Can AI be integrated with SAP, MES or other existing systems?

Yes. SAP S/4HANA has OData APIs, MES systems usually OPC-UA. We build pre-processing steps in front of these systems — AI extracts or classifies, the system receives structured data as always. We do not touch core transactions.

What about data sovereignty for sensitive production data?

Construction drawings, machine parameters and supplier contracts usually do not belong in a US cloud. For such use cases we use on-premise models (Llama, Mistral) or EU hosting. For uncritical tasks (general translation, copywriting) frontier cloud is fine.

Industry — Manufacturing

AI in industry and manufacturing

Q: What is a realistic initial investment?

For one productive use case in a mid-sized company (50 to 500 employees) typically 25,000 to 70,000 € first-project cost — depending on data quality and integration depth. BAFA funds the consulting phase, ZIM or go-digital fund parts of implementation. We help with funding combinations.

Specifications, maintenance knowledge, multilingual supplier chains — the areas where actually takes load off manufacturing SMBs. Five concrete , with an honest dividing line to predictive maintenance and auto-control.

The daily reality AI has to fit into

Order volume volatile, specifications twenty-four hours old and already in the mail pile. Engineering and costing work on different versions because no one has read the entire spec first. Service request on machine 17 at 2 am — the colleague still remembers the fault from 2022 but can't how it was solved.

in manufacturing in 2026 is first and foremost a tool for the quoting phase, maintenance knowledge and supply-chain communication. Control and plant technology remains classical — what pays off are the recurring writing and search tasks that today slow down costing, sales and service.

Prerequisite in all cases: plugs into ERP, CAD and MES, it doesn't replace them. The biggest levers lie before and after the order, not on the shop floor itself.

Five places where AI in manufacturing really makes sense

For each case: the situation today, where plugs in, a pointer to the matching setup tiers, and an honest trade-off.

01. Quote preparation from spec or RFQ

Situation today

A customer inquiry with a spec arrives as a PDF — sixty pages with requirements, tolerances, quantities, delivery terms. Sales needs the engineering briefing, engineering needs the quantity profile, costing needs the rest. Four people read the same document, each pulls different information.

Where AI helps

extracts structured specifications from the spec (dimensions, tolerances, quantities, material requirements, delivery terms) and creates an inquiry entry in the ERP — with references to the source passages in the original PDF. Sales, engineering and costing work on the same structured base.

Matching setup tiers

→ Document processes — structured extraction from technical PDFs

What it can't do

Costing remains an engineering task. delivers the structured briefing, it doesn't cost — and with complex tolerances, the engineer has to read over it anyway.

02. Maintenance and service knowledge base

Situation today

Manufacturer docs in a cabinet, own maintenance histories in an Excel sheet, spare-part data sheets on the server, undocumented tips in the heads of two service technicians retiring in three years. When a plant stops at night, people phone rather than search.

Where AI helps

Your own knowledge base with manufacturer docs, maintenance histories, spare-part data sheets and the service technicians' experience notes — semantically searchable. “When did we have the 3-phase fault on machine 17?” finds the entry from 2023.

Matching setup tiers

→ Knowledge management — hybrid retrieval with vector and full-text

What it can't do

Lives on service discipline. Without notes, you have a nice search over old data sheets — the value only comes when the experience notes really land.

03. Multilingual supplier and customer correspondence

Situation today

Supplier in Italy sends a delivery reply in Italian, customer in the Czech Republic asks about a tolerance detail in English, the engineer replies in German. Translation ping-pong via DeepL, copied three times, one mistake slips through.

Where AI helps

translates incoming correspondence, summarises it in a structured way and suggests a reply draft — in the recipient's language. The employee reviews and signs off, the original stays in the file with both languages.

Matching setup tiers

→ Back office — mail routing and pre-classification

What it can't do

For contracts and complaints, a human must read over the translation — makes tone and nuance errors that can cost money in a complaint.

04. Incoming mail routing across sales, service and purchasing

Situation today

The sales catch-all inbox gets RFQs, complaints, service requests, supplier replies and three newsletters a day. The assistant sorts, the sales lead resorts, service has its own inbox separately — and the RFQ from Thursday gets noticed on Monday.

Where AI helps

classifies incoming mails (RFQ, complaint, service, supplier, newsletter), files them in the right ERP or CRM inbox and suggests follow-up tasks — with prioritisation for RFQs with a deadline.

Matching setup tiers

→ Back office — email routing and pre-classification

What it can't do

No automatic ordering or shipping release. Classification as suggestion, human confirms — otherwise you risk an RFQ being classified as newsletter and lost.

05. Audit and standards research

Situation today

ISO 9001 re-certification is due, the auditor asks for the documented procedure for supplier evaluation. The QMR works through three folders because the procedure manual isn't in sync with real processes. The standards database sits on the shelf.

Where AI helps

AI-supported knowledge base with ISO/DIN/IATF standards, your own procedure documents and audit reports — semantically searchable. “What do our procedure instructions say about supplier evaluation compared to the ISO 9001 requirement?” delivers the comparison.

Matching setup tiers

→ Knowledge management — self-hosted RAG

What it can't do

Legal and audit-relevant assessment stays with the QMR and auditor. delivers preparation, no procedure manual and no audit response.

What in manufacturing isn't (yet) worth it

Four promises that burn more money than they save in a typical manufacturing SMB:

Predictive maintenance without sensor history

Predicting machine failures works with historical sensor data over years. Anyone with only a handful of machines or who would first have to retrofit sensors invests better in clean maintenance plans and manufacturer service than in an prediction layer.

Generative design in small businesses

AI-supported CAD generation is a highly specialised market in 2026 with its own tools and five-figure licence costs. For classic job-shop or series production, the entry rarely pays off — established CAD workflows are more results-oriented here.

AI plant control via LLM

Control and PLC logic does not belong in a language model. Machinery directive, functional safety, IEC 61508 set clear limits — and established control solutions deliver more, with clear liability.

Fully automatic dispatch

Operations-research methods usually beat significantly in classic dispatch. can deliver a briefing on bottlenecks and exceptions, the planning itself runs better in existing ERP and MES modules.

What needs to be thought through for AI in manufacturing

Four pillars against which every industrial setup is checked:

ISO 9001, ISO 14001, IATF 16949

Quality and environmental management standards require documented procedures and traceability. workflows must deliver audit trail and be integrated into the existing QM system — otherwise shadow processes emerge that surface in the next audit.

Machinery directive / regulation 2027

With the new EU Machinery Regulation (application from 2027), stricter requirements apply, especially for components in safety-relevant functions. Anyone planning here must factor in conformity — not later.

Product liability and export control (dual-use)

Products with parts can need to be reassessed under product liability law. For export-controlled goods (dual-use), strict requirements apply anyway, components further shift the picture.

GDPR on employee and supplier data

Personal data in supplier or customer correspondence falls under GDPR. DPA with the vendor, EU region and clear data-domain separation are mandatory setup, not bonus.

Tools that already run in manufacturing

doesn't replace these systems — it plugs in. Where the interface typically sits:

ERP systems

SAP Business One, abas, proAlpha, Sage 100, Microsoft Dynamics 365 Business Central, Odoo — typically plugs in via REST , file import/export or middleware

CAD and PLM

AutoCAD, SolidWorks, Inventor, Creo (CAD); Teamcenter, Windchill, PRO.FILE (PLM) — typically processes PDF exports or bills of materials, no native CAD manipulation

MES and shop floor

Hydra (MPDV), COSCOM, gfos, FORCAM — briefings plug in via report exports or MES , not into real-time control

Accounting and interfaces

DATEV export from ERP, ZUGFeRD/X-Rechnung for B2B invoices — typical plug-in points for receipt workflows

How manufacturers typically get started

Anyone starting without experience has two clear candidates — and an area where the spectacular promises should be ignored for a long time.

Typical entry point 1 — multilingual correspondence

Low risk, immediate effect in sales and purchasing. Prerequisite: a maintained catch-all inbox, clear rules on who reads over which translation.

Typical entry point 2 — quote preparation from RFQ

Biggest lever on sales and costing, especially with recurring inquiry patterns. Needs some preparation (which fields get structured), but pays off on every RFQ.

Don't start on the shop floor

Predictive maintenance, control, auto-dispatch — the most spectacular promises, at the same time the most expensive dead ends without a data base. First clean up admin and correspondence workflows, then maybe move on.

Funding for manufacturing SMBs

consulting funding covers the conception and introduction phase. For investment-heavy innovation projects, ZIM (Central Innovation Programme for SMBs) or BMWK funding come into play — especially for projects with a development part. Alongside there's “go-digital” and regional digital-bonus programmes. Mittelstand-4.0 competence centres offer free initial consulting in many regions.

→ Details on BAFA funding

FAQ

Frequently asked questions about AI in industry

Classically in four areas: incoming documents (orders, delivery notes, certificates), maintenance documentation (classifying damage photos, finding repair instructions), quality inspection (image data) and sales inquiries (quotes from RFQs). Predictive maintenance is often harder because clean sensor data is missing.

Yes. SAP S/4HANA has OData APIs, MES systems usually OPC-UA. We build pre-processing steps in front of these systems — extracts or classifies, the system receives structured data as always. We do not touch core transactions.

Construction drawings, machine parameters and supplier contracts usually do not belong in a US cloud. For such we use models (Llama, Mistral) or EU hosting. For uncritical tasks (general translation, copywriting) frontier cloud is fine.

Not with the hardest task. We recommend a clearly measurable area (e.g. delivery note capture or quote draft from customer inquiry), pilot in one department, measure hour savings and error rate, scale only afterwards. Typically 6 to 10 weeks to a productive pilot.

For one productive in a mid-sized company (50 to 500 employees) typically 25,000 to 70,000 € first-project cost — depending on data quality and integration depth. funds the consulting phase, ZIM or go-digital fund parts of implementation. We help with funding combinations.

Ready for the next step?

Free intro call, no strings attached. In 30 minutes you'll know whether and how AI can help your business.

Book a call Check eligibility