AI in everyday work
AI in manufacturing: three workflows that keep knowledge in the company
Your most experienced employee retires in two years, the specification for the next quote runs to 140 pages, and the audit is due in spring. This article walks through three workflows to show how an AI colleague works with your company’s knowledge, what can be connected to your ERP, and where the limits are.
In brief
- Only 17% of SMEs have a defined digital strategy, compared with almost 50% of large companies: that is the state of the 2020 Swissmem survey.1
- In the Swiss tech industry, AI close to production is still rare: half of the companies have not considered it, 10% are running pilots.2
- Kisuno works with your company’s knowledge. Service reports, test protocols, notes on bills of materials and change requests become a knowledge base: a collection of your own documents that you can question in plain sentences.
- Your design data stays in Switzerland or on your own server and does not train any models. Your ERP can be connected if it has a documented interface. An interface is the point where two programs exchange data; the Model Context Protocol (MCP) is the open standard an AI program uses to reach it.
Starting point: the strategy is missing, the knowledge is retiring
Only 17% of SMEs have a defined digital strategy, compared with almost 50% of large companies. That comes from the 2020 Swissmem survey on Industry 4.0, with 267 companies from German-speaking Switzerland and 54 from the French-speaking part.1 With artificial intelligence the gap between announcement and shop floor is wider still. In 2024, ETH Zurich, together with Swissmem and Next Industries, surveyed more than 200 Swiss tech companies. Half have not even considered AI close to production, and 10% are running pilots.2
There are good reasons for the hesitation. Swissmem names security risks and data protection questions, regulatory requirements, and data structure and data quality as the hurdles.4 Putting design data into somebody else’s cloud buys you a problem that no efficiency gain makes up for. And anyone with an archive that grew over thirty years knows that data quality does not appear at the push of a button. The workable answer is not to wait, but to start small and under control.
Because a clock is running, and it hangs in the HR office. The loss of knowledge when experienced specialists leave is regarded as a major problem for Swiss industry. Swissmem presents the research project Digitaler Wissenserhalt im Unternehmen run by HES-SO Valais-Wallis: it captures the practical knowledge of long-serving staff and makes it usable for training newcomers.3 That is exactly what your workshop lives through every week: the answer is in a document nobody can find any more, and the colleague who knows it by heart has two years left before retirement. A language model, that is a program that reads text and writes text of its own, can make such documents findable again.
In food production there is a legal track on top. Self-monitoring according to the HACCP principles is an obligation of the business.5 Industry guides to good practice are approved by the Federal Food Safety and Veterinary Office; they are a recognised aid, but the applicable food law takes precedence over them. The basis for this is Article 80 of the Swiss Foodstuffs and Utility Articles Ordinance.6 The obligation stays with the business. The paperwork around it does not have to.
- Service reports, test protocols and change requests from decades of work
- A knowledge base inside your own company, with permissions per tool
- A question in everyday language, from the shop floor or the office
- An answer with its source document, checked by the specialist
Use case 1: practical knowledge stays in the company
A machine built in 2011 has stopped at the customer’s site. The service technician is looking for the cause and is certain the same fault has occurred before, except that it is written up in a service report nobody can find. Two doors down, the design engineer asks why a variant was built with that particular bearing arrangement back then. The same employee can answer both questions, and he retires in two years.
Service reports, test protocols, notes on bills of materials and change requests go into the knowledge base exactly as they are: as PDFs, as scans, as a folder structure that grew over time. Questions are asked in everyday language, the way you would ask a colleague, and not with search commands: which fault has already occurred on which series, why that bearing arrangement was chosen, what the acceptance protocol said. The answer comes with a reference to the document it is based on. So you look it up in seconds instead of having to take it on trust.
- A question in everyday language, from service or engineering
- Reports, protocols and change requests, with permissions per tool
- An answer with its source document, checkable in seconds
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. Start with one machine type or one series and check the first answers against cases you know by heart. The quality of the answers grows with the quality of the archive, and the business sees on its own example what orderly filing is worth from now on.
Use case 2: examining the specification before the quote
The customer’s requirements specification arrives with 140 pages. The submission deadline is in ten days, and the only person who can read it is the senior engineer who is fully booked anyway. The document goes into the AI colleague, and out comes a structured requirements list: what is required, which chapter it appears in, which standard is cited and which tolerance is named.
In a second step that list is matched against your own specifications, standards lists and earlier bids from the knowledge base. What your portfolio covers is shown as covered, with a reference to your own specification. What it does not cover is flagged. What is worded ambiguously goes onto the list of questions for the customer, with the exact spot in the specification. The senior engineer then reads the flagged passages instead of all 140 pages, and the decision whether to bid rests on a basis rather than on a feeling.
- Customer specification, exactly as it arrives
- Your own specifications, standards lists and earlier bids
- Requirements list with flagged gaps and open questions
What deliberately does not happen: the AI colleague makes no commitment and replaces no technical clarification. It delivers a draft with sources, and a language model can miss a requirement or classify it wrongly. That is why the list belongs in the hands of the specialists who judge what is feasible. The gain is that the review starts on day one instead of day seven.
Use case 3: quality assurance and audits in food production
The IFS Food or BRCGS audit is due in March. The question catalogue is mirrored against your own HACCP and quality management manual, and out comes a gap list: which question the manual answers, where there is only a cross-reference, where there is nothing. Quality assurance works through that list instead of reading the manual from the start three times over.
The second pile is supplier documentation: specifications, certificates, certificates of analysis. The AI colleague summarises them, flags expiry dates and names what is missing. The third case is multilingual label and declaration drafts as well as operating instructions. They are produced with your company glossary, that is with your own list of agreed terms, and use the same terms across every language. The order of ingredients and the highlighting of allergens are check points on the quality assurance list, not results produced by the machine.
- Audit questions, supplier documents and recipe files
- Comparison against your own HACCP and quality manual
- Gap list, document overview and label draft for approval
The boundary here is drawn sharply, and rightly so. Self-monitoring is a statutory obligation of the business.5 The AI colleague writes drafts and finds gaps; sign-off has to come from quality assurance. With labels and operating instructions, product liability is added on top: professional approval belongs explicitly to a person, and a machine-written text does not on its own satisfy any legal requirement. This article is not legal advice.
Connectors: what can be linked to your systems
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. According to a 2023 Profondia analysis of the roughly 13,500 largest Swiss companies, the most widely used ERP in Switzerland is Abacus with 17%, ahead of SAP with 16% and Microsoft Dynamics with 8%.7 The table shows how far the AI connection to common software in manufacturing and food production 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.
| Software | Interface | Connection to Kisuno |
|---|---|---|
| Abacus | REST API with OAuth2, access through the partner programme | Connection via MCP possible; access runs through the Abacus partner channel |
| SAP Business One | Service Layer (OData), documented | Connection via MCP possible: items, stock, business partners, orders, invoices |
| Microsoft Dynamics 365 Business Central | API documented | MCP server available from Microsoft; read-only out of the box in Business Central online, write access can be enabled individually |
| Odoo | Open interface, documented | Ready-made MCP connections available from the community: manufacturing orders, bills of materials, stock |
| Microsoft 365 (Outlook, Teams, SharePoint) | Graph API, openly documented | MCP server available from Microsoft; email, calendar and files directly addressable |
| proALPHA, Opacc | Interfaces exist, documentation only for customers and partners | Connection via MCP possible; the route runs through the vendor |
| CSB-System | No publicly documented API | The route runs via exports and documents |
Is your software missing from the list? The three workflows above work with documents and need no system connection at all. To begin with, an export from your MES, that is from your production control system, or from your test equipment system is enough. How a connection via MCP works in principle, and which systems are prepared, is described on the integrations page.
Trade secrets: your design data stays in your hands
The decisive point in this industry is not primarily data protection law but manufacturing and business secrecy. Article 162 of the Swiss Criminal Code makes the betrayal of such secrets a criminal offence where a statutory or contractual duty of secrecy exists; it is prosecuted only on complaint.8 Exactly those secrets sit in your design data, your costings and your customer specifications, often protected by a non-disclosure agreement on top. Anyone supplying medtech, pharmaceuticals or aviation also carries their customers’ contractual requirements against any flow of data into third-party clouds. This article is not legal advice.
Hence the core statement: your design data stays in your own instance. An instance is the installation that runs for your business alone, with permissions per tool and a log for every access. It runs in ISO 27001 certified data centres operated by Green in Lupfig (Switzerland) or on your own server. On offer are 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. 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. Customer data does not train any models. Telemetry, that is the reporting back of usage data, is switched off by default. If customer requirements mean that nothing at all may go outside, you run the instance in your own data centre. The LibreChat base is available under the MIT licence; the repository with our own extensions follows shortly.
For personnel data the revised Swiss Data Protection Act applies on top. The business remains responsible even when an AI tool processes the data. If a provider is involved, that requires a data processing agreement, a contract setting out what this provider may do with the data. The data processing agreement and the subprocessor list are published openly on the website, the technical measures in the Trust Center.
Frequently asked questions from manufacturing and food production
Our documentation grew over decades and is anything but consistent. Is this still useful for us?
Yes, with one honest caveat. The knowledge base takes PDFs, scanned reports and entire folder structures, but the quality of the answers grows with the quality of the archive. Start with one machine type or one series, check the answers against cases you know, and expand afterwards. A clean-up project up front is not necessary, and it would never be finished anyway.
Does the AI understand our technical drawings?
Text and tables yes, geometry only to a limited extent. Bills of materials, title blocks, stamps, revision notes and handwritten remarks on the drawing are read; a language model does not assess a drawing in the sense of a design review. That is enough for finding documents and notes again, and not enough for approving a design.
Does our design knowledge stay protected?
Permissions apply down to individual tools, in the order user, role, group and server default, and every access is recorded in the exportable audit log. Managed operation runs in Switzerland, self-hosting runs on your own server; in neither case does customer data train any models. If you have to meet contractual requirements from your customers, you run the instance inside your own company.
What does it cost?
The managed instance costs CHF 7 (EUR 7.50) per person per month. Through the end of the current month: 50 percent off with a twelve-month term, that is CHF 3.50 (EUR 3.75) per person per month. Testing is free and not limited in time: every new account starts with a starting credit of CHF 50 (EUR 55) for model usage, with no payment details. If you self-host, the software costs nothing, permanently and from a single user. The LibreChat base is available under the MIT licence, a free licence with no licence fee; the repository with our own extensions follows shortly. All details are on the pricing page.
Can we start with a single department?
Yes, and that is the usual route: service or quality assurance with a clearly defined knowledge base. You manage users, permissions and the shared monthly budget in the admin panel, and the budget has a soft limit so that costs stay predictable. Further departments follow once the first one is visibly working.
Does this replace our ERP or PLM, and does it work with Abacus, SAP Business One or Odoo?
No, nothing is replaced. Your ERP remains the leading system for master data, orders and bills of materials, your PLM for revisions and releases. The AI colleague works on top of them: it reads, summarises and drafts. The AI connection to Abacus, SAP Business One, Dynamics 365 Business Central or Odoo runs via MCP and their documented interfaces; the table above shows what is possible for each system. Permissions per tool and a log for every call apply throughout.
We write multilingual operating instructions and labels. How far can we rely on this?
The draft is produced with your company glossary and uses the same terms across every language. Professional approval remains mandatory: product liability hangs on instructions and labels, and a language model can produce errors. For machinery the operating instructions are part of conformity, and the language version for the destination country is a legal requirement; approval remains with the manufacturer. No draft is a valid declaration on its own. This answer is not legal advice.
See it on a case of your own
Bring a customer specification, a stack of service reports or your audit question catalogue, and we will walk through the workflow together. Or sign up: because demand is high, we hand out access to the managed Swiss instance in stages.
Sources
All figures and legal references in this article come from the following sources, last checked in August 2026.
- Swissmem, survey on Industry 4.0 in the Swiss mechanical and electrical engineering industry (surveyed in 2020, published on 16 November 2020; 267 companies in German-speaking Switzerland, 54 in the French-speaking part), digital strategy in SMEs and large companies: swissmem.ch
- Study by ETH Zurich with Swissmem and Next Industries covering more than 200 Swiss tech companies on the state of AI (surveyed in 2024), reported by MaschinenMarkt Schweiz on 14 November 2024: maschinenmarkt.ch
- Swissmem, research project Digitaler Wissenserhalt im Unternehmen run by HES-SO Valais-Wallis, capturing practical knowledge, presented at the Swissmem Industry 4.0 conference in January 2024: swissmem.ch
- Swissmem, artificial intelligence: current applications and challenges in Swiss industry: swissmem.ch
- Federal Food Safety and Veterinary Office, guides to good practice for food businesses and self-monitoring under HACCP: blv.admin.ch
- Article 80 of the Swiss Foodstuffs and Utility Articles Ordinance (LGV), industry guides and their approval by the BLV: fedlex.admin.ch
- Profondia analysis of ERP use among the roughly 13,500 largest Swiss companies, reported by Swiss IT Magazine in July 2023: itmagazine.ch
- Article 162 of the Swiss Criminal Code (StGB), violation of manufacturing or trade secrets, prosecuted only on complaint: fedlex.admin.ch
Last reviewed: