AI in SMEs: The Nap is Okay

Meinung & Provokation · 21. Mai 2026 · Joseph Flesh

AI in SMEs is being slept on. Why that's rational in the short term – and what roadmap CEOs need now before it gets really expensive.

Three weeks ago, I was in Gütersloh in the shipping zone of a mechanical engineering supplier. Next to me was Thomas, CEO, 58, gray sleeveless sweater, tired eyes. It smelled of cardboard, cutting fluid, and that slightly burnt coffee that only office machines in industrial areas can produce. Thomas pointed to a stack of delivery notes and said: "Klaus, we've been talking about AI in SMEs since January. But yesterday, our ERP pulled three prices incorrectly again." I laughed. Briefly. Then I stopped.

My thesis is simple and won't please everyone: German SMEs are sleeping on AI – and in the short term, that's okay. Not elegant. Not visionary. But often more economically sound than what some consultants are currently selling as a breakthrough.

That sentence has a painful second part: This nap cannot last 24 months. Anyone who proudly declares in 2026 that they have "only released ChatGPT for texts" but has no data strategy, no clean CRM, no process ownership, and no plan for productive AI use cases, is no longer just sleeping. They are falling into a hole.

AI in SMEs: Why most are wrong

For months, I've been hearing the same accusation. At trade fairs. In advisory boards. In those IHK breakfast rounds where the butter pretzels are often better than the slides. "SMEs are too slow." True. Well, almost.

Slow isn't automatically stupid. Over the past two decades, I've visited factories of Trumpf in Ditzingen, Wittenstein in Igersheim, Festo in Esslingen, and small contract manufacturers on the Swabian Alb. The good companies there have a virtue that sounds almost old-fashioned in the AI discourse: they calculate. Not at a pitch deck level, but with machine hour rates, scrap rates, inventory, payment terms, and the question of whether the customer from Wolfsburg will swallow the price increase.

Anyone who tells these people that they immediately need to set up a large Gen-AI program has never heard the noise on a late shift. The forklift beeps in reverse, a worker taps a fixture with a rubber mallet, compressed air hisses somewhere. And in the office next door sits Andrea, Head of Sales at a hidden champion in Bielefeld, struggling with 4,800 duplicates in the CRM. And from this material, an AI assistant is supposed to build reliable sales forecasts? Please.

Most are wrong because they treat AI as a technology issue. But it is first and foremost an operational issue. Who has access to what data? Who decides when the model and the master disagree? Who is liable if an AI systematically filters out applicants in recruiting? Who pays the cloud bill if a pilot project suddenly turns into 600 users?

I'll put it bluntly: German SMEs aren't stupid – they just know that it's dangerous to scale the world's most expensive technology with data garbage and half-baked regulations.

The uncomfortable truth about AI in SMEs

The number that stuck with me comes from Passion4IT's "AI Readiness Check DACH 2026," which refers to Gartner and IDC analyses: Around 60 percent of all AI initiatives fail due to an inadequate data foundation, not the technology itself. This is not a footnote problem. This is the core.

In March 2025, I looked into a data room at an automation company near Heilbronn. No glossy demo. A real server room, too warm, fans like a hairdryer starting up, a yellowed sticker from 2016 on the door. Markus, IT manager, showed me three export files from ERP, MES, and the service system. Three item number logics. Two time formats. One customer with four spellings. "And now sales wants an AI for next-best-offer," Markus said. He said it without anger. That was worse.

McKinsey reported in the Global AI Survey 2023/24 that only about 11 to 15 percent of companies worldwide achieve significant, measurable EBIT increases with AI. BCG came to a similarly sober conclusion for Industrial Goods in 2024: Only 20 to 25 percent of AI pilot projects achieve the desired business impact. The rest? Slides, demos, frustration. Sometimes also a nice dashboard that displays numbers every morning that no one trusts.

Now let's take the typical German SME with 180 employees, 72 million euros in revenue, an EBIT margin between five and eight percent, two large customers, an SAP B1 or Proalpha system, an aging workforce, and a sales manager who still tracks his pipeline in Excel because the CRM "doesn't quite fit." For this company, a failed large-scale AI project is not an annoying innovation pain. It can eat up a year's margin.

SMEs are not sleeping on AI, but on data quality

This is the sentence I would most like to nail to the wall for CEOs. Not with anger. With a small stainless steel sign, like you find at Kärcher or Phoenix Contact in the visitor center. AI is rarely the first problem in SMEs. The first problem is item master data, quotation processes, machine history, duplicates, authorization concepts, process discipline.

Releasing ChatGPT for emails doesn't make a company AI-ready. It might make emails a bit smoother. Sometimes longer. At a company in Augsburg, Sabine, commercial director, told me in April 2025 that since the release of Gen-AI, her team produces more text but doesn't have less work. "People are now formulating more beautifully, but approval still rests with me," she said. On her desk lay three contract folders, a half-eaten apple, and a printout of the EU AI Act with yellow highlights.

That's the crux of the matter. AI without process decision is cosmetics. AI without clean data is gambling. AI without responsible parties is theater.

I'd rather have clean master data for a year than three months of AI circus. Our customers don't pay for experiments, but for functioning systems.

— Jens, COO of a special machine builder from Pforzheim

Why waiting can be rational for AI in SMEs

There's a hard truth that's unpopular in Berlin panels: Not every company has to be at the forefront of every technology wave. Family businesses, in particular, have often learned that the second buyer gets the better machine. The first pays tuition. The second buys the revised version, with stable software and fewer teething problems.

It's no different with AI. Large corporations, software companies, and startups are currently bearing some of the experimentation costs. SAP is pushing Joule into its product world, Microsoft is integrating Copilot into M365, Dynamics, and Power Platform, Intershop from Jena is talking about AI-powered agentic B2B commerce for SMEs at its 2025 general meeting. Sounds fancy. But in many companies, they're still happy if the webshop doesn't mess up customer-specific prices after an update.

In June 2025, I was at a supplier near Nuremberg. The B2B shop was cleanly integrated, after four years of project work, not four sprints. In the warehouse, a scanner beeped every few seconds, and blue markings for walkways were stuck to the concrete floor. Stefan, Head of E-Commerce, told me: "Agents? Klaus, we'd be happy if every customer could find their framework agreement number correctly." That sounds small. It isn't. It's securing revenue.

Anyone who says in such a situation that we'll wait for autonomous AI agents is not being backward. They are protecting delivery capability. And delivery capability trumps almost every buzzword in B2B.

The regulatory trap: EU AI Act, AGG, and liability

The EU AI Act was finally passed in 2024, with obligations phased in until 2026 and 2027. High-risk AI in HR, quality control, or safety-critical production processes brings documentation and audit requirements. Corporations digest such things with legal departments, compliance teams, and external law firms. SMEs often respond with a simple reflex: "We'd rather leave it for now."

I don't consider this reflex cowardly. In HR, it's even healthy. The platform anymize warns of a special problem with internal promotions: If all cohort data is available in the company, the AGG reversal of the burden of proof can become particularly well-substantiated – or particularly dangerous. Translated into factory canteen German: If your AI produces promotion scores and a pattern based on gender, age, or origin becomes visible, then shrugging your shoulders is no longer enough.

In September 2024, I sat in Stuttgart with Claudia, HR manager of an automotive supplier, at a round table made of light wood. Outside, the S-Bahn passed by, inside coffee cups clinked. "We are not productively testing AI in recruiting," she said. "Not until I know who will explain in court why candidate A was rejected." I didn't find that innovation-hostile. I found it mature.

But: The strongest counter-argument against my nap

Now comes the part where some CEOs become too comfortable for me. The strongest counter-argument is: In the short term, waiting may be rational, but economically and strategically, it can be ruinous. There's no getting around that.

The McKinsey Global Institute described additional annual productivity gains from AI of 0.2 to 3.3 percentage points in 2023 and 2024, depending on the industry. The OECD and the EU Commission have been warning for years about structural competitive gaps if countries adapt AI too slowly. Germany is already struggling with high energy costs, bureaucracy, weak demand, and an investment fatigue that you can almost smell in some factory halls. If the AI productivity lever is then left unused, caution eventually turns into self-harm.

And no, SMEs are not slow everywhere. Trumpf uses AI in laser systems, predictive maintenance, and smart factory applications. Krones works on plant optimization, quality assurance, and service planning. DMG Mori is advancing condition monitoring and digital twins. Festo has been showing for years that AI in learning systems, condition monitoring, and quality assurance doesn't have to sound like science fiction. Schaeffler uses AI for predictive maintenance, production, and logistics. These companies sometimes speak less loudly than US tech corporations, but they are working.

Anyone in mechanical engineering, plant construction, or component manufacturing who sleeps on the service business now will not only lose a little margin later. They will lose touch with the customer. VDMA, BCG, and Accenture analyses from 2023/2024 show effects in projects around service, spare parts, and remote monitoring such as five to ten percent more service revenue or 20 to 40 percent fewer unplanned downtimes. These are not LinkedIn likes. This is money.

The follower strategy has an expiry date

I only defend waiting if it is active waiting. Passive waiting is dangerous. Active waiting means: cleaning up data, nailing down processes, prioritizing use cases, training employees, observing providers, starting small productive deployments. Passive waiting means: saying "we'll look into it" and hoping the spook will pass.

The spook won't pass. Microsoft Copilot won't disappear from offices. SAP Joule won't politely ask if SMEs are psychologically ready. Intershop, Salesforce, HubSpot, ServiceNow, and numerous specialized providers are integrating AI into process software. Anyone who then has no own data rules, no roles, and no idea of the benefits will still get AI. Just as a shadow process.

What I see in practice

I see two types of SMEs. One talks little and works diligently. In Heilbronn, a sales manager showed me in May 2025 how his team, with the help of AI-powered lead scoring, had filtered out 620 target accounts from 18,000 existing and trade fair contacts. No fireworks. But after nine months, there were three times as many qualified initial appointments in the calendar, without a new sales employee. The man's name was Ralf, he came from Ludwigsburg, wore safety shoes, and had more order on his ThinkPad than some corporate strategies.

The other type of SME plays theater. There's an AI working group, a kickoff presentation with a robot image, and three people secretly using ChatGPT for quotes because the official tool is still stuck in data protection. A poster about digital transformation hangs in the hallway, below it a printer that has been reporting a paper jam for two days. Am I exaggerating? Only a little, unfortunately.

In sales, the situation is particularly absurd. Many CEOs believe that AI in sales is a matter of text generation. Wrong. The hard levers lie in Ideal Customer Profile, account prioritization, trigger recognition, quote follow-up, pricing hints, and clean handover between marketing, sales, and service. A nice cold email doesn't save a bad target customer list.

Amplifa ICP Playbook A practical playbook to clearly define target customers, sharpen data, and not unleash AI on the wrong accounts in sales.

That's why I consider the topic of ICP underestimated. Anyone who doesn't have a clear answer to which customers are truly profitable, which triggers indicate purchase readiness, and which industries only keep the field sales busy, should not apply AI to lead generation. Otherwise, they automate scatter loss. Putting the cart before the horse, just with a cloud subscription.

AI in Sales: Where SMEs cannot wait

In sales, my patience runs out faster than in production. An AI in quality control can increase scrap or give false security with bad data. An AI for account research, CRM hygiene, and quote prioritization is less risky if it is properly contained. There, you can learn without immediately endangering a production line.

A CSO from Nuremberg, let's call him Martin, told me in February 2025 after a VDMA evening in Frankfurt: "If I wait another year, I'll know my customers less well than my competitor." He didn't mean some US corporation. He meant another German supplier, 90 kilometers away, who was already evaluating visit reports, scanning spare parts histories, and deriving sales signals from service cases.

That's the point. AI doesn't have to control production immediately. It can first prevent sales from wasting time on the wrong accounts. It can make dormant quotes visible. It can recognize that a customer has ordered less for six months, even though their industry is growing. It can tell the field sales team which five visits next week are likely to generate revenue – and which will only cost coffee.

Amplifa for Lead Generation For B2B teams that want to build a reliable pipeline from data, target customer profiles, and triggers – without AI folklore.

Why a pure inbound strategy is too thin in 2026

Anyone who still relies on a pure inbound strategy in 2026 will have no pipeline in five years. I'm pretty sure of that. The purchasing departments of OEMs won't leisurely download a whitepaper just because an SME puts "Innovation" in the headline. They compare delivery capability, price, certificates, data integration, and response time. And they are doing so increasingly digitally.

Outbound isn't dead. Bad outbound is dead. Out of 1,200 emails, a mechanical engineer from East Westphalia received four replies in autumn 2024. One of them was a complaint. After an ICP refinement, industry clusters, and triggers from investment announcements, job advertisements, and service events, there were 31 real conversations from 480 targeted contacts. No miracle. Just craftsmanship.

What needs to happen now

I wouldn't advise any CEO to set up an eight-figure AI program tomorrow just because a competitor posts a photo of their innovation lab on LinkedIn. But I would advise them to take a brutal inventory this week. Not in a strategy hotel. In the company. With sales, IT, production, controlling, and a person who really knows the data – often not the department head, but the woman who has been pulling exports for 14 years.

1. Check data situation – make item master data, customer master data, service history, offer data, and CRM duplicates visible, don't sugarcoat them.
2. Choose two low-risk use cases – such as account prioritization in sales or internal knowledge search for service technicians.
3. Define responsibilities – a process owner, an IT manager, a specialist department, a clear budget.
4. Involve compliance early – especially in HR, quality control, and safety-relevant processes due to the EU AI Act and AGG.
5. Measure ROI in small increments – not with visions, but with appointments, lead time, scrap, service revenue, or saved hours.
6. Stop or scale after 90 days – no eternal pilot project with monthly steering committee and no results.

The most important step is the first, and it sounds unsexy: check the data situation. I know that sells worse than an AI agent with a demo video. But if customer names, machine serial numbers, and offer statuses are incorrect, no matter how expensive the platform, it cannot produce reliable decisions. It will only make more expensive guesses.

Amplifa Pipeline Check An approach for CEOs and sales managers who want to know if their pipeline's data foundation, target customer logic, and AI deployment are even viable.

FAQ: Is AI in SMEs really too expensive?

Yes, if implemented incorrectly. No, if you start with clear use cases. A Gen-AI system that is frequently used in quote calculation, technical documentation, or customer communication can quickly become expensive due to cloud subscriptions, consultants, internal project time, and rework. With EBIT margins of five to ten percent, this is not pocket money. A limited use case in sales or service, cleanly measured, is a different league.

FAQ: Which AI use cases fit first?

In many SMEs, I would not start with autonomous production control. Too delicate. Too many interfaces. Too high liability. Better are use cases that condense knowledge and prepare decisions: summarizing service tickets, identifying spare parts opportunities, clustering existing customers by potential, making technical documentation discoverable, prioritizing sales visits. Humans decide, AI pre-sorts.

FAQ: What about the lighthouses like Trumpf and Festo?

Trumpf, Festo, Schaeffler, Krones, or DMG Mori prove that AI works in industry when data foundation, engineering culture, and process discipline come together. However, they do not refute that many smaller companies are not yet ready. They rather show the gap. And precisely this gap is dangerous if it becomes an excuse.

The price of the nap

The minimum wage is set to rise from 12.82 euros to 13.90 euros in 2026 and 14.60 euros in 2027, according to the Minimum Wage Commission's decision. In industry, relevant wage costs are already far above this. Skilled workers are becoming scarcer, administrative work is not getting cheaper, customers expect faster responses, and banks are looking more closely at digital maturity for financing and succession. Anyone who believes they can continue to operate office work, sales preparation, and service coordination as they did in 2018 will not be overrun by AI. They will be crushed by costs.

Private equity firms have long been asking about automation potential. Family offices too. In January 2025, I sat in Munich with Nora, an investment manager at a family office, in a meeting room overlooking wet asphalt and Leopoldstraße. "If a target company doesn't have a data and AI roadmap, we discount that," she said. Not dramatically. Not with a Silicon Valley grin. More like someone evaluating a machine that's overdue for maintenance.

Major customers are also becoming more impatient. Automotive, aerospace, medtech, mechanical engineering – everywhere, demands for digital supply chains, quality data, traceability, and response times are growing. Today, the OEM politely asks about interfaces. Tomorrow, the supplier will be excluded from the shortlist because a competitor provides better data. This will not be recorded as an "AI problem" in the minutes. It will be recorded as "lack of process capability." Sounds harmless. Is more deadly.

My actual provocation

German SMEs are sleeping on AI – and that's okay if they're currently changing their mattress. Meaning: if they're cleaning up data, clarifying processes, understanding compliance, and preparing initial use cases with real ROI. It's not okay if they pull the covers over their heads and hope the next technology wave passes them by again.

I understand the fatigue. ERP implementations, MES projects, CRM rollouts, cloud migrations, energy crisis, supply chain stress, interest rate turnaround, skilled labor shortage – many organizations are not innovation-averse, they are exhausted. At a company near Ulm, Peter, production manager, told me in October 2024: "We're not afraid of AI. We're afraid of the next project that remains half-finished." That sentence is more honest than 80 percent of the keynotes I heard in 2025.

Nevertheless: exhaustion is not a strategy. SMEs don't have to chase every hype, but they need to know where they want to stand. In 90 days. In twelve months. Especially in two years, when AI functions are no longer sold as an extra, but are embedded in every software running in the office and on the factory floor.

When I asked Thomas in Gütersloh what he was doing with AI now, he zipped up his gray fleece jacket and pointed to the shipping documents. "First, we're cleaning up this mess," he said. Then he grinned. "But this time, so a machine can read it later." Outside, a truck started, diesel fumes in the rain. Perhaps AI in SMEs begins exactly like that. Not with a robot on stage, but with a CEO who finally takes his data seriously.