Which AI Tools Pay Off for Manufacturing SMEs?

There is an AI tool for almost every step in manufacturing today. So the question for a manufacturing SME is not whether tools exist, but which ones pay off at your size. This post sorts the categories by the condition under which each pays off.

---

Key points

• Categories, not vendors. We name the condition, not a ranking.
• Document extraction is the most mature category. One process, clearly bounded.
• OCR is not extraction. Reading and assigning are two steps.
• Natural-language access needs a clean data basis.
• AI agents prepare decisions, they do not make them.
• Fixed steps belong in the program.
• AI assistants help knowledge work when someone checks.
• The data foundation decides, not company size.
• On-premises is a serious option.
• Machine vision: pilot first, then decide.

---

Where manufacturing SMEs stand today

The sector-specific Swiss figure is now two years old. It comes from the 2024 study by ETH Zurich, Swissmem and Next Industries, across 200 Swiss tech companies. More than half had not yet considered AI in manufacturing. About a quarter were planning, testing or already applying it. We worked through what to do about that in How to Get AI Going in Manufacturing. The newer surveys are broader, and each measures something different:

Together these sources support one claim: adoption is rising and falls off as company size drops. A figure for manufacturing SMEs specifically cannot be derived from them.

There are no newer adoption figures. The AXA/Sotomo survey of March 2026 no longer asks the AI-usage question, and neither Destatis nor Bitkom has published a new wave since.

Document extraction: OCR is not extraction

OCR turns pixels into characters. Which of them is the article number, which the quantity and which the delivery date is decided by a second step. That is where many document-capture projects stall. The company buys text recognition, gets text, and still retypes into the ERP.

OmniDocBench (CVPR 2025), the academic reference benchmark, compares OCR methods against vision-language models. It finds no overall winner, but strengths and weaknesses by document type. The methods diverge most on tables. The line items on a purchase order are a table. A method can read every character correctly and still assign quantity and price to the wrong line.

A Swiss CNC contract manufacturer receives more than 50 orders a week. The effort for order entry fell from 20 hours to 2 hours (case study). The flow in detail, and where the human approval sits, is the subject of AI Agents in Order Intake.

Pays off when: documents arrive regularly and in varying layouts. With five suppliers on stable forms, a template is enough.

Natural-language access to production data

The data has long been sitting in machines, MES and ERP. Access is what is missing. Anyone who cannot query the database files a ticket and waits. A natural-language interface translates the question into a query and returns a table or a chart. We use talk.factoriq for this and have described the architecture in detail. Where this tier sits within a whole architecture is shown in The Model Comes Last.

Pays off when: the data is structured and several people need evaluations regularly. Without a clean data basis, a language interface only moves the problem. Once that basis exists the rollout is short, a matter of hours rather than months. The same data basis is also the precondition for predictive maintenance.

AI agents: when several steps hang together

An AI tool handles one step. An AI agent chains several. It reads the supplier email, checks the order in the ERP, finds the affected production orders and puts the result in front of a person. How we build such agents is described on our AI agents page.

There are no reliable adoption figures for agents, neither for Switzerland nor for DACH manufacturing. The closest official number is Destatis 2025: 26% of AI-using companies with 10 to 49 employees use AI to automate workflows or decision-making. Our own experience comes from our own house: iosys runs its back office on agents and documents the architecture publicly in agent-atlas.

For planning there is a limit we state openly. An agent does not plan. Sequence, capacity and dates are an optimisation problem that a deterministic solver handles better and more transparently than a language model. The agent takes on the work before and around it: gathering unstructured input, spotting the exception, preparing the options. Your planner decides. The next step, where an agent drives the solver itself, is one we will take once it has proven itself in our own operation. That is the staged path we set out in Engineered AI: targeted automation first, then monitoring, then autonomy inside defined boundaries.

An agent gets more reliable where fixed steps are held as a script or a defined skill. The alternative is letting the language model re-derive them on every run, which costs tokens and varies between runs. Code over inference is the principle: what can be expressed as a program belongs in the program. That is the same logic as the split above, where the solver computes and the agent prepares.

The pattern underneath is a division of labour. The model takes the open part, reading intent and drafting a proposal. A fixed program answers the questions that have one right answer. Whether a part number exists, whether two quantities reconcile, whether a date falls inside the agreed lead time. The agent does not judge those, it calls something that checks them. What comes back is either right or a fault you can locate.

The expensive failure is the quiet one, a plausible order confirmation carrying the wrong quantity. A fixed check catches that class before anyone sees it. The same division loosens the tie to one provider, because the more the program carries, the less the result depends on the model above it.

One point that matters at purchasing time: agents connect to third-party systems through the Model Context Protocol. Since December 2025 that protocol has been held by the Agentic AI Foundation under the Linux Foundation. That lowers the risk of tying yourself to a single vendor when connecting to your ERP.

Pays off when: the process is already described and your systems have an interface. Without both, you are building an agent around a problem you do not yet understand.

AI assistants: useful in knowledge work

Translating, drafting and summarising save real time, in manufacturing too. Quotation text, a complaint in a foreign language, a long supplier thread reduced to its point. The catch is checking. Take the result unchecked and you move the effort from writing to correcting.

They do not carry an automated process. There is no connection to your systems and no record of who approved what. The same document can come out differently on a second run. Expect these tools to be in the building already. People bring their private accounts to work, usually without IT knowing. So the question is rarely whether you introduce them, but whether you know which data leaves the building.

Pays off when: a person checks the result. As an automated process, not yet.

The data foundation before any tool

The strongest evidence comes from KfW Research (No. 533, February 2026, reference period 2022 to 2024). Once you control for innovation and digitalisation activity, company size and sector matter only marginally. A company with no digitalisation activity reaches a modelled 19% probability of using AI. With a digitalisation strategy it reaches 35%.

So the groundwork decides, not your size. Here is what that groundwork looks like: at maxon it was a data bridge from the production machines into MES and ERP. At DUAP it was unlocking an undocumented ERP into a Production Control Center with ten live dashboards. Not an AI project, but the precondition for one. The maxon lessons are in iosys bridges the shopfloor data gap for maxon.

On-premises is a real option

Document extraction runs on your own hardware today. Open-weights document models with 3 to 8 billion parameters have been available since autumn 2025. They include DeepSeek-OCR and olmOCR-2 from the Allen Institute for AI. That size class needs no hyperscaler API.

Swiss companies also have Apertus, the fully open language model from EPFL, ETH Zurich and CSCS. It comes in 8B and 70B sizes, with Swiss German. ETH explicitly calls it infrastructure rather than a finished product.

This is a live concern. In the AXA/Sotomo survey of March 2026, among 336 Swiss SMEs, 57% worry about their dependence on mostly US IT solutions. 15% have already moved to European alternatives, and 13% plan to.

The legal position: under 18 U.S.C. § 2713, added by the CLOUD Act in 2018, a provider must disclose data in its possession, custody or control. This applies whether the data sits inside the United States or outside it. A Swiss region run by a provider under US jurisdiction is therefore not automatically beyond US legal reach. This describes legal reach rather than routine access, and it is not legal advice.

We therefore run the solution on-premises, in Switzerland or in the EU on request, to guarantee your data sovereignty and data security.

Pays off when: data residency is the binding constraint. Otherwise not.

What does not (yet) pay off

Running cost. Flat rates are giving way to usage-based billing. What was a fixed monthly figure in the pilot grows with use later. So cost a category only once you can estimate a year of operation.

Machine vision in quality control. The only institutional figure comes from Destatis 2025: image recognition is used by 16% of AI-using companies with 10 to 49 employees. Everything else is vendor-published and contradicts itself, in places by an order of magnitude. Put your own sample parts through a pilot installation before you buy.

The question before any tool decision

A simpler question comes before the selection: is this process standard for you, or is it differentiation? Where it is standard, buy. Where it carries your competitive edge, building pays off. A standard product would pull you back to the industry average there.

That classification belongs before the tool selection. It also explains why a category works for your neighbour and not for you.

Which AI category pays off, and under what conditionCategoryPays off whenDocument extraction in order intakeDocuments arrive regularly and in varying layouts. With five suppliers on stable forms, a template is enough.Natural-language access to production dataThe data is already structured and several people need evaluations regularly.AI agents for recurring exceptionsThe process is already described and your systems have an interface. Agents prepare, people decide.Everyday AI assistants (drafting, translation)A person checks the result. Knowledge work yes, automated process not yet.Data foundation (interfaces, dashboards)Always first. Without it no other category works. Not an AI project, but the precondition for one.On-premises / Swiss or EU operationData location is the binding constraint. Not a default.Machine vision in quality controlRun your own sample parts through a pilot first. Evaluate, do not buy.Pays off todayPreconditionEvaluate, do not buy

Conclusion

No tool list holds for every manufacturing SME. An order of operations does: the data foundation first, then the category with the clearly bounded process, then everything else. For most manufacturing SMEs, that is document extraction in order intake.

We build that automation with factorIQ and connect it to your existing ERP, with no system change. Book a call if you want to know which category will carry first at your company.

---

Sources

• ETH Zurich, Swissmem and Next Industries, AI in the Swiss tech sector: reality lags the hype, 2024, 200 Swiss tech companies - ethz.ch

• Federal Statistical Office of Germany (Destatis), Enterprises using artificial intelligence technologies by employee size class, reporting year 2025 - destatis.de

• AXA Switzerland / Sotomo, KMU-Arbeitsmarktstudie 2025, fielded March 2025, 300 Swiss SMEs - axa.ch

• AXA Switzerland / Sotomo, KMU-Arbeitsmarktstudie 2026, fielded 11 to 18 March 2026, 336 Swiss SMEs with five or more employees - axa.ch

• Bitkom, Industrie 4.0: 42 Prozent der Unternehmen setzen KI in der Produktion ein, March 2025, 552 industrial companies with 100 or more employees - bitkom.org

• KfW Research, Einsatz von Künstlicher Intelligenz vor allem in Unternehmen mit hohen Innovations- und Digitalisierungsaktivitäten, Fokus Volkswirtschaft No. 533, February 2026 - kfw.de

• Ouyang et al., OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations, CVPR 2025 - arxiv.org/abs/2412.07626

• ETH Zurich, EPFL and CSCS, Apertus: a fully open, transparent, multilingual language model, September 2025 - ethz.ch

• DeepSeek-AI, DeepSeek-OCR: Contexts Optical Compression, October 2025 - arxiv.org/abs/2510.18234

• Poznanski, Soldaini and Lo (Allen Institute for AI), olmOCR 2: Unit Test Rewards for Document OCR, October 2025 - arxiv.org/abs/2510.19817

• Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation, 9 December 2025 - anthropic.com

• iosys, agent-atlas: architecture snapshots of our own multi-agent operation - github.com/ryaniosys/agent-atlas

• 18 U.S. Code § 2713, Required preservation and disclosure of communications and records (CLOUD Act, 2018) - law.cornell.edu

• iosys case study, From 20 to 2 Hours: AI-Powered Order Processing - Read case study

• iosys case study, Shopfloor <-> ERP: Real-Time Data Exchange - Read case study

• iosys case study, Live Shopfloor Data from a Legacy ERP - Read case study

• Simon Spätli, Where AI Agents Belong in Data Engineering: The Correctness Layer, July 2026 - ssp.sh

• The Generative Programmer, 9 Principles That Separate Useful Skills from Markdown Essays, June 2026 - generativeprogrammer.com

• iosys blog, How to Get AI Going in Manufacturing - Read article

• iosys blog, Engineered AI: How Process Automation Redefines Manufacturing in 2026 - Read article

• iosys blog, Ask Your Machines: Natural Language Access for Your Factory Data - Read article

• iosys blog, AI Agents in Order Intake: What Works Today - Read article