AI Costs: What SMEs Really Pay

KI & Automatisierung · 24. Juni 2026 · Anthony Filipiak

AI costs often explode after the pilot. Check cloud, data, integration, and governance before your ROI window closes.

AI costs are the expenses for software, implementation, and operation of AI systems. That's roughly what every budget deck I see from managing directors in mechanical engineering, automotive supply, and electrical engineering states. Not quite true. In practice, AI costs are primarily the costs for everything that was already broken, distributed, unclear, or politically inconvenient – data quality, interfaces, responsibilities, cloud contracts, works councils, security, change. My prediction for 2026 surprises many CFOs: It won't be the AI pilot that busts the budget, but the attempt to integrate it into SAP, MES, PLM, CRM, and real sales processes. Those who only compare license prices are calculating a construction site based on the price of a doorknob.

I'm not writing this as an analyst with a clean chart from London. I'm writing it as Anthony Filipiak, CEO & Co-Founder of Amplifa, based on conversations with managing directors, CSOs, investors, and sales managers in DACH SMEs. In March 2025, Andrea, Head of Sales at a hidden champion in Bielefeld, told me: “We won the AI case before we even started – on PowerPoint. We then lost it in the database.” I know that sound too: not a rocket launch, but the tired clicking through 17 Excel exports from three factories.

AI Costs in the Status Quo: The Headline is Too Cheap

The big market story is quickly told. McKinsey, BCG, PwC, and other firms see the global AI market reaching between 600 and 900 billion US dollars annually by 2030, depending on whether you count only software and services or include infrastructure, cloud, and GenAI platforms more broadly. In 2023, McKinsey alone described an additional economic potential of 2.6 to 4.4 trillion US dollars per year for generative AI. IDC predicted worldwide spending on AI-related systems of more than 630 billion US dollars by 2028. Sounds big. Is big. But that number means little to a managing director in Paderborn, Villingen-Schwenningen, or Linz when his cloud bill jumps from 18,000 to 52,000 euros per month after a Copilot rollout.

Headlines talk about productivity. Purchasing sees subscriptions. IT sees tokens, storage, network, logging, identity, backup, and new admin roles. Sales sees a tool that supposedly pre-qualifies offers, but without a clean ICP, it initially sorts the same wrong accounts faster. Well, almost. Sometimes it even sorts them worse, just more convincingly phrased. According to Eurostat, about 13.5 percent of companies in the EU used at least one AI technology in 2024; in Germany, the rate was significantly higher for larger companies but visibly lower for small and medium-sized enterprises. The problem is not a lack of curiosity. The problem is that many SMEs buy AI as an add-on and then realize they've bought an operating architecture.

Hyperscalers are investing billions, and this is not a charity program. Microsoft, Google, AWS, Meta, and Oracle are building data centers, buying NVIDIA GPUs, securing energy contracts, and bundling AI functions into existing platforms. Microsoft has publicly signaled investments in the double-digit billions around OpenAI and its own cloud capacities; Meta announced Capex plans of over 35 billion US dollars for 2024, strongly driven by AI infrastructure. These sums don't just disappear as a footnote. They translate into prices, packages, consumption models, minimum purchases, and lock-in. SMEs don't just pay for artificial intelligence. They also pay the electricity bill for the new platform economy.

What's Missing in Many Business Cases Today

In the business cases I see, there are almost always three clean lines: license, implementation, expected savings. Then an upward arrow. What's missing? Data cleansing, testing effort, role model, operating model, security review, legal review, training, error costs in the first few months, coordination with the works council, audit trails, model monitoring, vendor change risk. For a manufacturer from Baden-Württemberg with around 1,200 employees, whose name I cannot disclose, the approved AI pilot in sales was 180,000 euros. The actual effort until productive use across nine countries was almost 690,000 euros after 14 months, if internal FTEs were calculated at 95,000 euros full cost. This is not an outlier. This is the norm with honest accounting.

Trend 1: AI Costs Shift from Software to Infrastructure

The first trend is brutally simple: the visible license becomes smaller than the invisible operation. Large language models don't cost money because a chat window looks pretty. They cost money because every prompt consumes computing power, because every API call goes through identity, logging, and compliance, because data not only has to be read but also versioned, checked, stored, and, if necessary, deleted again. Inference costs of 0.002 to 0.03 US dollars per 1,000 tokens seem ridiculous as long as five people are testing. With 800 users, 20 workflows, long technical documents, offer drafts, sales research, and automatic summaries, this becomes a cost block. Quietly. Monthly. Opex.

For manufacturing use cases, the effect is even harsher. Computer vision in quality inspection devours data volume. Predictive maintenance requires historical sensor data, new data pipelines, and often edge hardware in factories where the control cabinet is older than the youngest data engineer. A medium-sized automotive supplier near Stuttgart calculated a vision pilot at 240,000 euros in 2024. The camera and model component was not the problem. Lighting conditions, data labeling, scrap classification, network segmentation, and the question of who is liable at 02:17 AM if the system incorrectly releases a batch became expensive. Brose, Schaeffler, Bosch, and Continental can throw entire teams at such questions. A supplier with 350 employees cannot do that on the side.

Therefore, I don't think much of AI roadmaps that start with “collecting use cases” and only then consider infrastructure. That's like planning a new production line and looking at the factory floor later. Yes, you need concrete use cases. But anyone who wants to scale AI in SMEs in 2026 must think about cloud, data architecture, rights, monitoring, and cost control together early on. Otherwise, the pilot wins the demo – and loses in operation.

“That doesn't work for us if I only find out after the pilot that we need three interfaces, two new roles, and a cloud cost model that no one in controlling understands.”

— Markus, CSO of a mechanical engineering company in Nuremberg, conversation in April 2025

Markus is right. And he describes the blind spot of many providers. In the demos, you see the result: a qualified account, a summarized requirements specification, a prioritized opportunity. What you don't see: authorization concepts, duplicates, old field logics in CRM, product names that are spelled differently in three countries, and sales regions that have grown historically like cables behind a control cabinet. To understand AI costs, you have to go beneath the surface. Not into the hype.

Trend 2: Data Governance Becomes a Budget Driver, Not a Side Project

The second trend is less sexy and therefore more dangerous: data governance is becoming one of the biggest cost blocks. Salesforce describes in its guides on data governance tools and enterprise data warehouses that AI can automatically classify data, detect errors, and flag compliance risks. That's useful. But before a tool can classify data cleanly, a company needs to know what data it owns, who is responsible for it, what it means, and whether it can be used. Sounds trivial. It's not. I've seen CRM systems where “industry” was a free text field and “mechanical engineering” appeared in eleven variants, including “Mech. Eng.”, “Mechanical Engineering DACH”, and “OEM maybe”.

An enterprise data warehouse or a data lakehouse in an SME doesn't just cost a platform license. Typical implementation budgets, depending on the starting situation, range from 0.5 to 3 million euros. In addition, there are annual operating costs in the six-figure range. Data governance tools with AI functions for many medium-sized companies roughly range between 50,000 and 300,000 euros per year, plus internal roles: Data Owner, Data Steward, Governance Lead, sometimes an ML-Ops profile. Those who don't plan for this later call it “unexpected complexity,” which was actually foreseeable work.

From our implementations, we know: In B2B sales projects with industrial customers, an average of 62 to 74 percent of the operational work in the first 8 to 12 weeks goes not into AI logic, but into data access, field mapping, duplicate logic, ICP refinement, and approval processes. The most common bottleneck is not a bad model. It's the question of whether “Kärcher Dealer,” “Alfred Kärcher SE & Co. KG,” and “Kaercher Export” are the same economic account, whether the Winnenden location or a subsidiary is meant, and whether sales is even allowed to sell this structure. This work is not in any glossy report. But it determines whether AI in sales generates appointments or just data salad with better grammar.

Data governance sounds like control. In reality, it's growth protection. Without clear data responsibility, a company scales wrong assumptions. A sales manager from Augsburg, Stefan, told me in June 2025 after a workshop: “We thought we had a lead problem. We had a definition problem.” Exactly. If marketing, sales, service, and product management use different terms for the same customer type, AI cannot magically create clarity. It can industrialize ambiguity.

Why the AI Act Makes Costs Not Just Legal

The EU AI Act is often dismissed as a legal matter in SMEs. Mistake. It changes the operating model. Risk classification, documentation, transparency, monitoring, human oversight – these are not just paragraphs, but work packages. Many European companies are now calculating additional compliance costs of 5 to 15 percent of the AI project budget, depending on the risk class, data type, and industry. For a 700,000-euro program, that's not peanuts. And if AI slips into quality control, credit decisions, HR, or safety-related processes, a “tool” becomes a system with proof obligations.

In DACH, there's another factor: co-determination. Anyone who introduces AI on the shop floor, evaluates performance data, or builds automatic recommendations for clerks quickly encounters works councils, data protection officers, and labor law. I'm not saying that's bad. Not quite – sometimes it's tough. But it's reality. A factory in southern Germany doesn't smell like a pitch deck, but like cooling lubricant, pallet wood, and old responsibilities. Anyone who wants to introduce an AI system there must explain what it does, what it doesn't do, what data is stored, and who decides in case of conflict.

Amplifa ICP Playbook A practical framework to clearly define target customers, signals, and sales priorities before AI automation.

Trend 3: Integration Eats the ROI Story

The third trend is the most expensive: integration. Not as a buzzword, but as a bill. AI is supposed to integrate with SAP, Microsoft Dynamics, Salesforce, HubSpot, Siemens Teamcenter, Dassault, MES, PLM, PIM, DMS, CPQ, and sometimes a self-built Access tool from 2009. Interface development for ERP, MES, or PLM quickly costs 150,000 to 500,000 euros for larger projects. In addition, there are testing, rights, roles, security hardening, data migration, monitoring, and maintenance. Analyst reports then state “productivity boost through AI integration.” In practice, the calendar says: alignment Monday 9:00, data model Wednesday, escalation Friday.

I'm blunt here: Anyone who still believes in 2026 that AI can deliver substantial ROI in SMEs without clean system integration is confusing demo with operation. Of course, there are isolated productivity gains. An employee has an email formulated. A service team summarizes tickets. A salesperson researches faster. Nice. But the big leverage only arises when AI intervenes in the process: prioritize accounts, recognize conversation triggers, prepare offers, write back data, trigger the next action, measure success. A chat window is not enough for that. You need architecture for that.

An example from sales: Markus's team, a technical component manufacturer from Franconia with around 85 million euros in revenue, booked three times as many qualified initial appointments in 9 months – without a single new sales employee. Not because a language model wrote particularly charming emails. The leverage lay in the interplay of ICP, account signals, CRM hygiene, outreach logic, and clear exclusion criteria. We automated less than originally planned and cut more. Away with irrelevant industries. Away with accounts without triggers. Away with people who were in the CRM but never influenced specifications. AI accelerated. Strategy decided.

Backup, Recovery, and Resilience: The Forgotten Block

One cost block is particularly often forgotten: resilience. AI systems generate logs, versions, prompts, outputs, model states, training data, evaluation data. If a system provides recommendations for offers, quality decisions, or service cases, someone will later want to know why. Not sometime. In the audit. In the complaint. In court, perhaps. Salesforce refers to AI agents in enterprise backup solutions that can reduce service costs and ensure data integrity. Nice. But extended backup and recovery solutions with AI functionality quickly cost 50,000 to 200,000 euros per year for medium-sized environments, plus storage, network, and process design.

For manufacturers with multiple plants, this becomes an operational question: Which data remains in the plant? Which goes to the cloud? Which must remain in Europe due to customer contracts? What happens if a line stops because an AI-supported inspection system is unavailable? At Festo or Phoenix Contact, the structures exist to systematically answer such questions. For a toolmaker with 220 people, it often depends on an IT manager who simultaneously handles firewall, ERP updates, telephone systems, and now also AI governance. This is not a technology deficit. This is organizational reality.

How High Are Hidden AI Costs Really?

The short answer: For medium-sized manufacturing companies with 100 to 5,000 employees, serious AI programs over 3 to 5 years often have total costs of 1 to 5 million euros. Not per chatbot. For a program with multiple use cases, data platform, integration, governance, operation, and change. A small pilot can cost 50,000 euros. A cleanly productively implemented use case with system integration often costs 250,000 to 900,000 euros. A portfolio of sales, service, quality, and supply chain quickly exceeds that. Honestly? I don't know for every company. But I do know that the first number is almost always too low.

The second answer is more uncomfortable: ROI is not only extended by costs, but by time. If a pilot takes three months and integration twelve months, then amortization is not visible in the quarter, no matter how pretty the demo was. In sales and lead generation, it can go faster because data sources are often more accessible than machine and quality data. But even there: Without ICP, without a signal model, without CRM feedback, and without sales acceptance, AI remains activity theater. Many companies celebrate more generated leads. The board asks about revenue six months later. Then it gets quiet.

For production use cases, the value is often higher, but the path is longer. Predictive maintenance can reduce downtime. Computer vision can reduce scrap. Supply chain optimization can reduce capital tie-up. But each of these cases requires data history, process understanding, clear responsibilities, and an error tolerance strategy. A wrong lead costs time. An incorrectly released part costs money, customer trust, and in the worst case, safety. Therefore, governance and testing are not bureaucracy. They are insurance against expensive stupidity.

What This Means for European SMEs

For European SMEs, the question shifts from “Are we doing AI?” to “What AI can we afford to operate?” This sounds defensive, but it's strategic. Large corporations like Siemens, Bosch, Schaeffler, Trumpf, DMG Mori, or Webasto can build platform teams, finance internal AI labs, and offset errors across portfolios. The classic hidden champion with 250 million euros in revenue cannot do that to the same extent. They have to make sharper decisions. Less playground. More capital discipline.

I see three implications. First: AI budgets are moving from innovation departments to the lines. Sales, service, operations, and engineering must take on their own profit responsibility. Second: CFOs will demand AI FinOps – i.e., cost control for tokens, cloud, user licenses, storage, integration effort, and external partners. Third: Purchasing will have to learn that the cheapest license price is rarely the cheapest operation. In a conversation in Munich in May 2025, Julia, CFO of an electronics manufacturer, said: “I won't sign any more AI contracts until someone writes down the exit costs on one page for me.” Good sentence. Should be on every boardroom wall.

This will also be relevant for investors. Those who invest in medium-sized industrial companies often look at EBITDA, working capital, order backlog, and degree of automation. In the future, AI TCO will be part of due diligence. Not as a slide on future viability, but as a hard examination: What platform dependencies exist? What is the company's data quality? Are there internal roles? Is AI Opex properly measured? Has sales become more productive through automation or just louder? I would not evaluate an industrial company in 2026 without asking these questions.

DACH is More Cautious – And That Can Be an Advantage

DACH companies are often considered slow when it comes to AI. That's partly true. Decision-making processes are longer, data protection is taken seriously, works councils have a say, and a managing director in East Westphalia prefers to ask twice before pushing customer data into an unfamiliar model. This annoys providers. Me sometimes too. But it can be an advantage if it leads to better TCO discipline. The USA is faster in rollout. Europe needs to be better in operation. That's not a consolation prize, that's a strategy.

The mistake would be to confuse caution with stagnation. Anyone who still relies on a pure inbound strategy in B2B sales in 2026 will have no pipeline in five years. But anyone who blindly applies AI automation to bad data builds a machine that processes the wrong accounts faster. Both are expensive. The art lies in the sequence: understand the market, sharpen the ICP, build the data model, integrate processes, then scale. Not the other way around.

Amplifa Product Amplifa combines ICP, account signals, and AI-powered sales processes for predictable B2B pipeline in SMEs.

Preparation: 7 Steps Before AI Costs Spiral Out of Control

I don't like checklists when they replace thinking. This one is meant to force thinking. If you are a managing director, investor, or strategy manager in a manufacturing company planning an AI program in 2026, you should clarify these seven points before budget approval. Not afterwards. Beforehand.

1. Calculate 5-year TCO instead of pilot budget: Include licenses, cloud consumption, API costs, storage, backup, external partners, internal FTEs, training, governance, security, and exit costs. A pilot budget of 150,000 euros is not investment logic if the rollout ties up 900,000 euros.
2. Conduct a data inventory with responsible parties: Define which data sources are relevant, who owns them, what their quality is, and which fields are business-critical. In sales, this means, for example: account structure, industry, revenue class, triggers, contacts, buying center, exclusion criteria.
3. Evaluate use cases by integration level: An isolated assistant is to be calculated differently than a system that writes to SAP, Salesforce, Teamcenter, or MES. Evaluate each use case by data access, process proximity, risk, number of users, and operational effort.
4. Set up FinOps for AI: Define cost limits, monitoring, consumption reports, and responsibilities for tokens, cloud, storage, and platform licenses. Without FinOps, AI Opex becomes a smokescreen in the monthly financial statements.
5. Don't delegate governance to legal: Risk classification, documentation, audit trails, model monitoring, and human oversight require an operating model. Legal reviews. The business operates.
6. Monetize change costs: Include training, productivity dips, process restructuring, and new roles. If a sales process is changed by AI, this affects target systems, compensation, leadership, and forecasting.
7. Clarify exit and lock-in before signing the contract: Check how data can be exported, which models are interchangeable, which integrations become proprietary, and what a vendor change costs. The most expensive contract is often the one you can't get out of.

A Simple TCO Model for Managing Directors

When I work with managing directors, I like to use a rough formula. It's not scientifically perfect. But it prevents self-deception. Total costs over five years = external project costs + software licenses + cloud and compute + data platform + integration + internal FTE + governance and compliance + security and resilience + change + risk buffer. The risk buffer should not be 5 percent. For complex integrations, I consider 20 to 30 percent more realistic. If you flinch now, you've understood the point.

AI Costs in Sales: Why Amplifa is So Strict Here

Since we at Amplifa work in sales, I see AI costs there particularly sharply. Many companies buy AI for sales because they want more pipeline. Understandable. But pipeline doesn't arise from text generation. Pipeline arises from the right target customers, clear signals, timing, conversation triggers, clean handover, and consistent follow-up. AI can support each of these steps. It can also inflate each of these steps. An SDR team that previously processed 500 mediocre accounts can process 5,000 mediocre accounts with AI. The calendar gets fuller. Revenue doesn't necessarily.

What we specifically see at Amplifa: For industrial companies with complex products, the appointment rate rarely increases through more personalization alone. The strongest effect comes from better account selection. In projects in 2024 and 2025, the difference between broadly automated outreach and ICP-sharp, signal-based outreach was often a factor of 2.4 to 3.7 in qualified appointments per 1,000 target accounts. The cost per appointment did not fall because AI wrote cheaper texts, but because fewer wrong accounts were contacted. This is an uncomfortable finding for tool providers. For managing directors, it's worth its weight in gold.

An example: A manufacturer of automation components, not Festo, but in a similar market segment, wanted to double its outbound activity with AI in January 2025. The original goal was 12,000 contacts per quarter. We cut to 3,800 relevant accounts, based on industry, machine park indicators, location structure, hiring signals, certifications, and triggers from investment announcements. Result after two quarters: fewer emails, more conversations, better acceptance in sales. The CSO, Daniel from Ulm, then said: “I wanted automation. What I got first was discipline.” That's exactly how it should be.

Amplifa for B2B Pipeline Strategy, data, and AI-powered implementation for medium-sized B2B sales teams with complex target markets.

Why Investment Sums Give False Security

Investment sums are impressive. 80 billion US dollars in venture capital in a peak year. Double-digit billions for hyperscalers. Series A and Series B rounds between 20 and 200 million US dollars for AI startups. Corporate venture arms of Siemens, Bosch, BMW, or Schneider Electric financing Industry 4.0 and AI startups. That sounds like market validation. It is. But it says nothing about whether a medium-sized manufacturer has its own cost structure under control.

Investors finance growth, not your data cleansing. Hyperscalers finance infrastructure that later needs to be sold. SaaS providers bundle AI into packages because it facilitates expansion. Analysts write about market sizes because markets are easier to model than internal friction. No one sits on your steering committee in the end when the CIO says that the PLM system cannot be properly connected, sales does not use the recommendations, and data protection still requires a risk analysis. These costs are not hidden because they are invisible. They are hidden because no one likes to put them on the first slide.

The Provider Mechanics: Bundling, Consumption, and Lock-in

The market mechanics are clear. Microsoft bundles AI into M365, Dynamics, and Azure. Salesforce integrates Einstein and Data Cloud. ServiceNow, Oracle, SAP, Google, and AWS build AI functions into platforms. This is convenient. It lowers entry barriers. But it also increases lock-in. Once workflows, data models, authorizations, and automations are deeply embedded in a platform, a change becomes expensive. Not impossible. But expensive enough that it is never seriously discussed in many boardrooms.

In addition, there are complex pricing models: user-based, transaction-based, token-based, volume-based, sometimes combined. Purchasing is trained on license negotiations, not on variable AI consumption. CFOs are now more familiar with cloud FinOps than five years ago, but AI FinOps is still young. If a sales team suddenly scales automatic account research, call summaries, email variants, CRM updates, and signal evaluations, consumption does not grow linearly with the number of users. It grows with process design. That's a different logic.

Regional Perspective: DACH, Europe, Global

DACH has a special AI cost structure. Industrial density is high, processes are complex, data landscapes are historical, customer requirements are strict. A machine builder in Baden-Württemberg has different integration problems than a US SaaS company. SAP is deeply rooted, MES systems are individually adapted, PLM is critical, Excel lives on (sometimes very vibrantly), and customers like BMW, Mercedes-Benz, Airbus, or Siemens demand proof, certifications, and reliability. That costs money.

Europe exacerbates the governance issue. The AI Act introduces risk classification and transparency obligations. GDPR remains. Industry regulations are added on top. In France, the Nordics, and Benelux, I see more funding programs and partly faster digital groundwork. In Germany, I see more skepticism and often better process discipline once a project has been approved. Globally, the USA and Asia are more aggressive. Large industrials and tech companies invest 10 to 100 million US dollars per year in AI programs. China, Korea, and Japan are driving manufacturing AI at high speed. But speed without governance is not free. It shifts costs into risk.

For DACH SMEs, this means: You don't have to imitate the hyperscaler. You have to win selectively. A machine tool manufacturer doesn't have to start 40 AI use cases. It might need five: offer support, spare part identification, quality analysis, supplier risk, sales signals. If these five are cleanly integrated, they beat 40 pilots in an innovation newsletter. I know that sounds unromantic. Good.

Business Impact: Where AI Really Makes Money

The most important business question is not: Which AI technology do we use? It is: Which bottleneck becomes economically smaller? In sales, this can be a lack of qualified conversations. In service, it can be the processing time of technical inquiries. In production, it can be scrap or downtime. In engineering, it can be the time until variant evaluation. Any other discussion is tool fetishism.

In medium-sized manufacturing companies, I see four value levers that can justify AI costs. First: better market coverage with the same sales team. Second: shorter response times for technical inquiries. Third: fewer errors in repetitive document and data processes. Fourth: better decisions on inventory, quality, and maintenance. These are not fantasies. But each lever needs measurement points before the start. If no one knows how many qualified conversations per 1,000 accounts are generated today, no one can later prove that AI in sales has worked.

In a customer project in mechanical engineering, we first measured the baseline before automation: 1,000 target accounts, 42 relevant responses, 11 qualified appointments, 3 real opportunities. After ICP sharpening and AI-powered signal prioritization, there were 27 qualified appointments and 8 opportunities for a comparable number of accounts. Not perfect. But measurable. The managing director was ultimately less interested in the model than in whether order intake picked up six months later. Rightly so. AI is not an end in itself. It is an expensive tool that either solves a bottleneck or just looks modern.

FAQ: When is AI Worth It Despite High Costs?

AI is worthwhile if an economically relevant process has enough repetition, enough data access, and enough decision volume for automation or assistance to have a noticeable effect. An offer process with 20 complex offers per year may not be a good first AI case. A spare parts process with 18,000 inquiries per year is more likely. A sales team with an unclear ICP first needs strategy. A sales team with a clear ICP but too little market coverage can get real leverage through AI-powered lead generation. The order decides.

FAQ: Which AI Costs Are Most Often Forgotten?

Most often forgotten are internal personnel costs, data cleansing, integration tests, governance, training, and ongoing monitoring. Directly after that come backup, audit trails, security hardening, and vendor change costs. In conversations with CFOs, I often hear: “We have external costs under control.” I even believe that. Internal costs are the problem because they run through calendars, meetings, and delays, not through a clean invoice.

FAQ: Should SMEs Train Their Own Models?

Mostly no. Not at the beginning. Training your own foundation models is nonsense for almost all medium-sized manufacturers. Too expensive, too complex, too far away from the bottleneck. More sensible are existing models, retrieval approaches, clean data spaces, clear rights, and industry-specific workflows. Fine-tuning can make sense later if enough data quality, usage volume, and governance are available. Anyone who starts with their own model without clarifying their data terms builds a monument to technical vanity.

Personal Forecast: 2026 to 2028 the Market Divides

My forecast for the next two to three years: The AI market in European SMEs will divide into three groups. Group one will continue to pilot, use funding, build internal presentations, and scale little. Group two will buy large platform packages, then suffer from Opex, integration, and acceptance. Group three will start smaller, prioritize harder, calculate TCO cleanly, and treat AI as an operating system for specific bottlenecks. This third group will not be the loudest. It will have the better pipeline, the more stable processes, and the more resilient margins.

I also believe that AI costs will get their own name in boardrooms. Today, they disappear into IT, digitalization, sales efficiency, or operations. By 2028, many SMEs will have AI Opex reporting, similar to cloud FinOps. Not because controlling enjoys new columns. But because variable model costs, platform licenses, data maintenance, and governance would otherwise grow into a gray block. And gray blocks are the enemy of any investment discipline.

SMEs have an advantage that many underestimate: proximity to the problem. A managing director at a manufacturer in Heilbronn often knows his bottlenecks more precisely than a corporate board member with 14 transformation teams. If this proximity is combined with data clarity and capital discipline, AI can bring a lot. But not as magic. As work. As a cost center with a return claim. As a system that you operate.

A few days ago, a sales manager from Cologne sent me a message after an internal AI review. Just one sentence: “We cut three use cases today and feel faster for the first time.” That's perhaps the most mature AI strategy I've seen this week.