AI in manufacturing: keeping the knowledge when the people leave

Starting point: the knowledge leaves, the evidence obligations arrive

At the end of the first half of 2026, German manufacturing employed 5.29 million people, 144,100 or 2.7% fewer than a year earlier.¹ The documentation has not shrunk with it, it is simply spread across fewer shoulders.

It looks similar across Europe. In Sweden, 87% of the technology and industrial companies surveyed report difficulties recruiting engineering skills in electrical engineering, energy and automation.² In Norway, 62% of the member companies of Norsk Industri report unmet skills needs and name lost market share as a consequence.⁴ In Spain, official statistics counted 13,868 vacancies in industry alone in the fourth quarter of 2025.⁶

From the other direction come the requirements. The German supply chain due diligence act applies to companies with, as a rule, at least 1,000 employees in Germany. Until the start of 2024 the threshold stood at 3,000.⁷ The supplier with 60 people is not covered, but the questionnaires of its large customers land on its desk anyway. Cybersecurity follows the same pattern: in Italy the NIS2 baseline measures must be in place by October 2026, and whoever supplies affected customers receives the requirements by contract.⁸

With AI itself, the biggest gap runs not between countries but between large and small. In Sweden, 30.7% of companies with 10 to 49 employees use at least one AI technology, against 71.9% from 250 employees upwards.³ In France, only 15% of the small and medium-sized businesses in the food industry, the country’s leading industrial sector, use AI solutions.⁵

Use case 1: the knowledge of the retiring generation becomes queryable

The maintenance technician who has known for twenty years how this press behaves after a tool change retires in two years. Much of his knowledge is already inside the company, just scattered: in service reports, fault and inspection records, setup sheets and the handwritten settings in the folder next to the machine. A knowledge base built from those documents answers the night shift’s question. It names the source, that is the document and the page the answer comes from. Which setting worked for this batch last time? How was this fault fixed?

The difference to a chat in the browser is the reference. The answer comes from your documents and shows which report it came from. It does not come from the model’s memory, that is from what the program learned in general. What is not written down it will not find, and outdated documents produce outdated answers.

What the AI colleague does not do: it does not set up any machine. It shows what was set last time and who documented it. The decision is made at the machine.

Use case 2: preparing customer audits and supplier questionnaires

A large customer sends a supplier questionnaire with 120 items: supply chain due diligence, proof of origin, information security. In parallel, preparation for the next IFS or BRCGS audit is running. Almost every answer is already somewhere inside the company, in last year’s self-declaration, in a certificate, in a procedure. It is just that nobody finds it in the week before the deadline.

The AI colleague maps the questions to the existing documents, produces a gap list for everything with no evidence on file, and files draft answers with source references for the rest. The work shifts from searching to reviewing. The gap list is often the bigger gain: it shows before the audit what is missing.

This is the sharpest limit. Every answer in a supplier questionnaire is a commitment to your customer, every audit document is evidence towards a certification body. Both go out after technical approval, never automatically.

Use case 3: instructions, labels and manuals in the language of the shift

In food production and in assembly, shift teams work in mixed languages. The hygiene rule, the setup instruction and the short version of the operating manual are generated from the approved document in the language of the employees. The key element is a company glossary: machine names, part designations and technical terms are fixed once, so that the same assembly is not called something different in every version.

The same glossary carries the declaration and label texts and the operating manual for export. It carries the other way too: the Spanish supplier specification and the Italian test certificate become readable without an outside translation portal.

The limit applies here too: translations are drafts. For safety-relevant briefings and legally required information, the approved version remains authoritative, and it is the qualified person who approves it. Permissions apply down to individual tools: whoever may not see the costing does not see it through the AI either. Every access is recorded in the audit log, a record of who opened which document and when.

Connectors: what can be linked to your ERP

An AI colleague is only as useful as what it is allowed to reach. MCP is an open standard that lets an AI program address another program. What it needs is a documented interface, that is a published connection point where two programs exchange data. Every single tool is released separately, and every call is logged. Default setting: reading tools free, writing tools only with confirmation. The table shows how far the AI connection to common software reaches today, checked in August 2026. The middle column gives the technical names of those interfaces; the column that matters for you is the right one.

What you deliberately will not find here: the machine and MES layer, that is production control. Controllers, shop floor data capture and detailed scheduling were not examined. An integration promise that breaks in the first conversation costs more than it brings. The benefit arises in the office and in work preparation. If your software is missing, the knowledge base and audit preparation still work: they only need documents. More under integrations.

The separation also has a legal reason. If AI is built into a regulated product as a safety component, it can qualify as a high-risk system under the AI Regulation. Under Regulation (EU) 2026/1744 its obligations apply from 2 December 2027, and from 2 August 2028 for embedded systems.⁹ An assistant in the office does not normally fall under this.

Data protection and secrecy: specifications stay in the plant

Customer specifications, recipes, costings and inspection data are trade secrets, often bound by a non-disclosure agreement with your customer on top. For personal data the General Data Protection Regulation applies. The business remains responsible even when an AI tool processes it. If a provider is involved, that requires a data processing agreement, a contract setting out what this provider may do with the data. This article is not legal advice.

If data goes to a country outside Europe, Chapter V of the Regulation applies as well. You then need a legal basis for the transfer and, depending on the instrument you choose, an assessment of the legal situation in the destination country. If storage and model processing take place within the European Economic Area, that step falls away. This does not make the processing lawful automatically, but it does make it easier to justify.

Kisuno runs in ISO 27001 certified data centres or on your own server. Customer data does not train any models. Telemetry, that is the reporting back of usage data, is switched off by default. If nothing at all may go outside, you run your own instance in your own data centre. An instance is the installation that runs for your company alone. LibreChat, the base, is MIT licensed, and the repository with our own extensions follows shortly. The data processing agreement and the subprocessor list are published openly on the website. The choice covers the current models from OpenAI (ChatGPT) and Anthropic (Claude); they run in the AWS data centre in Frankfurt. AWS is a US company, so the US CLOUD Act applies even with servers in Frankfurt, and we say so openly. How that path runs in detail is set out in the article on the Anthropic models on Amazon Bedrock. For customers with higher data-protection requirements, open frontier models such as GLM, Kimi and DeepSeek are available, operated by European companies on servers in Europe; the CLOUD Act does not apply there.

One more point the European statistics show clearly: in Norway, 71% of AI-using companies rely on off-the-shelf software, in Sweden 62%.³ Such tools are available quickly, but they do not know the business: no service reports, no inspection records, no audit answers.

Frequently asked questions from manufacturing and food production

Sources

All figures come from these sources, last checked in August 2026, with the country each was collected for.

1. Federal Statistical Office of Germany, press release N057 of 14 August 2026 (DE): destatis.de
2. Industrirådet, report on the skills needs of industry (SE, survey May to October 2025, 1,500 companies), at Industriarbetsgivarna: industriarbetsgivarna.se
3. Eurostat, dataset isoc_eb_ai, survey year 2025: AI use by size class and how AI is sourced (SE, NO): ec.europa.eu/eurostat
4. Norsk Industri on the NHO skills barometer 2025, 24 February 2026 (NO): norskindustri.no
5. ANIA, Nos chiffres clés, on the food industry (FR): ania.net; AI use by sector in the Baromètre France Num 2025: francenum.gouv.fr
6. INE, Encuesta Trimestral de Coste Laboral, fourth quarter 2025 (ES): ine.es
7. German Supply Chain Due Diligence Act, section 1, scope (DE): gesetze-im-internet.de
8. Agenzia per la cybersicurezza nazionale, deadlines of the Italian NIS2 transposition (IT): acn.gov.it
9. European Commission, regulatory framework for artificial intelligence, with the deadlines as amended by Regulation (EU) 2026/1744: digital-strategy.ec.europa.eu