data is "any information relating to an identified or identifiable natural person", in the words of Article 4(1) of the GDPR, and Article 9(1) names special categories, among them "data concerning health", whose processing is prohibited unless an exception applies. If question 6 mentions such data, your data-protection officer confirms the legal basis, and our guide to a DPIA for an internal LLM shows the assessment that may be needed. Companies that keep a cloud assistant for general work and run confidential work locally need a rule staff can follow, which our guide to hybrid AI with confidential data local sets out.

Room, downtime and who looks after it

The answer to question 9 picks the build. A rack server with two RTX PRO 6000 Server Edition cards draws about 1.8 kW at the wall by our estimate and runs on one 230 V 16 A feed. Four cards at 600 W draw about 3.4 kW, more than the 2.94 kW we plan per 16 A feed, so they need 32 A, three-phase power or a 450 W limit per card. A rack server is loud and needs cooling for its full load; our guide to a GPU server in an office explains when a workstation fits a room where people work. Without any suitable room, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

Our estimates: cards, processors and 15 per cent for the rest of a server, as in our two- and four-card guides; workstation from our guide to GPU servers in an office. A measurement on the delivered system confirms the figure.

Downtime decides whether one server is enough. If the order desk stops when the AI stops, plan a second server so that either one carries the peak; if a day without the assistant is only inconvenient, one server with good spare-parts terms will do. Our guide to high-availability LLM inference on two nodes describes the set-up with two servers.

Question 11 is the one companies without IT staff most often leave open. Someone installs updates, renews certificates, watches the disks and adds users. If nobody in-house can do it, a support agreement covers it: after the project our Private AI/ML service offers ongoing support under an agreed SLA, or trains your team to take it over. Our guide to an AI server without an IT department lists who does what.

Success criteria and a filled-in example

Question 12 turns the project into something you can check. Pick one or two tasks and measure them before the AI arrives: the time per complaint, per quote or per translated page, and a test set of 20 to 50 real questions with known answers. The same measurement after a pilot shows whether it works. Our guide on what a private AI pilot should measure adds the technical figures your supplier collects at the same time.

A filled-in example, a manufacturer with 120 staff, all figures invented for illustration:

Example values; peak by the example values of our company-size guide (120 × 40 % × 6 per hour × 30 s × 2 ≈ 5), our estimate.

Five requests at the peak are far below the 19 conversations one card holds at 32K, so card one runs gpt-oss-120b and card two, split with MIG, runs the search models and room for a speech model later. The company could start with one task and add the others on the same server later; our guide to first AI use cases in a manufacturing SME helps choose that task.

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 AI servers to order with two, four or eight RTX PRO 6000 Server Edition cards, assembled and tested before delivery, 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. Sizing is free of charge. Models, document search, sign-in and the connection to your systems are our Private AI/ML service, with our engineering partner Vixen.UNO; if a question in the template is hard to answer, describe what you know and we work through the rest with you in the first call.

FAQ