AI Investments: Hidden Costs in Mid-Sized Businesses

Marktanalyse · 16. Oktober 2026 · Anthony Filipiak

AI investments cost more than licenses show. Review data, integration, and operations before your pilot becomes an expensive cost block.

On October 14, 2026, Europa Press reported, citing the Stanford AI Index, that global private AI funding reached approximately $344.7 billion in 2025 — with about $285.9 billion in the USA and only around $20.9 billion in Europe. The same data set reveals public European programs like InvestAI with €20 billion and larger public-private AI projects up to €200 billion. This sounds like a catch-up race. Well, almost. For managing directors in DACH mid-sized businesses, the headline about AI investments is less important than the question of how much of these billions actually reach their own factory, their own sales, their own margins — and how much remains as infrastructure costs with Microsoft, AWS, Google, Nvidia, SAP, or Siemens.

My surprising prediction: In the next two to three years, many mid-sized manufacturing companies will spend more money on data cleansing, integration, governance, and model operations than on the actual AI software. Not just slightly more. Significantly more. Anyone building an AI business case in 2026 that ends with license fees and API costs is sugarcoating the pilot and ruining the rollout.

I'm not writing this as an analyst from a glass office. I'm writing it based on conversations with managing directors, sales managers, and IT managers at manufacturers, suppliers, and technical service providers who are realizing that AI doesn't fail at the demo, but on the Monday morning after. At the SAP bill of materials. At the MES, which was maintained differently than expected. At the works council. At old item numbers that smell like a damp archive basement — and in some factories, are still exactly there.

AI Investments in the Status Quo — Lots of Money, Wrong Interpretation

The visible market is exploding. According to Menlo Ventures, enterprise spending on generative AI rose from $11.5 billion in 2024 to $37 billion in 2025. Of this, approximately $19 billion was for AI applications and about $18 billion for infrastructure; Foundation Model APIs accounted for the largest block within infrastructure at $12.5 billion. This figure is important because it shows where the initial margin flow goes — to model providers and platforms, not automatically into productive processes at Brose, Schaeffler, Trumpf, Phoenix Contact, or a toolmaker with 850 employees in East Westphalia.

Europe is investing. But differently. The gap in late-stage growth capital is brutal: around $12 billion in late-stage AI funding is cited for Europe in 2025, compared to approximately $141 billion in the USA. This is not just a VC problem for founders in Paris, Munich, or Stockholm. It also determines whether European providers can finance their own base models, industrial cloud infrastructures, and vertical platforms — or whether DACH companies will have to build their AI layer on US infrastructure and later contend with price increases, data residency issues, and interface changes.

In mid-sized businesses, there's an additional factor missing from many market graphics. The calculation doesn't start at €29 per user per month. It starts with the question of whether the data from SAP, MES, SCADA, PLM, CRM, quality management, Excel lists, and local Access databases even fit together. Thomas, CFO of an Augsburg-based mechanical engineering supplier, told me in March 2026: "The vendor's demo was clean. Our data wasn't." That's the sentence I hear more often than any vision of autonomous factories.

What Analyst Figures Show — and What They Conceal

Analyst reports are useful. I read them. I also use them in client discussions because investors and advisory boards need numbers. But they often cut the market in the wrong place. Visible are funding rounds, cloud revenues, software licenses, and API consumption. Less visible are internal working hours of process engineers, data migration from 14 years of production history, works council approvals, new role models in SAP, audit documentation for the EU AI Act, security tests against prompt injection, and the quiet work of a quality engineer relabeling 18,000 images because the camera was misaligned in 2021.

That's the real cost block. Not sexy. No exhibition stand at Hannover Messe will prominently display it. At Hannover Messe in April 2026, I heard almost the same pitch in two halls: industrial copilots, secure data spaces, AI for maintenance, AI for sales, AI for engineering. What I heard less often: Who maintains the ground truth after a product change? Who pays for parallel operation? Who is responsible for a model if it runs technically but slowly deteriorates in terms of content?

Trend 1 — AI Investments Shift from Models to Operating Costs

The first trend is uncomfortable: the model becomes a smaller item. Not always in the initial quote, but in the real three-year budget. A manufacturer can access language models, search documents, classify emails, or evaluate technical manuals for a few thousand euros a month today. But as soon as the system becomes production-relevant, the cost shift begins. Then it's about data quality, interfaces, rights, logging, monitoring, high availability, and acceptance. Sounds like IT minutiae. It's not. It's the difference between a demo for the board and a system that doesn't fall apart during a night shift in Plant 2.

In manufacturing, data is rarely located where consultant slides suggest. SAP knows orders, materials, delivery dates. MES knows machine states. SCADA knows signals. PLC controls provide cycles, error codes, temperatures. Quality management knows test plans and complaints. Maintenance knows disruptions, but often as free text. Sales and service know customer context, but that's in CRM fields, PDF quotes, email threads, and sometimes in the head of Bernd, who has been managing key accounts for 19 years. Anyone who wants to build reliable AI from this doesn't just pay for tokens. They pay for translation work between systems, teams, and realities.

I'm being deliberately blunt here: treating AI rollout as a software purchase burns budget. The companies that win treat AI as an operating model. Not as a tool. A predictive maintenance model must capture machine signals, classify states, generate maintenance orders, reserve spare parts, touch personnel planning, and evaluate false alarms. If only the first step works, you have a nice graphic. Not a maintenance strategy.

We don't have too little AI. We have too few clean handovers between data, process, and responsibility.

— Andrea, Head of Sales Operations at a mechanical engineering supplier, Bielefeld

Andrea told me this sentence in May 2026 after a workshop on account prioritization and quote data. Her team wanted to use AI to identify better target customers. The license was approved. The data situation was not. In the CRM, 38 percent of industry fields were empty or incorrect, quotes from two product lines were only available as PDFs, and for 11 key accounts, it was unclear whether revenue development came from price increases, volume growth, or spare parts business. This is not an exception. This is everyday life.

The Hidden Costs — Where the AI Business Case Breaks

When I talk to managing directors about AI investments, the same question almost always comes up first: What does the tool cost? Wrong question. The better question is: What does the state cost in which the tool can operate economically? Not quite right — both questions are necessary. But the second one determines the ROI.

Data preparation is the first big chunk. A mid-sized factory with 500 to 2,000 employees can easily spend several hundred thousand euros on a well-planned data and integration project before the first use case scales. Not because providers are greedy. But because historical data is missing, timestamps are misaligned, item numbers are inconsistent, sensors drift, shift notes were manually added, and process changes were never properly documented. In one factory, the server room still smelled of warm dust when the IT manager explained to us why the oldest production database only runs on a virtual Windows server that no one wants to touch. That's exactly where AI costs begin.

Integration is the second big chunk. API development, middleware, ERP connection, MES coupling, historian databases, edge computing, role models, monitoring, logging, backups, version management for models and data. This is not a footnote. In a visual quality inspection project, the software might cost €100,000, while cameras, lighting, industrial network, labeling, QMS integration, training, model maintenance, QA acceptance, and escalation processes push the total amount into a different league. Anyone who only sees the software price is looking at the tip of the cost iceberg.

Then comes MLOps. Many business cases calculate development but not operations. Models need to be monitored. Data and concept drift need to be detected. New model versions require testing. False alarms need to be analyzed. Cloud or GPU usage fluctuates. Documentation must be auditable. Support needs accountability. A model can continue to run technically even if raw materials, suppliers, machine parameters, or product variants have changed. It doesn't lie then. It's just not as good as everyone believes.

The fourth chunk is organizational. AI rarely replaces an entire process immediately. Usually, parallel operation emerges. Employees check suggestions. Quality engineers control misclassifications. Maintenance personnel validate maintenance forecasts. Buyers review automatically generated purchase suggestions. Lawyers check generated content. This control layer is not a sign of backwardness. In production, a wrong suggestion can mean scrap, downtime, or liability risk. So, personnel requirements often don't decrease at first. They shift.

And then compliance. The EU AI Act forces companies to clarify more precisely whether a system merely recommends or influences decisions — for example, in quality, safety, personnel, or production control. In addition, there's access protection, data residency, supplier vetting, prompt logging, model logging, emergency procedures, and protection against data leakage. An internal knowledge system that processes design data, recipes, customer prices, or service reports is not a nice chatbot. It's a new target for attacks.

Trend 2 — Europe Becomes Strong in Applications, Dependent in Infrastructure

The second trend: Europe will not win by every mid-sized company building its own Foundation Model. Honestly? I don't know why this fantasy still appears in some strategy rounds. For most DACH manufacturing companies, it's economically nonsensical. The strength lies in applications on proprietary data: production parameters, quality patterns, service histories, design variants, quote histories, technical specifications, machine states, and customer behavior.

The competitive landscape is clearly stratified. Hyperscalers like Microsoft, Amazon Web Services, and Google sell cloud, model, and data platforms. Model providers like OpenAI, Anthropic, Google, Meta, and European providers compete on Foundation Models and APIs. Industrial platforms like Siemens, SAP, Dassault Systèmes, PTC, Schneider Electric, and Bosch integrate AI into engineering, ERP, automation, and maintenance. Specialized providers handle quality inspection, predictive maintenance, production planning, document automation, or industrial copilots. System integrators like Accenture, Capgemini, Deloitte, T-Systems, and regional specialists take on the hard work of integration.

This creates a new dependency for customers. If $12.5 billion in enterprise infrastructure spending is allocated solely to Foundation Model APIs, that's not just a market indicator. It's a warning signal. API prices change. Token consumption increases. Models are deprecated. Proprietary interfaces shift. Data protection requirements tighten. A purchasing manager from Stuttgart told me in June 2026 after a discussion about AI-supported quote qualification: "I don't want to find out in three years that our sales knowledge is trapped in a pricing model." That's exactly the point.

From our implementations, we know: For industrial customers in the DACH region, the first bottleneck is almost never the language model. In the last 12 months at Amplifa, in projects with mechanical engineers, technical service providers, and B2B suppliers, we have seen that 30 to 45 percent of relevant CRM and market data had to be cleaned, supplemented, or re-mapped before productive use. Particularly common: incorrect industry classification, duplicate company locations, outdated contacts, unmaintained buying center roles, and quote data only available as PDFs in folder structures. This takes time. But without this step, AI in sales produces pretty lists — and bad priorities.

This point is bigger than sales. It applies to production, service, purchasing, and engineering. Owning proprietary data does not yet mean having an advantage. The advantage only arises when this data is clean, accessible, contextualized, and legally usable. DACH companies have a real opportunity here because their process expertise is deep. Wittenstein understands mechatronic systems. Festo understands automation. Trumpf understands lasers and sheet metal processing. Kärcher understands equipment fleets and service. But knowledge within the company is not automatically AI-ready knowledge.

Amplifa ICP Playbook A practical guide to cleanly structure target customers, data gaps, and prioritization in B2B sales — before AI scales wrong accounts.

Trend 3 — The Pilot Graveyard Becomes a Balance Sheet Problem

The third trend is the most expensive because it rarely appears on a single cost center: unscaled pilots. A pilot for visual quality control here. A chatbot for technical documentation there. A predictive maintenance idea in Plant 1. A sales assistant in the export team. A copilot for purchasing. Each pilot on its own seems justifiable. €50,000. €120,000. An internal team, two external consultants, a cloud budget, a proof of concept. After 18 months, the advisory board asks what's running productively — and the room falls silent.

The damage is not just from money spent. It's from lost management attention, disappointed departments, and a cynical organization. "We've had that before, it didn't do anything," Daniel, production manager at a plastics processor near Nuremberg, told me in February 2026. The sentence didn't come from technophobia. It came from fatigue. Three pilots, no rollout, no clear responsibility, no baseline metrics. The smell of warm plastic in the hall was more real than any AI roadmap in the meeting room.

Why do pilots end up in the graveyard? Mostly for the same reasons: no process owner, no reliable baseline, unavailable production data, missing integration, unclear liability for wrong decisions, no budget for operations, too strong a dependence on external project teams. And a particularly German classic: the pilot is secured for so long that it's no longer allowed to change anything. Then it's safe. And useless.

FAQ — Why do AI pilots fail despite good models?

Because a good model doesn't save a bad process. An AI system can only deliver economic benefits if data access, process responsibility, user acceptance, operating model, and escalation logic are clarified. In a scrap reduction project, a high detection rate is not enough if the line doesn't know who decides in case of a false alarm. For AI Sales, lead scoring is not enough if sales doesn't have clean ICP criteria and marketing continues to count MQLs based on form downloads. For predictive maintenance, a forecast is not enough if spare parts have a six-week delivery time.

I disagree with a popular thesis here: "We just need to experiment more." No. Mid-sized businesses have experimented enough. What's missing is the discipline to define the scaling conditions before the experiment. What data is needed? Which interfaces? Who operates the system? What happens in case of wrong decisions? Which metric needs to improve? At what threshold is it stopped? These questions are boring. That's exactly why they save money.

What Hidden AI Costs Specifically Mean for Mid-Sized Businesses

For managing directors, the core question is not whether AI works. Of course, it works in many areas. The question is whether it works in their own economic system. A manufacturing company with 1,200 employees, three plants, SAP ECC, an older MES, several special machines, decentralized quality assurance, and a sales department with 17 evolved Excel processes has a different cost curve than a SaaS company with a clean data warehouse. Ignoring these differences means importing Silicon Valley calculations into a brownfield world.

The economic threshold is often higher than the first pilot suggests. A single use case can start small. A plant-wide rollout requires data and process analysis, integration, cloud or hardware resources, training, security review, compliance clarification, and several months of operational stabilization. ROI should therefore not only be measured by personnel savings. In manufacturing, scrap rate, first-pass yield, unplanned downtime, OEE, maintenance costs per asset, energy consumption per production unit, lead time, planning quality, complaint rate, and time to root cause analysis count.

It's similar in sales. Anyone who understands AI only as email automation will scale bad mass outreach in 2026. Anyone who uses AI as a data and prioritization system can manage pipeline differently. Markus, sales manager at a plant manufacturer from Heilbronn, told me in April 2026: "We had enough leads. We just didn't know which ones were really ready to buy." His team didn't send more emails afterward. They tightened the ICP, maintained trigger data, re-evaluated existing customer potential, and corrected old industry labels. Result after nine months: significantly fewer irrelevant initial meetings and almost three times as many qualified appointments in the defined core segment — without hiring a new SDR.

That's the point: AI doesn't automatically create demand. It can focus attention, reveal patterns, accelerate processes, and prepare decisions. But it also amplifies garbage. Bad data gets worse faster. Wrong target customers are processed more consistently incorrectly. An unclear process doesn't become clearer just because a model looks at it.

The Brownfield Factor — Why DACH is More Expensive, But Not Without Opportunity

Germany has a strong industrial base and a lot of proprietary data that is valuable for AI. At the same time, IT and OT landscapes are fragmented. Old SAP instances, specialized software from the 2000s, machines with proprietary interfaces, local databases, production lines that were never built for real-time data. At a supplier in Baden-Württemberg in September 2025, an old industrial PC was still standing next to a machine, its fan sounding like a hairdryer running continuously. Part of the quality data needed for an AI project was on this computer. No one had drawn it on the architecture slide.

This doesn't make DACH weak. It just makes the path different. The advantage lies in proprietary machine data, complex manufacturing, quality expertise, engineering depth, industrial customer relationships, and automation networks. The disadvantages lie in lower access to growth capital, skilled labor shortages, slow procurement, data protection coordination, works council processes, heterogeneous brownfield systems, and a lack of scaling beyond individual plants. Those who accept this reality can make good decisions. Those who gloss over it with a cloud demo will pay later.

For most mid-sized companies, a hybrid model is more economical: sensitive data and critical production functions remain in controlled environments, standardized language or analysis functions run via cloud APIs where appropriate. This is less glamorous than "everything into the AI platform." But it better suits liability, availability, data protection, and cost control. Especially in supply chains near Webasto, Schaeffler, Phoenix Contact, or Festo, trust is not a decorative issue. It decides on approvals.

Calculating AI Investments Correctly — TCO Instead of License Comparison

I would calculate every AI business case in mid-sized businesses over three to five years. Not over six months of pilot runtime. The TCO must include license, API consumption, cloud, hardware, data preparation, integration, training, operations, governance, security, model monitoring, support, process transformation, and exit costs. Yes, that makes the number bigger. That's exactly the point.

A cheap API price can quickly be overcompensated by usage peaks, data migration, monitoring, new compliance requirements, and vendor switching costs. Particularly insidious are systems that seem cheap at first but don't allow prompts, logs, fine-tuning data, or interfaces to be cleanly exportable. Then the company later finds itself in a negotiation where the vendor changes the price list and its own portability only exists on paper. No CFO likes this moment. Some experience it anyway.

TCO is also important because public investment sums in the market must not be confused with mid-sized business budgets. When the EU announces €20 billion for InvestAI or public-private partnerships up to €200 billion are discussed, a large part flows into data centers, semiconductors, research, and large projects. That helps the ecosystem. But it doesn't automatically pay for cleaning your item master data, connecting MES in Plant 3, or qualifying your maintenance technicians for model validation.

Preparation — 7 Steps Before the Next AI Pilot Starts

If I could give a managing director in a mid-sized company only one piece of advice, it would be this: Don't start with "We need AI." Start with an expensive, measurable problem. Scrap. Downtime. Energy consumption. Planning deviation. Lost quotes in the core segment. Complaints. If the problem isn't measurable, the pilot will later become political.

1. Define a hard baseline. Before the first model, it must be clear what the current scrap rate, OEE, downtime, lead-to-opportunity rate, quote turnaround time, or complaint rate is. Without a baseline, every success is an opinion.
2. Map data sources with owners. SAP, MES, SCADA, PLM, CRM, QMS, Excel, local databases, and supplier systems need technical owners. Not "IT." Specific names. Specific responsibility.
3. Check brownfield capability before vendor selection. Ask about real integrations into ERP, MES, PLCs, quality, and sales data. A demo with sample data says little about Plant 2 in the Czech Republic or spare parts sales in Austria.
4. Plan operations before the pilot. Who monitors drift? Who tests new model versions? Who reacts to false alarms? Who documents decisions for audit and compliance? If the answer only comes after go-live, it's too late.
5. Specifically define human control. Low-risk decisions can be automated more. Critical production, quality, safety, or personnel decisions require escalation paths. A blanket four-eyes principle slows everything down; no control makes it dangerous.
6. Contractually secure portability and exit rules. Models, prompts, training data, logs, interfaces, and evaluations must remain exportable. The cheapest provider is expensive if you can't switch later.
7. Plan for scaling from Plant Two. A solution that only works on one machine with an enthusiastic project manager is not yet a platform. Check variants, languages, roles, data models, and support across locations.

Amplifa Product Amplifa connects market, company, and sales data so B2B teams can prioritize better target customers and make data-driven pipeline decisions.

What Investors Should Evaluate Differently

Investors often look at two things in AI projects in mid-sized businesses: automation potential and software costs. I would focus more on data access, integration capability, and operational discipline. A company with a mediocre model but a clean data model and a clear process owner can create more value than a company with a strong model and chaos in master data, roles, and responsibilities.

For due diligence questions, "Do you use AI?" is no longer enough. The better question is: Which productive AI use cases have been running for more than six months with measurable KPI impact? How are model errors documented? How often are data sources checked? What dependence exists on individual cloud or model providers? Is there a budget for operations or only project budget? Those who can answer these questions cleanly are further along than many companies with a big AI roadmap.

This is especially true for companies with multiple plants. Value is not created in the first successful use case, but in the ability to transfer it in a controlled manner. A predictive maintenance model on one line is interesting. A repeatable data and operating standard for ten lines is enterprise value. An AI sales system for one region is nice. A unified ICP, signal, and prioritization model across DACH, Benelux, and France changes pipeline management.

FAQ — Which AI Investments Pay Off First?

First, use cases with clear cost impact, existing data, and limited risk pay off. In production, these are often scrap reduction, root cause analysis, energy optimization, maintenance prioritization, or quality documentation. In sales, it's ICP sharpening, account prioritization, trigger detection, quote analysis, and pipeline hygiene. I would be cautious with projects that touch many systems, have high liability, and lack a clean baseline. They can be important. But they are rarely good starting points.

For example: an AI assistant for technical documentation can quickly deliver value if documents are centrally available, access is cleanly regulated, and answers are supported by sources. A system that automatically changes production parameters is a different risk class. A lead scoring model on well-maintained account data can show impact after weeks. An autonomous sales agent that generates quotes without clear approval will eventually produce prices, delivery times, or commitments that someone has to explain. Usually the sales manager.

FAQ — How High Are Realistic AI Costs in Mid-Sized Businesses?

The honest answer: It strongly depends on the process and the data state. A limited pilot can start with a five- to low six-figure budget. A plant-wide rollout with data integration, security, training, operations, and compliance can quickly cost several hundred thousand euros; for multiple plants or regulated products, the amount increases significantly. Anyone who only calculates licenses underestimates the effort. Anyone who first checks data state, interfaces, operating model, and risk can cut projects smaller and still build them scalably.

I don't like blanket percentage rules, but I often see a pattern: The visible software item is prominent in the first quote, but later internal time, integration, and operations dominate. In sales projects at Amplifa, we see that companies achieve the greatest leverage not through more automation in outreach, but through better account selection and clean signals. In production projects, the leverage often lies not in the neural network itself, but in whether the plant, QMS, and maintenance see the same state.

Amplifa for Pipeline and Account Prioritization For mid-sized B2B teams who want to use AI not as a gimmick, but as a data system for better sales decisions.

My Personal Forecast Until 2028

By 2028, the market will sort itself out. Many generic AI tools will become interchangeable. Pricing power will remain with infrastructure, data access, and deeply integrated workflows. European companies will become stronger in applications, especially where industrial know-how, proprietary data, and regulatory trust matter. For Foundation Models and hyperscaler infrastructure, Europe will continue to struggle. Perhaps longer than politicians admit in press releases.

For DACH mid-sized businesses, I see three groups of winners. First, companies that take their data foundation seriously and don't clean it only when the AI project is already on fire. Second, providers who master brownfield integration and don't just sell pretty interfaces. Third, management teams who understand AI as an operational capability — with owners, KPIs, budget, and exit rules. These companies will not be the loudest. They will simply have better margin reports.

My counter-prediction: The biggest AI mistake in the coming years will not be investing too little. It will be investing incorrectly — in pilots without operations, in tools without a data foundation, in platforms without portability, in automation without process clarity. Anyone who still relies on a pure inbound strategy in sales in 2026 will have no pipeline left in five years. Anyone in production who only looks at model accuracy but ignores data and process costs may have a strong algorithm and still more downtime than before.

I don't believe in mid-sized businesses as victims of the AI wave. Their processes are too deep, customer relationships too robust, technical data too valuable for that. But I also don't believe in the narrative that a few licenses and an internal copilot are enough. The visible AI investments are just the tip of the cost iceberg. Beneath them lie data, integration, control, operations, and responsibility. That's where it's decided whether AI creates value in mid-sized businesses — or just produces new budget lines.

Perhaps that's the most sober prediction: The winners will not be those who can show the most AI projects in 2027. They will be those who have the least to explain about why their pilots didn't amount to anything.