Standard Software or an AI Agent: How the Mittelstand Should Choose Where Software Budget Goes in 2026

A German precision engineering company sat through five SaaS demos in three weeks. CRM, configure-price-quote, document management, supplier portal, sales enablement. Each vendor promised the same thing: a faster sales process. After the fifth demo the managing director asked his sales lead a single question - which of these would actually let our team stop copying data between systems? The answer was none of them. Each one was its own island. The team would still copy data, just into more screens.

That is the trap most Mittelstand companies fall into in 2026. The instinct when a process hurts is to look for software that solves it. The catalogue keeps getting longer - companies with more than 1,000 employees now run an average of 177 SaaS applications⁷. The result is not less work. It is the same work, distributed across more interfaces, with more licence costs and more integration debt.

Standard software and AI agents solve fundamentally different problems. Confusing them is expensive. This article gives you the framework to choose correctly the first time - and to recognise when the right answer is not standard software at all.

Why the Question Matters Now

For 25 years the answer to a process problem was “buy software for it”. That instinct produced the Mittelstand IT landscape we have today: an ERP, a CRM, a document management system, a project tool, a planning tool, a time tracking tool, a payroll system, ten line-of-business applications. Useful in isolation. Painful in combination. AI agents change the question because they sit above all of that and can do the things software was never designed to do.

Three forces are reshaping the buying decision

• SaaS sprawl is a measurable problem - Companies use an average of 106 SaaS applications, and large enterprises run 177⁷. Each new tool adds licence cost, integration overhead, training burden, and switching costs. The marginal value of application number 178 is close to zero.
• SaaS prices are climbing fast - Enterprise SaaS price increases at renewal now run 10 to 20 percent annually as standard, with vendors like Salesforce and ServiceNow pushing 15 to 25 percent¹². Custom-built alternatives that were uneconomic five years ago are economic now.
• AI agents reach production maturity - Gartner predicts that 40 percent of enterprise applications will feature task-specific AI agents by end of 2026, up from less than 5 percent in 2025¹. The technology is no longer experimental in production deployment terms.

The German Mittelstand context

German SMEs face a specific version of this question. The IT landscape is heavy with industry-specific Mittelstand software (proAlpha, ABAS, Sage, DATEV, Lexware) layered on top of SAP or Microsoft Dynamics. Most processes already have software. The question is rarely “should we buy something for this”. The question is “should we buy yet another tool, or should we make the tools we have actually work together”.

• Existing licence stack is locked in - Pulling out a core ERP or CRM is a multi-year project. Adding to the stack is the path of least resistance even when it is the wrong path.
• Industry-specific Mittelstand software is non-trivial - These tools encode decades of industry know-how. Replacing them with a generic platform usually produces worse outcomes, not better ones.
• Process-level pain is often integration pain - The thing that hurts is not that the CRM is bad. It is that the CRM, the ERP, and the supplier portal do not talk to each other. Buying a fifth tool does not fix the integration. An AI agent often does.

What Standard Software Actually Is

Standard software is a finished product. The vendor designs the data model, the workflow logic, the user interface, and the integrations. Customers pay a per-user fee to use what was built. Customisation is bounded - usually limited to configuration of fields, workflows, and dashboards within the limits the vendor allows.

The defining characteristics

• Pre-defined logic - The behaviour of the software is fixed by the vendor. Your process must adapt to fit. This is not a flaw - it is the design. Standardisation is what makes the per-user economics work.
• Per-seat pricing - Cost scales with how many people use the tool. A 50-person team using a 100-euro-per-user-per-month tool pays 60,000 euros per year regardless of how much value the tool actually generates.
• Configuration, not customisation - You can set fields, build dashboards, define workflows within vendor-allowed boundaries. Anything beyond those boundaries requires custom development - which most SaaS vendors discourage or prohibit.
• Vendor-controlled roadmap - The features you get are decided by the vendor based on what makes sense for their entire customer base. Your specific need only gets built if enough other customers also need it.
• Periodic updates - The vendor releases new versions, sometimes deprecates features you depend on, sometimes changes the UI in ways that break user habits and require retraining.
• Network effects on integrations - Popular standard software has a large ecosystem of pre-built integrations. Less popular tools are isolated and require custom integration work.

Where standard software is genuinely brilliant

The critique of standard software is not that it is bad. It is that companies pick it for the wrong reasons - inertia, vendor familiarity, demo polish - rather than for the reasons it actually wins.

• Commodity processes - Bookkeeping, payroll, time tracking, basic CRM, expense management. These processes are essentially the same across all companies. Reinventing them is wasteful. Standard software wins.
• Compliance-driven processes - Tax filing, HR record keeping, e-invoicing in Germany - these need to follow specific regulatory rules that are documented in standard software. Building this yourself is duplicating audit-tested logic at higher risk.
• Network-effect processes - Email, video conferencing, messaging, identity providers - the value is in compatibility with the rest of the world. Custom alternatives have no network. Standard wins.
• Predictable, structured workflows - When the process is genuinely the same every time and the data is genuinely structured, the configurability of standard software is enough. There is no reasoning required.

The hidden cost layers Mittelstand buyers underestimate

• Implementation cost is 30 to 50 percent of the first-year total - For ERP-class systems, implementation often equals or exceeds the licence cost in year one¹³. Mid-tier SaaS tools have a smaller implementation tail but it is rarely zero.
• Integration cost compounds - Each new tool needs to talk to the others. Integration work is recurring (the integrations break when either side updates) and rarely scoped accurately upfront.
• Switching costs grow over time - 74 percent of SaaS buyers now evaluate switching costs before purchase, up from 47 percent in 2018¹². Once a tool becomes the system of record for any process, removing it costs more than installing it did.
• Per-user pricing punishes growth - As headcount grows, licence cost grows linearly even when the tool is doing the same job. Custom-built alternatives flatten this curve.

What an AI Agent Actually Is

An AI agent is not another app on the shelf. It is a goal-directed system that uses a language model to reason about what to do, calls tools (APIs, databases, document processors, your existing software) to get work done, and produces an outcome rather than a screen. Where standard software gives users an interface to operate, an AI agent operates on behalf of users against the software they already have.

The defining characteristics

• Goal-directed, not script-directed - You define the outcome (close the supplier query, qualify the lead, reconcile the invoice). The agent decides the steps. If one path fails, it tries another. If the case is unfamiliar, it reasons about what to do or escalates with full context.
• Operates across systems, not within one - A single agent can read from your CRM, write to your ERP, query your document management system, and send an email - all within a single workflow. No swivel-chair work for the user.
• Handles unstructured inputs - Emails, PDFs with variable layouts, free-text customer messages, contract clauses. Approximately 80 percent of enterprise data is unstructured - most of it invisible to standard software³.
• Adapts without redeployment - When a new exception type appears or a process detail changes, the agent typically handles it without code changes - because it reasons about each case rather than following a fixed script.
• Pricing tied to value delivered - Most agent vendors price per process, per outcome, or as a flat platform fee - not per user. The cost lines up with the work the agent does, not with how many people watch it.
• Augments rather than replaces - An AI agent does not remove your CRM, ERP, or DMS. It uses them. Your systems of record stay; the work surface above them changes.

What an AI agent is not

• Not a chatbot - Chatbots answer questions. Agents act on goals. A chatbot can tell you what your order status is. An agent can investigate why an order is delayed, contact the supplier, update the customer, and reschedule the delivery.
• Not a workflow tool - Workflow tools (n8n, Make, Zapier) follow fixed rules. They cannot reason about exceptions or unstructured data. They are useful primitives an agent might call - not a substitute for one.
• Not custom code - Custom software needs new code for each new case. An agent uses a language model that handles new cases as they come up. The build profile is closer to a configuration than a from-scratch development project.
• Not magic - Agents are not infallible. They make reasoning errors, particularly on highly specialised domain tasks. Human-in-the-loop checkpoints are essential for high-stakes decisions. Gartner predicts more than 40 percent of agentic AI projects will be cancelled by end of 2027, mostly due to inadequate governance¹¹.

The economic model is different - which is why direct price comparison fails

Standard software is sold per user. An AI agent is sold per process or per outcome. A per-user-vs-per-process comparison is apples to oranges. The right comparison is total cost of work removed.

• Standard software cost - Licence per user x users x months + implementation + integration + ongoing admin. Scales linearly with headcount.
• AI agent cost - Platform fee + LLM inference cost per task + implementation + monitoring. Scales with the volume of work, not the number of people.
• The comparison that matters - Cost per unit of work removed. If a process consumes 30 hours per week of human time across the team, the question is what each option costs per hour of work removed - not what each option costs per seat.

Six Fundamental Differences

The differences between standard software and an AI agent are not surface-level. They are architectural and economic. Understanding them is what makes the buying decision defensible.

Difference 1: The unit of work

Standard software gives you a screen. The user types, clicks, and validates. The work happens because a person operates the tool. If the person is busy, sick, or new, the work waits or gets done badly.

An AI agent receives a goal. It does the work and produces a result. The user reviews exceptions and approves outputs. The work happens whether or not anyone is at a desk.

Difference 2: Who adapts to whom

Standard software has a fixed model. Your data goes into its fields. Your workflow follows its sequence. If your reality does not fit, you either work around the tool or pay for customisation that the vendor allows.

An AI agent is built around your reality. The fields, the documents, the systems, the exceptions - those are inputs to the agent, not constraints it imposes on you. The agent fits the process, not the reverse.

Difference 3: The pricing model

Standard software charges per user. This works well when the value scales with users (more sales reps in a CRM means more deals tracked). It works badly when value scales differently - for example when a process touches many people lightly but the actual work is performed by a small number of specialists.

An AI agent typically charges per process or per outcome. The cost is decoupled from headcount. A team of 5 specialists processing 10,000 items per month and a team of 50 generalists processing the same volume pay roughly the same - because the work is the same.

Difference 4: What kind of data each can handle

Standard software works on structured data: defined fields, expected formats, predictable types. Anything that arrives outside that frame becomes manual work, an exception, or a separate tool.

An AI agent reads emails, extracts data from PDFs with variable layouts, processes contract clauses, classifies free-text messages. About 80 percent of enterprise data is unstructured - and standard software cannot see most of it.

Difference 5: Cross-system reach

Standard software operates within its own boundaries. An ERP knows about ERP data. A CRM knows about CRM data. When a process spans both, someone copies data between them or an integration pipeline handles it (and breaks when either side updates).

An AI agent is multi-system by design. It can read your CRM, check your ERP, look up a document in DMS, and send an email - within a single workflow. The integration is not a brittle pipeline; it is the agent reasoning across tools.

Difference 6: Where competitive advantage can live

You cannot get a competitive advantage from standard software. Your competitors have the same tool. The tool can make you efficient, but it cannot make you different.

An AI agent can encode your specific process knowledge - the way your sales team qualifies leads, the way your service team handles exceptions, the way your buyers evaluate suppliers. That process design is yours. It can become an advantage your competitors cannot copy by buying the same software.

Where Each One Wins

The framing “standard software or AI agent” suggests a binary. In practice the question is process-by-process: which one wins for this specific workflow. The same Mittelstand company will rationally pick standard software for some processes and an AI agent for others - and reject both for processes that should be done differently entirely.

Where standard software wins

• Bookkeeping and accounting - DATEV, Lexware, Xero, QuickBooks all encode tax law and bookkeeping rules. Building this from scratch with an agent is reinventing audit-tested logic with new risk.
• Payroll - Tax tables, social security contributions, garnishments, mandatory reporting. Compliance-heavy, structured, regulated. Standard wins.
• Video conferencing and messaging - Network-effect tools where compatibility with the rest of the world is the whole point.
• Time and attendance - Standard rules, structured input, regulated reporting requirements. Workforce management software is the right answer.
• Basic CRM for small teams - For teams under 10 reps with simple sales motions, lightweight CRM (Pipedrive, HubSpot Starter) outperforms anything custom on cost-to-value.
• Expense management - Receipt capture, approval workflows, GoBD-compliant archiving. Tools like Mobilexpense, SAP Concur, or Spendesk handle this well.
• Identity and access management - Standard authentication, SSO, MFA. Custom IAM is a security risk; pick a vendor.
• HRIS for under 200 employees - Employee records, leave management, document storage. Personio, BambooHR, Workday for larger orgs.

Where an AI agent wins

• Supplier invoice processing - Variable formats, exception handling, cross-system matching against POs and goods receipts. Standard OCR and ERP modules cannot handle the variability.
• Customer service triage and resolution - Unstructured customer messages, classification, routing, draft responses. Standard ticketing software handles the queue but not the work itself.
• Sales lead qualification - Reading inbound emails, evaluating intent, pulling enrichment data, drafting outreach. CRM software stores leads; an agent qualifies them.
• Quote and proposal generation - Especially in variant manufacturing where every quote is custom. Reading the customer specification, calculating from BOM history, generating the document. Standard CPQ software cannot handle the long tail.
• Compliance reporting - Pulling data across systems, formatting for the regulator, flagging anomalies. Different from accounting; the work is integration, not booking.
• Document review for procurement - Comparing supplier contracts, NDAs, framework agreements. Reading clauses, flagging deviations from standard terms.
• Supply chain exception handling - Delivery delays, stockouts, reroutes. Pulling signals across ERP and logistics systems, triaging, escalating with context.
• Cross-system reporting and reconciliation - Numbers from ERP, CRM, and BI tools that should match but do not. Investigation work that humans currently do manually.
• Knowledge base search and synthesis - SharePoint with 10 years of documents, decades-old SAP customisations, expert know-how from people who have left. Search software finds documents; an agent synthesises across them.

Where neither one is the right answer

Sometimes the honest answer is “you do not have a software problem”. Buying tools to solve a process problem is the most expensive form of avoidance.

• Broken processes - If the process itself is poorly designed, no software fixes it. Fix the process first, then automate the right one.
• Politically blocked processes - If the friction is between departments rather than in the work, software is not the lever. Tooling makes a politically broken process worse, not better.
• Compliance-only processes that nobody actually uses - Some processes exist on paper and not in reality. Automating them encodes the fiction.
• Processes used by under three people once a year - Volume too low to justify automation cost. A spreadsheet and a checklist is the right tool.

The Cost Comparison Done Honestly

Most cost comparisons between standard software and AI agents are framed wrong. They compare licence to licence and miss everything that matters. Here is a complete frame.

What standard software actually costs

• Per-seat licence - 30 to 200 euros per user per month for most B2B SaaS, depending on category. Higher for ERP, lower for productivity tools.
• Implementation - 30 to 50 percent of year-one total cost for ERP-class systems, less for lighter tools¹³. Often under-scoped during procurement.
• Integration - Custom connectors to your existing systems. Recurring cost as both sides update.
• Training and change management - User onboarding, ongoing training as features change, retraining when the vendor updates the UI.
• Internal admin - Someone owns the tool, manages users, configures workflows, troubleshoots issues.
• Annual price increases - 10 to 20 percent at renewal is now standard; 15 to 25 percent for major vendors¹².
• Switching cost when you eventually leave - Often higher than the original implementation. Data migration, integration rewrite, retraining, parallel run period.

What an AI agent actually costs

• Platform or process fee - Typically 1,500 to 8,000 euros per month per active process for production-grade agents in the Mittelstand. Varies by complexity and volume.
• LLM inference cost - A few cents per task for most workflows. Accumulates with volume but stays modest unless processing volume is in millions per day.
• Implementation - 30,000 to 80,000 euros for a focused first deployment. Mostly process mapping, integration setup, and validation.
• Integration - One-time setup against existing system APIs. Less recurring than standard software integration because agents tolerate API drift better than scraping bots.
• Monitoring and feedback loop - Someone reviews exceptions, corrects errors, feeds learning back. Typically 0.2 to 0.5 FTE per active agent.
• No per-seat scaling - The team can grow without licence cost growing.

Three-year total cost on a representative process

Take a process where 12 people across your team spend a combined 30 hours per week on supplier invoice exception handling. Standard SaaS approach: a procure-to-pay tool at 80 euros per user per month for the 12 users, plus an additional invoice OCR add-on, plus implementation. AI agent approach: one agent handling the exception layer, integrated with the existing ERP.

Where the cost comparison flips

• Pure commodity processes - When the process really is the same as everyone else and exception rate is genuinely under 5 percent, standard software wins on cost too. There is little manual work for the agent to remove.
• Very small teams - Under 5 people on the process, the per-seat licence math works in standard software’s favour. Agent platform fees do not amortise across a small team.
• Compliance-driven processes - When the standard tool encodes regulated logic that you would otherwise have to build into the agent (tax filing, GoBD bookkeeping), the agent loses on compliance burden alone.
• High-exception, multi-system processes - Agents win cleanly. The labour cost being avoided is significantly larger than the licence delta, and standard software cannot remove it.

The Hybrid Reality Most Mittelstand Companies End Up With

Almost no Mittelstand company ends up running only standard software or only AI agents. The realistic state is hybrid: standard software for commodity processes, AI agents for the work that crosses between systems and handles the variability standard tools cannot. Understanding the hybrid pattern is what makes the architecture sustainable.

The architecture that works

• Standard software stays as the system of record - SAP, DATEV, your CRM, your DMS. They store the truth. Updates and audit logs live there.
• AI agents become the work surface above - Users interact with the agent for goal-driven work. The agent reads from and writes to the systems of record.
• Workflow tools handle stable connections - Where a deterministic workflow is enough (event A always triggers action B), n8n, Make, or similar bridge systems without an LLM in the loop. Agents call these as primitives.
• Standard software handles the front of standard processes - Bookkeeping, payroll, expense submission still go through their dedicated tools. The agent layer sits on top of complex, judgement-heavy work.

The most common hybrid patterns we see

• SAP plus AI agent for procurement - SAP MM stays as the system of record for purchase orders, goods receipts, invoices. The agent handles invoice exception matching, supplier query response drafting, and contract review against framework agreements.
• CRM plus AI agent for sales operations - HubSpot or Salesforce stays as the deal record. The agent handles inbound lead triage, qualification calls preparation, follow-up email drafting, and CRM data hygiene.
• DMS plus AI agent for knowledge work - SharePoint or M-Files keeps documents. The agent answers questions across them, drafts responses to frequent queries, and synthesises content for new use.
• DATEV plus AI agent for finance back-office - DATEV books the entries. The agent handles document intake, classification, supplier matching, and exception escalation before bookings happen.
• Helpdesk software plus AI agent for service - Zendesk or Freshdesk keeps the ticket queue. The agent reads inbound messages, classifies them, drafts responses, and resolves the cases that match known patterns - while routing complex cases to humans with full context.

How the responsibilities split in a hybrid architecture

The Decision Framework

Use this framework process by process. The output is a defensible decision you can take into a procurement committee or a board meeting. Skip the framework and you end up with the SaaS sprawl most Mittelstand IT landscapes already have.

Step 1: Classify the process

First answer: is this a commodity process or a specific process?

• Commodity - Most companies in your industry do this the same way. The process does not generate competitive advantage. It is the cost of doing business.
• Specific - The way you do this process is part of how your business works. The process knowledge is yours. Doing it the same as everyone else would erode your value.

If commodity, default to standard software. If specific, evaluate further with steps 2 to 4.

Step 2: Score the variability

How variable are the inputs and outputs of this process?

• Low variability - Same data formats, same fields, same case structure every time. Exception rate under 5 percent. Standard software fits.
• Medium variability - Some structure, some unstructured. Exception rate 5 to 15 percent. Hybrid - standard software with an agent layer for the variable cases.
• High variability - Documents in different formats, free-text inputs, exception rate over 15 percent. AI agent fits; standard software will leave most of the work manual.

Step 3: Score the cross-system reach

How many systems does this process touch end-to-end?

• One system - The process lives in one tool. Standard software inside that tool, or its native AI features, are usually sufficient.
• Two systems - Integration is feasible. Standard software with a workflow tool between them often works.
• Three or more systems - The integration burden grows fast. AI agent’s native cross-system reasoning starts to win.

Step 4: Score the volume and team size

How many people work on this process and how much volume runs through it?

• Small team, low volume - Under 5 people, under a few hundred items per month. Standard software per-user economics likely wins.
• Larger team, higher volume - 10 plus people, thousands of items per month. AI agent economics start to win as headcount licence cost grows.
• Specialist team, high volume - Small team handling huge volume - exactly where agents shine. Per-process pricing decouples from headcount.

Step 5: Read the decision matrix

How Superkind Fits

Superkind builds custom AI agents for the layer above your existing standard software. We do not replace your SAP, your DATEV, your Personio, or your Salesforce. We build the agent that sits on top, removes the work standard software cannot, and makes your existing stack actually work as one system.

Core capabilities

• Process-first build approach - We map the actual process before any code is written. The agent gets built around what your team really does, including the exceptions and edge cases standard software ignores.
• Native integration with Mittelstand standard software - SAP (S/4HANA, ECC, Business One), Microsoft Dynamics, DATEV, Personio, Salesforce, HubSpot, proAlpha, ABAS, M-Files, SharePoint. Connect via APIs, not UI scraping.
• Document intelligence without templates - Reads variable supplier invoices, contracts, technical specifications, customer emails. No per-supplier or per-document configuration needed.
• Cross-system orchestration in one workflow - A single Superkind agent can read from your CRM, write to your ERP, query your DMS, and send an email - all within one task. No swivel-chair work for the user, no brittle integration pipelines.
• Exception handling with full context - When an agent encounters an unusual case, it presents the human reviewer with what it found, what it tried, and what decision is needed. Reviews are 5 to 10 times faster than standard exception queues.
• Human-in-the-loop checkpoints by design - You define which actions require approval and at what confidence threshold. Agents escalate with context rather than stalling silently. Critical for high-stakes decisions and EU AI Act alignment.
• EU deployment and DSGVO alignment - Agents can run on EU cloud or your own infrastructure. Data does not leave your defined perimeter. Audit logs track every agent action for compliance documentation.
• 8 to 12 weeks to first production deployment - From process assessment to live operation on a focused first use case. Not a multi-year transformation; a focused scope you can validate against business outcomes.

Superkind vs alternative paths

The 90-Day Plan: From Decision to First Outcome

This plan covers running the decision framework on one process, picking the right tool, and reaching first production value. Use it to structure your internal discussion or align with your delivery partner.

Weeks 1 to 3: Decision and scoping

• Pick three candidate processes - Each one with significant pain (exception handling, manual rework, swivel-chair work, missed SLAs). Document current pain in numbers, not adjectives.
• Apply the decision framework - Score each process on commodity-vs-specific, variability, cross-system reach, and team size. Output is one process recommended for standard software, one for AI agent, one as a hybrid.
• Validate the recommendation with the team that owns the process - The people doing the work know whether your scoring is realistic. If they say it is wrong, listen.
• Define success metrics before any tool selection - Hours per week saved, error rate reduced, throughput increased. Numbers you can measure in 90 days, not hypothetical ROI projections.
• Confirm system access - For an AI agent, identify which system APIs are needed. For standard software, identify which integrations are needed. Surfacing API gaps now avoids surprises later.
• Brief the Betriebsrat where applicable - If the process generates employee performance data, start consultation early. Most works councils accelerate projects when involved early.

Weeks 4 to 8: Build or configure

• For standard software - Vendor selection (3 to 5 demos), procurement, contract, configuration of fields and workflows, integration to your existing systems, user training plan.
• For an AI agent - Process mapping detailed enough to brief development, prompt and tool design, integration setup, escalation thresholds and human-in-the-loop checkpoints.
• For both - Test against real historical data including exceptions, document the audit trail, validate against the success metrics defined in week 1, train the team that will work with the new tool.
• Document the rollback plan - If something goes wrong in production, what is the backup. This is often skipped in a hurry; include it.

Weeks 9 to 12: Production and learning

• Deploy to limited scope first - 20 percent of process volume, one product line, one region. Run in parallel with the existing process for two weeks before expanding.
• Review outputs weekly - For standard software, are users actually using it correctly. For an AI agent, are exceptions being handled well and are corrections feeding back. Both need active oversight in the first month.
• Measure against success metrics - Compare to the baseline you established. If numbers are not moving as expected, diagnose before expanding rather than after.
• Expand once metrics validate the deployment - Two to three weeks of stable operation at limited scope gives you the confidence to scale to full volume.
• Document the experience for the next process - What did you learn about your decision framework. Where did the recommendation prove wrong. The second decision is faster when the first is documented.

Related Articles

• RPA vs AI Agents: What German SMEs Get Wrong About Automation
• AI Agents vs Microsoft Copilot: When Custom Is Worth the Premium for the Mittelstand
• AI Agents for the Mittelstand: How Germany’s Hidden Champions Deploy AI Without Losing What Makes Them Great
• Which LLM Should the Mittelstand Choose? GPT, Claude, Gemini and Mistral Compared

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.