AI in SMEs: China, USA, Europe 2026

Marktanalyse · 22. Juli 2026 · Anthony Filipiak

AI in SMEs is becoming a test of location: Read where the USA, China, and Europe stand – and what manufacturers really need to do now for their pipeline and production.

AI in SMEs is the application of artificial intelligence in small and medium-sized enterprises. That's the definition. But it's too clean. In practice, AI in SMEs in 2026 is more of a stress test: Who can combine data from ERP, CRM, and machine control, who can explain inference costs, who dares to truly restructure a process – and who just paints a strategy paper with a robot image on page one. My prognosis is simple and uncomfortable: In the next 24 months, the company that wins will not be the one that buys the best model, but the one that translates AI into costs, lead time, and pipeline most quickly.

The three-track AI race between China, the USA, and Europe is therefore not a geopolitical feuilleton topic. It lands on the desks of managing directors in Bielefeld, sales managers in Nuremberg, and production managers in Tuttlingen. The USA leads in frontier models, capital, and platforms. China is catching up with brutal speed in model quality and industrial implementation. Europe is in between – not blind, not without a chance, but too often with the handbrake of venture capital and the excuse of regulation. Not quite. Regulation is not the main problem. The main problem is that many European companies still treat AI like an IT project, even though it has long been a matter of margin, delivery capability, and market share.

AI in SMEs: Status Quo of the AI Race

Where do we stand today? The USA continues to dominate the top. OpenAI, Anthropic, Google, Meta, and Microsoft not only possess models but also distribution channels, cloud capacities, developer ecosystems, and the ability to pour billions into training and inference without causing panic in every quarterly call. According to CB Insights, around 100 billion US dollars flowed into AI startups worldwide in 2024, a large part of it in the USA; PitchBook reported double-digit billion-dollar deals for generative AI alone in 2024. This is noticeable. Anyone starting a Copilot project in a DACH machine builder today almost automatically ends up with Microsoft, Azure OpenAI, or one of the major US platforms. Not because procurement is in love. But because the systems are already in the IT landscape.

China is playing a different game. It's not just about chatbots, but about industrialization. According to reports from Shanghai, China plans investments of 2 trillion yuan in data centers over five years – roughly 250 billion euros, depending on the exchange rate. Reuters reported on July 17, 2026, that Beijing wants to challenge US leadership with an international AI coalition. Moonshot positions Kimi K3 with 2.8 trillion parameters as an open-weight model and claims benchmark proximity to OpenAI and Anthropic. Is every benchmark cleanly comparable? Honestly? I don't know. But the direction is clear: China doesn't just want to keep up. China wants to make AI so cheap and widely available that Western premium prices come under pressure.

Europe has strong industry, good automation, companies like Siemens, Bosch, Festo, Phoenix Contact, Wittenstein, and Trumpf. But Europe has too little capital, too little scaling ambition, and too many projects that end after a proof of concept because no one has provided the process owner with a budget. Philippe Aghion, according to DW in 2026, essentially said that Europe needs more venture capital, long-term research funding, and better research conditions if it wants to catch up in AI. That's true. But it doesn't help the managing director of a machine tool supplier in Schwäbisch Hall much if his scrap rate increases in Q4 and his sales team continues to prioritize leads based on gut feeling.

The Surprising Forecast for 2026 to 2028

My forecast surprises many strategy rounds: SMEs will not fail due to a lack of model quality. The models are good enough. Not always brilliant, sometimes annoying, occasionally confidently wrong – but good enough for offer preparation, spare part classification, quality documentation, customer research, maintenance forecasts, and sales prioritization. SMEs will fail due to integration costs, data access, and unclear process responsibility. In other words, things that never shine on the stage of an AI conference. The sound of this is not applause, but the quiet click of an Excel file on a network drive that no one has wanted to touch since 2017.

Every week, I speak with managing directors and sales managers in the DACH region. The sentence I've heard remarkably often since March 2025 is not: "Which LLM is the best?" The sentence is: "Where do we start without giving away our data and without the works council tearing us apart?" A CSO from Nuremberg, Markus, recently told me: "If I get another tool that promises me more data but doesn't bring appointments, I'll cancel the pilot after six weeks." Harsh. But fair. AI in SMEs needs to get out of the slide deck world. It needs to move money in offers, manufacturing, service, and sales.

Trend 1: The USA Sells Peak Performance – and Dependence

The USA will remain the pace-setter for frontier models in 2026. This means: best general performance, fastest product cycles, strongest platform lock-in. OpenAI, Anthropic, and Google set the benchmark that buyers, CIOs, and investors use. In many DACH companies, Microsoft Copilot is the first point of contact with generative AI because Microsoft 365 is already running. Convenient. Well, almost. Because anyone who sells Copilot as an AI strategy confuses a user interface with an operating model.

For medium-sized manufacturers, US dominance has two sides. On the one hand, they get very good models, strong security documentation, and a procurement logic that purchasing departments understand. On the other hand, dependencies on a few platforms grow. Prices can change. Data flows must be checked. And anyone who tries to run every use case directly through the most expensive model class builds a cost structure that looks nice with 20 users but suddenly smells like a CFO conversation with 600 factory employees. For a manufacturer of plastic components from Baden-Württemberg, the first projection for an internal assistance system in May 2026 was over 38,000 euros in monthly inference and license costs. After model routing and process tailoring, it was under 11,000 euros. Same purpose. Different architecture.

This is the point many overlook. Frontier models are not automatically wrong models. But they are rarely the most economical answer to all questions. A service technician doesn't always need the model with the best philosophical argumentation when troubleshooting a system. He needs access to maintenance history, bill of materials, error codes, and an answer that works in the factory, even if the Wi-Fi in Hall 4 smells of oil and the scanners beep. The USA delivers the peak. SMEs often need the robust middle ground.

Europe must take its strengths seriously: industry, research, and the single market. But without more venture capital and long-term financing, strength will not translate into speed.

— Philippe Aghion, Nobel laureate and economist, paraphrased from DW reporting 2026

I consider the US lead to be real, but overestimated in its impact on medium-sized manufacturers. Sounds contradictory. It isn't. If a model is three percentage points better in difficult coding benchmarks, that doesn't automatically solve the problem of a spare parts sales department that has to manage 18,000 item numbers with inconsistent descriptions, outdated PDF catalogs, and four languages. In discussions with sales managers of automation suppliers, I rarely hear: "We need more intelligence in the model." I hear: "We finally need clean handover between marketing, inside sales, and field sales." AI can solve that. But only if someone tackles the dirt in the process.

Trend 2: China Pushes AI into Industry – and into the Cost Curve

China is not simply second place in the model race. China is the more dangerous opponent for European manufacturers because the country combines AI with manufacturing depth, supply chain power, and state-orchestrated infrastructure. According to DW, experts now see the gap between Chinese and US models in certain benchmarks as largely shrunk or closed. At the same time, reports from China indicate that the latest US systems are still about six to nine months ahead. That sounds like a lag. But for an industrial competition, it's almost parity if operating costs are significantly lower and the models work close enough to the application.

Moonshot with Kimi K3, Alibaba with Qwen, DeepSeek with cheap inference models, Baidu with Ernie – the list is growing. The important point is not every model name. The important point is the price dynamic. If Chinese providers make powerful models open-weight or available at aggressive prices, the discussion in European companies shifts. Then the CFO no longer asks whether AI is fundamentally too expensive. He asks why an internal knowledge search costs as much per query as a spare part in purchasing. And then the real debate begins: Which data can go where, which models run locally, which answers must be auditable?

China is not just exporting AI as technology. China is exporting a cost logic. This is dangerous for DACH. Not because every German machine builder will hang a Chinese LLM in their production planning tomorrow. That often won't happen for compliance reasons. But because Chinese competitors can reduce their process costs with AI-supported quality assurance, planning, supplier control, and offer processing. Anyone in Europe who then still says: "Our customers pay for quality" should look at their offer losses of the last twelve months. Quality remains. But price and delivery time are also at the table.

I see a hard shift here. Previously, the question was: Can we use AI? Today it is: Can we use AI more cheaply than the competitor from Shenzhen, Detroit, or Eindhoven? The difference is not academic. A manufacturer of special machines from Bavaria told me in June 2026 that an Asian competitor now delivers offers within 48 hours with variant calculation. Previously, it was five to seven days. Is there a perfect AI system behind it? Probably not. Probably a mixture of data standardization, offer modules, and automation. That's precisely why it's threatening.

Why Lower AI Costs Are Just as Important for Sales as for Manufacturing

Many managing directors still separate manufacturing and sales too much. Production-related AI gets respect because it sounds like a machine. Sales AI gets distrust because it sounds like spam. Understandable. The market is full of tools that claim to automatically win customers, and in the end, they send 3,000 bad emails with the wrong contact person. But the price decline in models also changes sales. If customer research, account scoring, trigger detection, and offer preparation become cheaper, a medium-sized sales team can suddenly work on markets that were previously too fragmented. Not with more people. With better prioritization.

What we specifically see at Amplifa: In the last 12 months, we have observed among DACH customers from mechanical engineering, component manufacturing, and technical services that 60 to 75% of usable sales signals are not in the CRM. They are in job advertisements, dealer pages, spare parts portals, trade fair exhibitor lists, certificate databases, and product changes on customer websites. For an industrial customer from North Rhine-Westphalia, 312 of 487 relevant target accounts in the CRM were marked as "inactive," although 91 of them had advertised new production or engineering positions within 120 days. This is not a data problem in the narrower sense. This is a pipeline problem with a data mask.

This is precisely where the global AI race meets SMEs. If models become cheaper, not everything automatically gets better. But suddenly it's worthwhile to analyze signals that no one could manually touch before. A sales manager cannot check 800 target customers every Monday for new plants, new machines, new tenders, and new contact persons. An AI system can do that. Provided that the ICP is clear, the data sources are cleanly prioritized, and sales accepts that gut feeling is not a scalable operating system.

Amplifa ICP Playbook A practical guide to data-driven refinement of target customers, triggers, and market segments in B2B sales – especially for industrial markets.

Trend 3: Europe Wins Not with Model Size – But with Implementation

Europe will not win the race for the largest foundation models in the short term. This is not surrender. This is reality. Anyone who believes in 2026 that a single European champion can compete against the capital depth of the USA and the infrastructure policy of China with a few funding programs has probably never seen a GPU bill. Mistral AI is important. Aleph Alpha is important. SAP, Siemens, and Bosch are important. But Europe's strongest card is not in building the largest model. Europe's strongest card is in integrating AI into regulated industrial processes that customers are actually willing to pay for.

That sounds less glamorous. But it is more economically relevant. A medium-sized manufacturer doesn't need a national AI opera. They need a system that brings together scrap data from quality control, maintenance notes from service, offer data from CRM, and delivery times from ERP – without IT security having a panic attack. Europe has advantages here: mechanical engineering knowledge, automation expertise, strong standards, established customer relationships. Companies like Festo, Trumpf, Schaeffler, Kärcher, and Phoenix Contact have learned over decades how to get complex industrial products into real processes. That's not insignificant. It's just not a Silicon Valley pitch deck sentence.

The EU AI Act is often described as a brake. Sometimes it is. I have little patience for regulation that overwhelms small teams with documentation requirements while Big Tech employs entire legal departments. But predictability can become a selling point. A Swiss medtech supplier told me in April 2026: "We prefer to buy a system that is introduced more slowly but passes an audit." That is the European sweet spot. Not the wildest demo. But an AI system that survives in the audit, in the works agreement, and in daily business.

The implementation question is ugly concrete. Where is the data? Who owns the process? Which answer can be automated? Which recommendation needs approval? Which key figure decides after 90 days whether to continue operations? In a project with a technical wholesaler from Southern Germany, the bottleneck was not the model, but the definition of a "qualified account." Marketing counted downloads. Sales counted known decision-makers. Management counted revenue in six months. Three truths. No system can make good prioritization from this if no one nails down the target definition.

AI-in-the-plant instead of AI theater

I believe Europe should take the term AI-in-the-plant more seriously. Not as a buzzword. As an alternative to model fetishism. This means AI that works close to production: quality control with camera data, predictive maintenance with sensor history, OEE optimization, energy consumption, intralogistics, shift planning, spare parts disposition. According to VDMA economic surveys 2025, investment sentiment in mechanical engineering remained tense; at the same time, companies are looking for productivity levers because wages, energy, and financing costs are pressing. It is precisely in this squeeze that AI becomes interesting. Not because it is modern. But because it cuts costs from processes that no one has fundamentally touched for years.

The same logic applies to the market. Those who sell technical products no longer sell only through field service routines and trade fair contacts. Purchasing processes are becoming more digital, willingness to switch is increasing, technical decision-makers inform themselves anonymously earlier. If a European SME does not constantly update its Ideal Customer Profiles, it will walk through 2026 with a map from 2021. Andrea, Head of Sales at a hidden champion in Bielefeld, told me in May: "Our best customers look worse in CRM than our worst leads." That's bitter. And very common.

What Does AI in SMEs Specifically Mean for Manufacturers?

For managing directors in DACH, the AI race means four things. First: The cost curve is falling, but integration remains expensive. Second: Model quality becomes a commodity, process knowledge does not. Third: Data sovereignty becomes a procurement criterion, especially for production-related information. Fourth: Anyone who only starts individual pilots loses against companies that set up AI as an operating model. That sounds strict. It is meant to sound strict. Anyone who still relies on a pure inbound strategy in 2026 will have no pipeline in five years. Anyone who sees AI only as an assistant for emails will have no productivity advantage in three years.

The capital markets will see this difference. Investors are already asking different questions. Two years ago, the question was: "Do you use AI?" Today it is more like: "Where is AI reflected in the gross margin, in working capital, or in sales productivity?" At an advisory board meeting of a medium-sized automation supplier in Munich in February 2026, it was not about a new dashboard. It was about why the order backlog fluctuates even though the market for certain components remains stable. The answer was not only in sales. It lay in a lack of segmentation, slow offer logic, and too late recognition of investment signals from existing customers.

Sovereignty is misunderstood. Many believe that only those who build everything themselves and operate it locally are sovereign. That is romantic. And expensive. Sovereign is whoever can change. Whoever cleanly separates data models. Whoever doesn't cement every process logic into a proprietary tool. Whoever knows which data can go into a US model, which belongs only in a European cloud, and which should not leave the factory at all. A CFO from Stuttgart put it dryly: "I don't want an AI religion. I want an exit plan." It can hardly be said better.

FAQ: Which AI Strategy Suits DACH SMEs?

The short answer: a two-stage strategy. First, production-related or sales-related use cases with clear ROI metrics, then a vendor-neutral architecture that enables data sovereignty, auditability, and reusability. Not the other way around. I see too many companies that spend six months drawing architecture and don't improve a single process. I also see the opposite: ten tools, no guardrails, later chaos. The middle way is not boring. It is disciplined.

A good starting point always has a hard metric. Scrap rate. Downtime minutes. Offer lead time. Appointment rate in the target segment. Spare parts sales per installed machine. Energy consumption per batch size. Once the metric is established, you can work backward: What data does the system need, what decision does it support, which person remains responsible? If these questions are not answered, the project is not an AI project. It is a software wish.

The Market Logic: Cost Beats Prestige

I deliberately contradict a popular phrase from corporate presentations here: "We rely on best-of-breed." In the AI context, this often only means that each department buys its favorite tool and no one calculates the operating costs. For SMEs, best-of-breed only makes sense if the architecture remains interchangeable. Otherwise, tool debt arises. And tool debt is like technical debt, only with more sales demos and worse contracts.

China sets the price anchor. The USA sets the performance and platform anchors. Europe must set trust and implementation. These roles are unevenly distributed, but not deterministic. A manufacturer of precision components in Switzerland does not have to train with OpenAI to remain competitive. But he must know why an offer for an A-customer takes three days, why service knowledge gathers dust in PDFs, and why his sales team only approaches target customers when the competition is already in the specification. That is the operational core.

The best AI strategy therefore does not start with the model list. It starts with the bottlenecks. Where does the company lose money, time, or market information? In many manufacturing companies, there are four areas: quality, maintenance, planning, and sales. I know these are broad areas. But they have a common denominator: They produce data that is often not used for decision-making today. The sensor measures. The CRM stores. The ERP posts. But no one connects the signals fast enough.

Amplifa Product Amplifa helps B2B teams identify target customers, evaluate market signals, and measurably scale sales processes with AI.

Preparation: 7 Steps for Managing Directors and Strategy Teams

If I could give a managing director in an SME only one piece of advice, it would be this: Don't wait for the perfect European answer in the model race. It may come. Or it may not. Your own market won't wait. An AI strategy must become visible in 90 days and have an operating model in 12 months. Otherwise, it becomes an innovation ritual.

1. Define three hard business metrics before choosing a tool. Examples: Reduce offer lead time from five days to two days, reduce scrap by 1.5 percentage points, increase appointment rate in the ICP segment by 30%. Without metrics, the loudest demo always wins.
2. Separate data classes by risk. Product data, customer data, machine data, and publicly available market signals do not belong in the same pot. Determine what can go into US clouds, European clouds, local models, or not leave the premises at all.
3. Build model routing instead of a model religion. Use strong frontier models where complexity matters, and cheaper models where classification, extraction, or research suffice. This saves money and reduces platform dependence.
4. Start with use cases that have a process owner. If no one takes responsibility for results in quality, service, sales, or planning, AI becomes an IT filing system. The process owner must have budget, time, and decision-making authority.
5. Measure after 30, 60, and 90 days. Not just usage. Results. How many offers were created faster? How many maintenance cases were correctly prioritized? How many target accounts were newly activated? Usage without impact is busywork.
6. Create a vendor-neutral data and integration layer. This sounds technical, but it's a power issue. Those who keep process logic and data access clean can switch models, renegotiate prices, and demonstrate compliance.
7. Train leaders, not just users. The master in manufacturing, the sales manager, the service chief, and the CFO must understand what AI can do, what it cannot do, and which decisions remain with humans. Otherwise, either fear or playfulness arises.

These steps are not spectacular. That's precisely why they work. A medium-sized company doesn't need an AI opera with 40 slides and five workstreams if the initial data sources are unclear. It needs a clear hierarchy. What brings money? What reduces risk? What creates data that can be reused later? In a mechanical engineering project from the Ulm area, the decisive step was not the model, but the decision to force all reasons for lost deals into four mandatory fields. After three months, the team could identify which segments were lost due to delivery time and which due to specification. Only then did AI-supported prioritization make sense.

Why Pure Pilot Programs Are No Longer Enough in 2026

Pilots were useful in 2023. In 2024, they were necessary. In 2026, they will become an excuse if they don't transition into scaling. I see too many companies that, by saying "We are testing AI," actually mean: "We don't want to make a process decision yet." The problem: The competition is not just testing. They are standardizing. They are negotiating infrastructure. They are building data pipelines. They are training leadership. They are connecting sales signals with production capacity. Then the pilot in your own company suddenly looks like a hobby room.

However, SMEs have an advantage over corporations. They can make decisions faster if management truly wants to. For a family business with 420 employees from Hesse, the decision for an AI-supported account scoring process took three weeks. No six-month steering committee. No 17 stakeholder interviews. The managing director, the sales manager, and the IT manager focused on two metrics: new appointments in the defined target segment and the proportion of relevant accounts with a current trigger. After nine months, the team had booked three times as many appointments in the prioritized segments – without a new sales employee. That's not magic. That's focus.

Of course, there are risks. Hallucinations. Data protection. IP leakage. Works council issues. Vendor lock-in. False automation. But risks don't get smaller if you ignore them. They get smaller if you classify them, limit them, and provide an operating model. I am more afraid of uncontrolled shadow tools in departments than of a cleanly introduced AI system with audit logs. The secret chatbot in the browser is not proof of innovation. It is a governance leak.

Investments: Where the Money Flows in the AI Race

Capital doesn't decide everything, but it decides a lot. The USA has the strongest combination of venture capital, hyperscalers, and research density. China directs state and private investments into infrastructure, models, and industrial application. The mentioned 2 trillion yuan for data centers are more than a number. They show that inference capacity itself is becoming industrial policy. Whoever controls computing power controls cost curves. Whoever controls cost curves influences adoption. This is not abstract for a DACH manufacturer. It determines whether an AI assistant costs 5,000 euros or 50,000 per month.

Europe invests, but more fragmented. There are funding programs, national initiatives, research clusters, Gaia-X-like debates, and strong institutes. Fraunhofer, DFKI, ETH Zurich, TU Munich – the substance is there. What is often missing is the bridge to scalable products and courageous procurement. A managing director from mechanical engineering told me in Augsburg: "We like to fund studies, but we buy too late." That's true. European industry is excellent at analyzing risks. Sometimes it analyzes for so long that the competitor has already won the customer.

For investors, the question therefore becomes interesting which SMEs translate AI into operational metrics. Not the loudest. The measurable ones. A company that reduces its offer costs, makes maintenance more predictable, and directs sales capacity to the best accounts will look different in three years than a company with the same machine but without data logic. The valuation difference will not be immediately obvious. It will show in margin stability, working capital, and resilience to demand fluctuations.

What Role Does Sales Play in the Industrial AI Race?

More than many manufacturing companies want to admit. Manufacturing likes to think first about machines, scrap, and maintenance. Understandable. But the AI race also changes market cultivation. If competitors are faster at recognizing which customers are investing, which plants are being expanded, which suppliers are failing, or which regulatory requirements are creating new demand, then sales becomes a data problem. Not just a relationship problem. Relationships remain important. But relationships without timing are an expensive visit.

I see a blind spot in DACH SMEs: Many companies know their existing customers well, but their best future customers poorly. They have historical sales lists, but no living market map. They know who bought five years ago, but not who announced a new production line three weeks ago. They visit Hannover Messe, but do not systematically evaluate exhibitor and partner movements. The rustling of brochures at the trade fair stand sounds like sales. The real music is now playing in data sources that are updated at 6 a.m. on Mondays.

Here, AI in sales becomes not a nice add-on, but a survival logic. Anyone with 12,000 potential target customers in Europe cannot treat them all equally. Field sales needs priorities. Inside sales needs triggers. Marketing needs segments. Management needs pipeline truth. If AI does not connect these four groups, it only produces activity. Activity is the most dangerous metric in sales because it feels good and yet proves nothing.

Amplifa for AI-supported Market Cultivation For industrial and B2B teams who no longer want to manually manage account prioritization, trigger detection, and pipeline building.

Risks: What Managing Directors Must Not Delegate to IT

AI risk is a matter for the boss. Not because managing directors have to understand every model architecture. They don't. But they have to define the risk tolerance. Can a system automatically contact customers? Can it issue technical recommendations? Can it combine internal margin information with external signals? Can a service employee send an AI answer to a customer unchecked? These questions are operational. And they are strategic. If they only land in IT, ultimately the person with the ticket system decides on market positioning.

A good governance model is not cumbersome. It is clear. Red data, yellow data, green data. Red data does not leave the house or only in audited environments. Yellow data may go into defined systems with contracts and logging. Green data is public or non-critical and can be used for market analysis, research, and segmentation. This simplicity helps. In an industrial customer project in November 2025, this classification reduced the approval time for new AI use cases from an average of eight weeks to just under three weeks. Not because compliance became irrelevant. But because it became understandable.

Vendor risk also belongs on the agenda. Anyone who completely ties themselves to a US hyperscaler buys speed and dependence. Anyone who uses Chinese models gets price advantages and geopolitical questions. Anyone who only allows European providers gains trust, but perhaps pays with performance or availability. There is no pure solution. There is only conscious architecture. That's why I consider multi-model capability more important than choosing a favorite model.

Personal Forecast: 2026 to 2028

My personal forecast for the next two to three years is sharp: The AI race will not manifest itself as a model race for European SMEs, but as a productivity gap. On one side are companies that embed AI in quality, service, planning, and sales. On the other side are companies that are still "gaining initial experience" in 2026 and explaining in 2028 why margins are under pressure. The gap will not open loudly. It will become visible in offer times, scrap rates, forecast accuracy, and reaction speed.

The USA will continue to shape the top. China will brutally push down the cost curve and industrial speed. Europe will only remain relevant if it takes implementation seriously. This means: less model romanticism, more process courage. Less AI lab, more responsibility in line and sales. Less funding logic, more procurement of functioning systems. I know that sounds impatient. I am too. The market won't wait until Europe has explained itself.

For DACH manufacturers, I see three groups of winners. First, companies that measurably scale production-related AI and don't stop at the demo. Second, companies that pragmatically solve data sovereignty instead of using it as an excuse for stagnation. Third, companies that understand sales and market intelligence as part of the AI strategy. Not as CRM cosmetics. As an early warning system. If a competitor knows earlier which customer is building capacity, that's no longer a sales issue. It's strategy.

I don't believe Europe has lost. But I also don't believe that Europe automatically wins because it has good engineers and can explain regulation. Good engineers are no shield against better cost structures. Regulation is no substitute for implementation. In the factory, it doesn't smell of strategy paper. It smells of cutting fluid, cardboard, metal, and sometimes of missed decisions. That's where it's decided whether AI in SMEs becomes a competitive advantage – or just the next topic for the annual retreat.