MES or AI Agent: Where the Boundary Runs on the Mittelstand Shopfloor in 2026

At 06:42 on a Tuesday morning, a press in a German Mittelstand metal-stamping plant threw a die-misalignment alarm. The MES did exactly what it was supposed to: logged the event, paused the order, sent an alert to the line supervisor. By 07:10 the supervisor had identified the root cause, but the production schedule still showed the original sequence. Three downstream cells continued setting up for parts that now would not arrive on time. By the end of the shift, four customer deliveries were at risk - none of them flagged anywhere outside the supervisor’s head.

That is not an MES failure. It is an MES being used for what it was never built to do. The MES recorded the event flawlessly. What it cannot do is reason across the schedule, the customer order book, the on-time-delivery commitments, the alternate routings, and the supervisor’s in-progress thinking. That is the work of an AI agent - and confusing the two is what wastes the next round of Mittelstand factory IT budget.

This article is not about whether MES is dead (it is not). It is about understanding precisely what each tool is built for, where the boundary runs, and what the hybrid architecture actually looks like in a real Mittelstand plant in 2026.

Why This Question Hits Hardest in Manufacturing

MES adoption in the German Mittelstand is patchy but mature. Some hidden champions run state-of-the-art MES platforms (MPDV Hydra, Industrie Informatik cronetwork, GFOS, Siemens Opcenter, AVEVA, Werum). Others run 15-year-old systems that nobody fully understands but nobody dares to replace. Either way, the MES sits at the centre of shopfloor IT - and the question of whether AI complements it or replaces it is now genuinely confusing for buyers.

The forces shaping the question right now

• MES vendors have all added AI - Every major MES vendor in 2026 markets some form of AI: anomaly detection, predictive scheduling, vision-based quality. The features are real but bounded. They do not solve the cross-system, exception-heavy decisions that define Mittelstand production reality.
• AI vendors claim MES territory - Some AI startups pitch “AI-native MES” that will replace traditional systems. The pitch ignores the regulatory, traceability, and ERP integration weight that real MES carries. Replacing a working MES is a multi-year project nobody should undertake on a vendor pitch.
• The data gap is the real bottleneck - Gartner forecasts that 60 percent of AI projects will be cancelled by end of 2026 due to inadequate data foundations⁴. 70 percent of manufacturers identify data quality as their biggest implementation obstacle⁴. The MES is usually where that data lives.
• The Mittelstand is under-invested - Horvath research shows mid-sized companies invest just 0.35 percent of revenue in AI, 30 percent below the global average⁴. VDMA finds fewer than 20 percent of German Mittelstand manufacturers have moved Industry 4.0 past pilot stage⁴. The instinct is to consolidate, not to multiply tools.

The German Mittelstand context

German manufacturing carries specific weight that makes the MES versus AI agent question different from a US or Asian conversation. The decisions you make have to fit five concurrent realities, not just one technical preference.

• SAP everywhere - Most German Mittelstand factories run SAP at Level 4 (S/4HANA, ECC, or Business One). The MES sits at Level 3 and integrates upward through ISA-95-aligned interfaces. Any AI agent that wants to add value has to respect this hierarchy.
• Mittelstand-specific MES - Industrie Informatik, MPDV, GFOS, iTAC dominate the Mittelstand mid-tier. Not Siemens-class enterprise MES, not light-touch tools. They encode decades of German industrial process logic. Replacing them is rarely the right move.
• Audit and traceability burden - Automotive (IATF 16949), pharma (GMP), food (HACCP), aerospace (AS9100) all require traceable, auditable production records. The MES is the audit trail. AI agents must respect it, not bypass it.
• Betriebsrat co-determination - Any system that monitors employee behaviour or performance is co-determination territory. Production AI agents that stay aggregate (machine, line, shift) are easier to deploy than agents that score individual operators.
• Skills shortage and demographics - 40 percent of German manufacturers cannot find AI-qualified workforce⁴. The architecture has to fit a small in-house IT team, not a 200-person enterprise IT function.

What an MES Actually Is

MES (Manufacturing Execution System) is the software layer between ERP and the shopfloor. ISA-95, the international standard for enterprise-control system integration, places MES at Level 3 - sitting above SCADA and PLC (Levels 1 and 2) and below ERP (Level 4). Its job is to translate ERP-scheduled production orders into machine-level execution and capture the result.

The defining characteristics

• Transactional core - Records production work order by order, station by station, lot by lot. The MES is the source of truth for what was actually produced - quantities, qualities, times, scrap, rework.
• Rule-based dispatching - Decides which work order goes to which machine based on configured logic. Optimisation is bounded by the rules; novel situations need human override.
• Machine data acquisition - Connects to PLCs, SCADA, and sensors to pull live machine state, OEE inputs, alarm conditions. This is the data layer most AI agents read from.
• ISA-95 integration to ERP - Pushes production results to ERP for back-end accounting, inventory, and financial reporting. Pulls planned orders down from ERP for execution.
• Traceability and audit - Maintains the genealogy of every lot, every batch, every part - which machines made it, which operators ran the shift, which quality results applied. Compliance-critical for regulated industries.
• OEE and KPI calculation - Calculates Overall Equipment Effectiveness, throughput, yield, and other KPIs from the data it captures. The numbers feed dashboards, leadership reports, and improvement projects.
• Operator UI on the shopfloor - Touchscreens at each station show the current order, capture confirmations and rework, surface alarms. The MES is the daily tool of every operator and supervisor.

Where the MES is genuinely brilliant

• Compliance-grade traceability - When the customer audit asks “which press, which shift, which operator made this part on this date with which raw material?”, the MES has the answer in seconds. Building that on top of an AI agent would be reinventing audit infrastructure.
• Real-time machine-level dispatching - Telling Press Line 3 which die change comes next, capturing the changeover start and end, dispatching the order to the operator screen. Deterministic, fast, fits the rhythm of the shopfloor.
• OEE calculation under load - The MES does the boring, structured calculation work that produces the daily OEE number. The number can be wrong if the data is wrong, but the calculation itself is solid.
• ERP integration - Production confirmation back to SAP, inventory updates, cost centre charging. ISA-95-aligned MES handle this through stable API contracts.
• The 24/7 system of record - The MES runs whether a planner is at her desk or not. It does not need a human in the loop for every transaction. That is exactly right for the structured, rule-based work it handles.

Where the MES hits its design limits

• Cross-system reasoning - The MES sees production data. It does not see customer emails, supplier delivery notifications, contract penalty clauses, ERP credit holds. Decisions that require this context fall through the gap.
• Unstructured input - An operator notes “Material seems off” in free text. A customer emails a delivery change. A supplier sends a PDF announcing a quality issue. The MES cannot read any of it.
• Exception handling beyond defined rules - The schedule says A then B then C. Reality breaks: the customer of B postpones, the customer of D has an urgent ad-hoc order. The MES needs a human to redo the schedule. There is no “just figure it out” option.
• Document-heavy back-end work - 8D reports, customer complaint responses, supplier corrective action requests, internal audit prep. The data is in the MES. The work is not the MES’s job.
• Cross-departmental orchestration - When production, sales, logistics, and quality all touch the same problem, no single MES coordinates them. The orchestration falls to the supervisor or planner who happens to be on shift.

What an AI Agent in Production Actually Is

An AI agent in a manufacturing context is not a chatbot, not a dashboard, not an MES replacement. It is a goal-directed reasoning layer that reads from the systems already in place (MES, ERP, CAQ, CRM, document systems, email), reasons about what to do toward a defined production goal, and either acts within bounded authority or escalates to a human with full context.

The defining characteristics

• Goal-directed, not script-directed - You define the outcome (rebalance the schedule, draft the 8D, predict the lot quality risk, escalate the customer impact). The agent figures out how to get there using available tools.
• Reads across systems - MES, ERP, CAQ, customer email, supplier portal, document management. The agent reasons about the full context, not just MES data.
• Handles unstructured input - PDFs, emails, free-text operator notes, contract clauses, supplier announcements. About 80 percent of enterprise data is unstructured - and most of it is invisible to the MES.
• Sits above the MES, not beside it - The MES stays as system of record. The agent reads from it and writes back through proper APIs. No bypassing, no replacing.
• Respects ISA-95 hierarchy - Production AI agents run at Level 4 (or as a layer between 3 and 4). They do not write directly to PLCs or machines. The MES remains the only system that talks to the shopfloor execution layer.
• Escalates with context - When the agent encounters a decision outside its authority or confidence, it escalates to a human with what it found, what it tried, and what it recommends. Reviews are 5 to 10 times faster than handling the raw exception alone.
• Pricing by use case, not by user - Agents are typically priced per active production use case (per agent, per process), not per shopfloor seat. The economics fit a small team handling high volume.

What an AI agent in production is not

• Not a chatbot for shopfloor - Operators do not need a chat window. They need work to flow. Agents act behind the MES UI rather than competing with it.
• Not a replacement for SCADA, MES, or PLC - The execution layer stays. Agents add the reasoning layer above.
• Not RPA on top of the MES - RPA scrapes UIs and breaks on every update. Agents integrate through APIs, reason about goals, and handle exceptions natively.
• Not magic - Gartner predicts more than 40 percent of agentic AI projects will be cancelled by end of 2027¹², mostly due to inadequate governance and unrealistic scope. Production AI needs human-in-the-loop checkpoints, audit logs, and rollback paths.

Six Differences That Matter on the Shopfloor

These differences are not philosophical. They show up in how each system behaves at 03:00 when something breaks, how procurement evaluates them, and what happens during an audit.

Difference 1: What it is for

The MES exists to record and dispatch production work according to defined rules. It is the system of record for what happened on the shopfloor. The AI agent exists to reason about decisions across the data the MES collects plus context the MES cannot see. The MES owns the truth. The agent owns the reasoning.

Difference 2: How it makes decisions

The MES decides through configured rules. If the order priority is X and the available machine is Y, dispatch to Y. The rules are predictable, deterministic, auditable. The AI agent decides through goal-directed reasoning. Given the goal of meeting customer commitments and the current state across systems, what is the best next action. Different paradigm, different outcomes on novel situations.

Difference 3: What data it sees

The MES sees what shopfloor sensors and operator inputs feed into it. Structured, machine-generated, station-specific. The agent sees the MES data plus everything else: ERP, CRM, customer email, supplier portal, contracts, free-text notes. The decision quality depends on the breadth of context, not just on the algorithm.

Difference 4: How it handles exceptions

The MES routes exceptions to the supervisor or planner. The exception queue grows in proportion to the variability of the production environment. The agent handles exceptions inline: it reasons about what happened, proposes a resolution, and either acts within defined authority or escalates with context. The supervisor reviews the agent’s analysis, not the raw exception.

Difference 5: Audit and traceability

The MES is the audit source of truth. Every part has a genealogy in the MES. The agent logs every decision it makes - what it found, what it considered, what it decided, what it escalated - but the production record itself remains in the MES. The agent creates an additional layer of decision audit, which is often valued during regulatory reviews.

Difference 6: How fast it deploys

An MES greenfield deployment takes 6 to 18 months and several million euros in a Mittelstand context. A first focused AI agent on top of an existing MES takes 8 to 12 weeks and 50,000 to 200,000 euros. The economics flip: MES is a long-cycle investment in infrastructure; agents are a short-cycle investment in specific outcomes.

Where Each One Wins

The framing “MES or AI agent” sets up a false binary. The real question is process-by-process: which one fits which job. The same Mittelstand plant will rely on its MES for some work and lean on an AI agent for other work - and the boundaries will run cleanly between them when both are deployed deliberately.

Where the MES wins clearly

• Work order dispatching - Right work order, right machine, right time. Deterministic, structured, audit-critical. The MES is built for this.
• Production confirmation back to ERP - When a lot is finished, the MES posts the result to SAP for inventory and accounting. Stable, regulated, no judgement involved.
• Lot genealogy and traceability - When a customer audit or recall investigation needs to trace a specific part backward through every machine, every shift, every raw material, the MES is the answer.
• Real-time machine data acquisition - Reading PLC and sensor data, calculating OEE inputs, surfacing alarm conditions. Latency-critical, deterministic, machine-level.
• Operator UI on the shopfloor - The touchscreen at each station, the order list, the rework reason capture. Built for the rhythm of production work.
• Shopfloor compliance documentation - GMP batch records, automotive PPAP traceability, food HACCP logs. Regulated, structured, audit-ready.
• SCADA-to-ERP integration under ISA-95 - The standardised, audit-aligned bridge between business systems and shopfloor execution.

Where an AI agent wins clearly

• Exception handling beyond defined rules - When the schedule breaks because reality does not match the plan, an agent reasons about the right resolution across MES, ERP, customer commitments, and supplier status.
• Schedule rebalancing across cross-system context - Customer urgency, supplier delivery, capacity availability, due-date pressure - the agent integrates inputs the MES does not see and proposes a new sequence with rationale.
• Predictive quality risk on a per-lot basis - Reading sensor data plus material lot history plus operator notes plus supplier quality history to flag a lot for inspection before it ships. The MES has the data; the agent has the reasoning.
• 8D report drafting from MES, CAQ, and engineering data - Pulling defect data, root-cause analysis from CAQ, supplier history from ERP, drafting the 8D for the quality engineer to review. Saves days per case.
• Customer complaint response drafting - Reading the customer email, finding the specific lot in the MES, drafting a fact-based response with affected quantity, root cause, and corrective action.
• Supplier complaint and SCAR letter drafting - Compiling production impact data, generating Supplier Corrective Action Requests with attached evidence.
• Audit prep across MES, CAQ, and engineering systems - Pulling the right records for an upcoming customer or certification audit. Hours instead of weeks of manual collation.
• Operator query handling - When an operator asks “what is the spec for this part?” or “why was the last setup changed?”, the agent reads engineering, MES, and quality data to answer in seconds.
• Cross-shift handover summarisation - End-of-shift summary covering open issues, in-progress orders, machine status, escalations, drafted automatically from MES and operator notes.

Where neither is the right answer

• Broken processes - If the production process is poorly designed, no software fixes it. Map and improve the process first, then deploy the right tool.
• Genuinely safety-critical control loops - Real-time machine safety, emergency stops, pressure relief. PLC and safety systems own this layer. AI does not belong here.
• Tiny volume, low value processes - Three setups a year for a low-margin product. Spreadsheet plus checklist beats automation.
• Politically blocked processes - When the friction is between departments, not in the work. Tooling makes politics worse, not better.

The Cost Comparison Done Honestly

Comparing MES cost to AI agent cost is the wrong frame. They solve different problems. The honest comparison is the cost of the gap each one leaves and the work each one removes.

What an MES actually costs in the Mittelstand

• Licence model - Per station, per module, sometimes per concurrent user. A typical Mittelstand mid-tier MES costs 1,500 to 3,500 euros per station for the base licence, plus modules for OEE, quality, traceability, document control.
• Implementation - The dominant cost line. 250,000 to 1,500,000 euros for a Mittelstand greenfield MES depending on plant complexity, number of lines, and modules. Often takes 9 to 18 months.
• Annual maintenance - 18 to 22 percent of licence value per year for support and updates, plus internal admin time.
• Integration to ERP and engineering - 50,000 to 250,000 euros for SAP integration, PDM/PLM connections, CAQ links. Sometimes ongoing as systems update.
• Operator training and rollout - 30,000 to 80,000 euros depending on operator headcount and shift coverage.
• Periodic upgrade projects - Major version upgrades every 5 to 8 years, typically 30 to 50 percent of original implementation cost.

What a production AI agent actually costs

• Platform / agent fee - 2,000 to 8,000 euros per month per active production use case, depending on complexity, volume, and integration scope.
• LLM inference cost - Cents per task. Even at thousands of tasks per day, total inference cost stays modest - low hundreds to low thousands of euros per month per use case.
• Implementation - 40,000 to 120,000 euros for a focused first deployment. Mostly process mapping, MES API integration, validation against historical cases.
• Integration maintenance - Lower than RPA or MES integration because agents tolerate API drift better. Typically 5,000 to 15,000 euros per year per active use case.
• Monitoring and feedback loop - 0.2 to 0.5 FTE per active agent for review, correction, and continuous improvement.
• No per-station scaling - The number of stations or operators does not change the agent cost. Per-process pricing fits per-process value.

Three-year total cost on a representative use case

Take an exception-handling use case: schedule rebalancing when a machine fails. Today, the supervisor on shift spends 30 hours per week resolving these cases manually across MES, ERP, customer order book, and shop emails. The MES alone cannot help (it routes the alarm; reasoning is the supervisor’s job). The agent reasons across systems and proposes resolutions for supervisor approval.

Where the cost comparison flips against an agent

• Stable, low-exception production - When the production environment is genuinely stable and the MES alone handles it cleanly, an agent adds cost without value.
• Very low volume processes - When the supervisor work being saved is a few hours per month, agent platform fees do not amortise.
• Bad data foundations - When the MES data is wrong or missing, the agent cannot reason reliably. Fix the data first; deploy the agent on the cleaner subset.
• Pure machine-control problems - PLC, SCADA, real-time control. Agent does not belong in this layer.

The MES + Agent Architecture That Works

Almost every Mittelstand factory that adopts production AI agents lands in a hybrid: MES stays as system of record, agents sit on top. The architecture is not a compromise; it is the only configuration that respects ISA-95, regulatory traceability, and the operational reality of the shopfloor.

The four-layer stack

• Layer 1: Machines, sensors, PLCs, SCADA - The physical execution layer. Real-time, safety-critical, deterministic. AI does not belong here.
• Layer 2: MES - The system of record for production. Captures data, dispatches work, records traceability, integrates upward to ERP. Compliance-grade.
• Layer 3: AI agents - The reasoning layer. Reads from MES, ERP, CAQ, CRM, document systems. Reasons about goals. Writes decisions back to MES through proper APIs.
• Layer 4: ERP and business systems - SAP, financial systems, customer order book, supplier portal. The MES integrates here under ISA-95. Agents read from here as additional context.

The most common hybrid patterns we see in the Mittelstand

• MES + agent for schedule rebalancing - MES dispatches and tracks. Agent monitors disruptions (machine failure, customer change, supplier delay) and proposes new sequences for the planner to approve.
• MES + agent for predictive quality - MES captures sensor and operator data per lot. Agent reads sensor history plus material lot data plus supplier quality history and flags risky lots for inspection before they ship.
• MES + agent for 8D and audit work - MES holds the production records. Agent pulls relevant data, drafts 8D reports, prepares audit binders, drafts customer communications. Quality engineer reviews and approves.
• MES + agent for shift handover - MES has the operational state. Agent summarises the shift end-to-end, surfaces open issues, drafts the next-shift briefing. Saves 20 to 40 minutes per shift.
• MES + agent for operator support - MES has the order, the spec, the recent events. Agent answers operator questions in seconds (“why did the setup change?”, “what is the tolerance for this feature?”) by reading MES + engineering systems.

How the responsibilities split in a hybrid architecture

The Decision Framework

Use this framework job by job to decide whether your existing MES is enough or whether an AI agent layer adds clear value. The output is a defensible recommendation you can present to operations leadership and IT in the same meeting.

Step 1: Classify the job

• Execution job - Dispatch a work order, capture a confirmation, record a measurement, route a part. Structured, repeatable, audit-critical. MES territory.
• Reasoning job - Decide what to do when reality breaks the plan, draft a document from cross-system data, predict a risk, summarise a shift. Cross-system, judgement-heavy, exception-prone. Agent territory.

Step 2: Score the data scope

• MES data only - The decision needs only what the MES already has (machine, lot, sensor, time). MES rules can probably handle it.
• Cross-system context needed - The decision needs MES plus ERP, CRM, customer email, supplier portal, contract terms, engineering documents. Agent territory.

Step 3: Score the variability

• Low variability - The case looks the same every time. MES rule logic is sufficient.
• High variability - The cases differ in small but meaningful ways. Rule maintenance becomes a permanent burden. Agent reasoning is more durable.

Step 4: Score the cost of delay

• Low cost of delay - The work can wait for the supervisor or planner to handle it next morning. No urgency.
• High cost of delay - Customer SLAs, line downtime, lot-at-risk decisions. Speed matters, and an agent that reasons in seconds beats a queue waiting for human attention.

Step 5: Read the decision matrix

How Superkind Fits

Superkind builds custom AI agents that sit on top of existing Mittelstand MES, ERP, CAQ, and engineering systems. We do not replace your MES. We build the agent layer that does what your MES was never built to do - reason across systems, handle exceptions, draft documents, escalate with context.

Core capabilities for production environments

• MES-native integration - APIs and data connectors for Industrie Informatik, MPDV Hydra, GFOS, iTAC, Siemens Opcenter, AVEVA, Werum, and SAP MII. We connect through stable interfaces, not UI scraping.
• SAP-native context - SAP S/4HANA, ECC, MII, PP/PI, QM, MM. Agents read production orders, customer commitments, supplier status, quality records as context for shopfloor decisions.
• CAQ and engineering integration - Babtec, Böhme & Weihs, iqs, plus PDM/PLM systems. Agents pull engineering specs, defect history, FMEA notes for quality and audit work.
• Document intelligence - Reads operator notes, customer emails, supplier announcements, contract clauses. No template configuration required.
• Cross-system orchestration in one workflow - A single agent reads MES + SAP + CAQ + email and produces a single decision or document. No swivel-chair work for the supervisor.
• ISA-95 alignment - Agents run at Level 4 or at the Level 3-4 boundary. They never write directly to PLCs or machines. The MES remains the only system that talks to the execution layer.
• Human-in-the-loop checkpoints - You define which decisions require approval and at what confidence threshold. Agents escalate with context rather than acting silently. Critical for high-stakes shopfloor decisions and EU AI Act alignment.
• Audit trail layered on top of MES - Every agent decision is logged: what it found, what it considered, what it decided, what it escalated. The agent log complements the MES audit trail rather than replacing it.
• EU deployment and DSGVO alignment - Agents run on EU cloud or your own infrastructure. Data does not leave your defined perimeter.
• 8 to 12 weeks to first production deployment - From process assessment to live operation on a focused first use case. No multi-year transformation.

Superkind vs alternative paths

The 90-Day Plan

This plan covers running the decision framework on a single shopfloor use case, validating MES data, deploying the agent in limited scope, and reaching first production value. Use it to align your operations leadership and IT.

Weeks 1 to 3: Use case selection and data audit

• Pick three candidate use cases - Each one with measurable pain (supervisor hours, missed SLAs, audit prep weeks, scrap escalations). Document current pain in numbers.
• Apply the decision framework - Score each use case on job type, data scope, variability, and cost of delay. Output: one use case clearly fit for an agent.
• Audit MES data quality - For the chosen use case, validate that the data the agent will reason on is clean enough. Most agent failures trace back to bad data.
• Confirm API access - MES vendor APIs (REST, SOAP, OPC-UA, file-based) for the data the agent needs. Engineering, CAQ, ERP integration confirmed.
• Define success metrics - Hours recovered per week, exception resolution time, audit prep duration. Numbers measurable in 90 days.
• Brief Betriebsrat where applicable - If the agent could touch employee performance data, start consultation early.

Weeks 4 to 8: Build and test

• Process mapping detailed - Document inputs, outputs, decision points, system touches, exception types. The work that makes the deployment succeed.
• Agent build against the process map - Prompt and tool design, integration setup, escalation thresholds, human-in-the-loop checkpoints.
• Test against real historical cases - Pull past examples (resolved by the supervisor), run the agent against them, compare outputs.
• Validate exception handling specifically - The hardest cases are the agent’s real test. Confirm escalations deliver useful context.
• Confirm DSGVO logging - Audit trail captures what the agent did, why, and what data it accessed.
• Train the team - Supervisors and planners learn how to review agent recommendations, correct errors, and feed back.

Weeks 9 to 12: Production and learning

• Deploy to limited scope - 20 percent of cases, one product family, one shift. Run in parallel with the existing process.
• Review every escalation and correction - Weekly cadence. What did the agent get wrong? Why? What is the correct answer?
• Measure against baseline - Hours recovered, resolution time, audit prep speed. If numbers are not moving, diagnose before scaling.
• Expand once metrics validate - Two to three weeks of stable operation at limited scope before scaling to full volume.
• Document lessons for the next use case - Where was the framework right, where was it wrong, what would you do differently.

Related Articles

• Standard Software or an AI Agent: How the Mittelstand Should Choose Where Software Budget Goes in 2026
• RPA vs AI Agents: What German SMEs Get Wrong About Automation
• AI Agents in Manufacturing: How German Manufacturers Cut Downtime, Defects, and Costs
• AI-Powered Production Planning: When APS Systems Hit Their Limits
• Predictive Maintenance for Hidden Champions: From Sensor Data to an Autonomous Maintenance Agent

Frequently Asked Questions

Henri Jung

Co-founder at Superkind

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.