First AI use cases in a manufacturing SME: choosing tasks and measuring the hours saved

The best first AI tasks for a manufacturer

The best first AI use cases in a manufacturing SME are office and engineering tasks that come often, take long per case, already exist in digital form and can be checked by a person in a short time. Typical candidates are entering customer orders from e-mails and PDFs, preparing quotations, handling complaints, writing manuals and translations, reading supplier certificates and declarations, and taking meeting minutes. Choose two or three by counting cases and minutes, measure the time per case before and after on a sample, and run them on one company AI server, so that each new task adds software, not hardware.

This guide is for owners and managing directors of manufacturers with roughly 20 to 500 staff, and for the heads of sales, quality, engineering and the office who will propose the tasks. It gives a ranking method, ten tasks with what the AI does in each, a way to measure hours saved, and the server that carries them. It does not give money figures for return on investment; the hours you measure in your own company are the basis for that calculation.

Many companies have not started yet. Eurostat reports that in 2025, 19.95 per cent of EU enterprises with ten or more employees used AI: 17 per cent of small enterprises (10 to 49 staff), 30.36 per cent of medium-sized ones (50 to 249) and 55.03 per cent of large ones. Among enterprises that had considered AI but did not use it, 70.89 per cent named a lack of relevant expertise, 52.52 per cent a lack of clarity about the legal consequences and 48.83 per cent concerns about data protection and privacy; only 20.68 per cent said it was not useful (Eurostat, "Use of artificial intelligence in enterprises", data extracted December 2025, read 25 October 2026). Those three reasons are what a first project must answer: someone who knows how, clear rules, and data that stays in the company.

Four questions that rank a task

For each candidate task, ask four questions and write the answers in one table. The first two give the size of the prize, the last two decide the order and the place.

1. How many cases per month? Orders entered, quotations written, complaints handled, certificates checked. Count them from the ERP, the mailbox or the quality software for a normal month.
2. How many minutes per case today? The working time of the person who does it, from opening the case to finishing it, without waiting times. A rough figure is enough for ranking; it is measured properly later.
3. Is the data ready? Are there examples in digital form that the AI can read: e-mails, PDFs, earlier quotations, old 8D reports, released manuals? Paper archives and knowledge in people's heads come later, with their own projects such as capturing expert knowledge before staff retire.
4. How confidential is it? Public, internal, customer-confidential (drawings, prices, specifications under NDA) or personal data (names, health data). The answer decides whether a cloud tool is acceptable at all or whether the task belongs on your own server.

Cases per month times minutes per case gives the time pool: the hours per month the task takes today. A task with a large pool, ready data and quick checking goes first. Add one more look at each task: the time a person needs to check the AI's result. Order entry is checked in seconds against the original; a quotation for a special machine takes much longer to check. Quick checking makes a good first task, because the benefit shows from the first week.

Ten AI examples for manufacturers, ranked by time pool

The table shows ten typical tasks with example values for a manufacturer of about 150 staff. Replace the example values with your own counts; the ranking is what matters.

Example values for illustration only, not measurements; time pool = cases per month × minutes per case ÷ 60.

In this example, quotations, questions to the company knowledge, order entry and manual sections have the largest pools. Order entry and supplier certificates are quick to check and have ready data, which makes them strong first tasks. Quotations have a large pool but need more checking and earlier quotations in a usable form. A sensible first set here is order entry for the sales office, complaints and audit preparation for quality, and a company assistant that answers questions from released documents for everyone: three departments see a result, on one server.

Each task has its own guide on this site, with the details that matter in practice:

• AI order entry from e-mails and PDFs into the ERP and AI quoting from drawings for job shops, or offer writing for machine and plant builders.
• AI in quality management: ISO 9001, audits and 8D, inspection certificates read and checked by AI and supplier declarations for RoHS, REACH and SCIP.
• Machine manuals under the Machinery Regulation, on-premise machine translation and meeting minutes with a local AI.
• Customer audits and supplier questionnaires, workplace risk assessments and a service assistant for machine builders.

Tasks to leave for later

Some tasks sound attractive and make poor first projects. Visual inspection at the line needs images of good and bad parts, labelled, and small cards at the station; it is a project of its own, described in our guide to visual inspection AI without data scientists. Predictive maintenance needs sensor data collected over months, and much of it runs on ordinary processors, as our guide on when predictive maintenance needs a GPU explains. Tasks whose result decides about safety, or about people, such as evaluating employees, need more care than a first project can give.

Tasks without digital data also wait. If the earlier quotations exist only as printed folders, the first step is scanning and sorting them, and the AI task follows when the data is there.

We supply AI servers built to order and set up the models and the document search under our Private AI/ML service; send us your list of candidate tasks with rough case counts, and our team tells you which ones one server carries and what it needs in the building.

Measuring the time saved with AI, before and after

Estimates of time saved are easy to write and hard to believe. A measurement is better, and it needs no special tools.

1. Before the start. For each chosen task, time 20 to 30 normal cases with start and end time, done the usual way, by the people who usually do them. Note the cases that needed rework or a question back to the customer.
2. After the start. When the AI is in use and the team has worked with it for a few weeks, time the same number of comparable cases. Include the time the person needs to check and correct the AI's result, and count the corrections.
3. Compare medians. Use the middle value of each sample, so that one unusual case does not distort the result. Hours saved per month = cases per month × (median minutes before − median minutes after) ÷ 60.
4. Record the quality side. Count errors that reached the ERP, the customer or the next department, before and after. A faster process with more errors has not saved time.
5. Repeat after changes. A new model, new document sources or new staff change the figures; measure again after each change.

The same sample gives a second benefit: real cases with checked answers, which become the test set for the next model change. Our guide to a private AI pilot and what to measure covers the technical figures the IT side collects in the same period, such as requests in flight and answer lengths, which size the production hardware.

Some benefits do not show as hours: a quotation that leaves on the same day, a complaint answered before the customer calls again, a new colleague who finds the right work instruction without asking. Note them as they happen, with the date and the case. They are often what convinces the rest of the company.

Who does what in the first AI project

A first AI project in a company without a large IT department needs few roles, filled by name.

• An owner per task. The head of the department where the task lives decides what the AI may do, approves the rules and reports the measured hours.
• A key user per task. The person who does the work today tests, gives feedback and checks the first results closely.
• One person for the platform. Inside the company or at a service partner, someone owns accounts, document sources and updates; our guide to an AI server without an IT department shows how this is split with a supplier.
• The management. It sets the usage rules, decides which data may go where, and involves the data protection officer and, where one exists, the works council early, as our guide on the works council and an internal AI assistant explains.

Users should also get a short introduction. Article 4 of the AI Act asks providers and deployers of AI systems to take measures to support the AI literacy of their staff, as our guide to AI literacy training describes. A written description of each task in business terms helps the quote and the project; our AI requirements template gives the headings, and the steps of an own AI server project shows what follows the order.

One server for the first tasks

The first tasks are small loads for an AI. A sales office enters a few hundred orders a month, quality handles a few complaints a week, and the assistant answers questions from staff during working hours. One server carries all of them, which is why the first project should size the server for the whole company.

Users-per-card rule (0.9 × driver-visible memory, less 3 GiB per card and less the weights): Qwen3.8-27B FP8 weights 30,866,866,928 bytes = 28.75 GiB (Hugging Face file list, read 25 October 2026) with FP8 cache, 1.14 GiB per 32K conversation; gpt-oss-120b about 65 GB = 60.8 GiB (Hugging Face file list, re-read 25 October 2026) with 16-bit cache, 0.5625 GiB per 16K conversation; power by our estimate (cards, processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

A conversation here is one request being answered at the same moment. A company of 150 people rarely has more than a handful of requests in flight at once, because most users are reading or typing, so the entry server leaves room for the next tasks; our guide to private ChatGPT servers by company size gives the arithmetic. A token is a piece of a word; 32K tokens hold a long order with its attachments or a full chapter of a manual. Qwen3.8-27B is a vision-language model under Apache 2.0 that reads images and documents, which suits scanned orders, certificates and drawings. gpt-oss-120b, also Apache 2.0, has 117 billion parameters, of which 5.1 billion are active per token, and serves the assistant.

The entry server has two NVIDIA RTX PRO 6000 Blackwell Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. It runs from one ordinary 230 V 16 A feed and belongs in a server room or a cooled IT room. Two cards are the sensible start: one model per card, and if one card fails, the other can carry the most important model until the repair. Without a server room, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU. Smaller firms find the same reasoning for 10 to 100 staff in our guide to an own AI server for a small company, and larger plants with inspection and simulation in our guide to an AI server for manufacturing.

Notes for the IT team

• Interfaces first. Order entry and certificates need a defined way into the ERP or the quality software, ideally one interface with a person's confirmation step; the AI never writes directly into master data.
• Document sources. The assistant reads released documents from the file share or document management system, with access filtered by the user's roles in the company directory.
• Measurement data. Keep the gateway log of requests with timestamps and token counts from the first day; it complements the business-side timing and is the basis for sizing the next expansion.
• Security. The server is not reachable from the internet; sign-in runs through the company directory, and the inference engine and its web front end are kept on current versions, as our guide to in-house AI security questions for management summarises.

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 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 models, the document search, access control and the connections your first tasks need. Name the two or three tasks you want to start with, and we propose a configuration for them.

FAQ