Agentic AI: Strategy for SMEs

KI-Strategie · 22. Juni 2026 · Anthony Filipiak

Agentic AI in SMEs: How managing directors can start with real use cases, governance, and ROI. Read where agents will have a real impact in 2026.

“If the agent is just a chatbot, I won't pay a single euro for it.” Thomas, managing director of an automation supplier from Augsburg, told me that in March 2025. It sounds rough. But it's precise. Because it describes the core of the market better than most glossy slides about Agentic AI: SMEs don't buy artificial intelligence; they buy less downtime, fewer manual handovers, more service revenue, and an organization that still functions even when three experienced clerks retire.

I'm writing this deep dive from my perspective as CEO and Co-Founder of Amplifa. Every week, we talk to managing directors, CTOs, sales managers, and digital officers in DACH companies with 50 to 500 employees. Not innovation tourists. But people who have SAP documents, quotation backlogs, service tickets, works councils, and a margin on their desks. Agentic AI is no longer a sci-fi topic for these companies. Well, almost. The productive cases are there, but they are selective, more narrowly defined, and significantly less magical than LinkedIn would have you believe.

Why Agentic AI is now reaching SMEs

German SMEs have long played with AI like a new cordless drill from the hardware store. Just try it out. Build a demo. Show it to the advisory board. Then back in the cupboard. Since early 2025, that has changed. Not because the models can suddenly do everything, but because three pressures are acting simultaneously: costs in service and back office are rising, skilled workers are lacking in operational roles, and customers expect answers in hours instead of days. For a machine builder like DMG Mori, an unresolved service case can trigger follow-up costs worldwide. For a supplier with 180 employees from East Westphalia, a stuck spare parts process is enough to make a major customer nervous.

In a corporate context, Agentic AI does not mean the freely roaming AI employee who opens their inbox in the morning and has independently rebuilt the company by evening. Anyone selling that is selling theater. What is meant are multi-step AI agents that pull information from ERP, CRM, DMS, email, sensors, or data warehouses, prepare a decision or execute an action, and adhere to defined escalation rules. So, an agent doesn't just read a ticket. It checks machine type, serial number, maintenance history, spare part availability, SLA, technician skills, and then suggests the next step. Sometimes it initiates it directly. Sometimes it waits for approval. That's precisely where the difference between a chatbot and a digital worker begins.

The timing has also become sharp regulatorily. Since February 2, 2025, the EU AI Act mandates AI competence for providers and operators of AI systems. Article 4 sounds dry but hits the nerve: companies must be able to demonstrate that people who use or supervise AI are qualified to do so. Modular Ops points out in its AI workshops for SMEs that violations of certain obligations can result in fines of up to 7.5 million euros or 1.5 percent of global annual turnover. This is not a fringe issue for legal. This is a board-level issue. And yes, also for the 120-person company with Microsoft 365, abas ERP, and a sales manager who says “Copilot” but actually means “automation.”

Agentic AI is not a tool project, but process politics

Most Agentic AI projects don't fail because of the model. They fail because of power. Who gets to decide what a standard case is? Who is liable if the agent suggests a wrong spare part number? Who loses visibility if a process suddenly no longer runs through five email forwards? Andrea, Head of Sales at a hidden champion in Bielefeld, told me three weeks ago: “Our CRM isn't bad, but it doesn't reflect how we really sell.” Exactly this sentence also applies to agents. If the real process is dirty, the agent won't be clean. It only makes the dirt visible faster.

Therefore, I don't think much of the question: Which Agentic AI platform should we buy? The better question is: Which process deserves autonomy? At Trumpf, a service process is structured differently than at a plastics processor with 85 employees in Franconia. At Phoenix Contact, data architecture, governance, and product structure are different worlds than at a component manufacturer whose spare parts knowledge is stored in PDF folders, heads, and old tickets. Nevertheless, both want the same effect. Less friction. More speed. Controllable risks.

From our implementations, we know: The strongest predictor of success is not the choice of model, but the existence of a real process owner with budget authority. In 18 projects and pre-projects that we have accompanied at B2B SMEs in the last 12 months, the fastest progress did not come from the largest data sets, but from the cleanest responsibilities. If service management, IT, and management jointly determine in week one which decisions an agent is allowed to make and which not, the cycle time drops dramatically. In a sales and service setup with 220 employees, we reduced the time to qualified contact with existing customers from an average of 6.8 days to 1.9 days, without hiring a new sales employee. The agent was not brilliant. The process was finally clear.

What an agent really needs to do

A productive agent has a job description. Not a prompt poem. I want to see input, output, permissions, escalation path, audit log, KPI, and shutdown criterion. For example, if a technical support agent works for a packaging machine manufacturer, it must be clear which data sources it is allowed to read: service tickets, bills of material, manuals, IoT error codes, customer contracts. Then it needs a list of allowed actions: create diagnosis suggestion, identify spare part, prioritize ticket, suggest service appointment. Not: independently grant discounts or send safety-critical repair instructions without approval. Sounds trivial. Not quite. This clarity is missing in many companies because processes have evolved historically and no one likes to touch them.

Agentic AI data situation: early productivity, few hard benchmarks

Anyone looking for reliable benchmarks for Agentic AI in European SMEs today will not find a clean OECD table with 2,000 companies and seven years of history. Honestly? I don't know if we'll even get such a table in the next two years. What we have are case studies, vendor reports, consulting data, and consistent patterns from projects in DACH. That's not enough for academic certainty. But it is enough for business decisions if you read numbers as corridors and not as laws of nature.

In mechanical engineering service, the picture is clearest. The Munich platform lytra explicitly positions AI agents for service processes in mechanical and plant engineering. The typical setup: 200 to 5,000 employees, high export quota, complex machines, after-sales under pressure. According to publicly described case patterns, agents for technical support, spare parts business, and deployment planning work together there. One agent detects or classifies faults, a second identifies spare parts and initiates ERP processes, a third plans technicians according to location, availability, and skill. This is not glamorous. It is valuable.

Typical impact corridors, according to market reports and project data in the DACH region, are 20 to 40 percent faster time-to-resolution in technical service, provided that first-level diagnosis is partially automated. For standard tickets, an automation rate of 30 to 60 percent is often mentioned. In back-office processes, consultancies like Modular Ops speak of 25 to 50 percent less processing time per transaction and 30 to 60 percent fewer errors in data entry when PDF, email, and ERP routes are cleanly connected. I only like such numbers if the conditions are also stated. Bad master data, no rights concept, no department owner? Then you can throw those percentages in the trash.

For data platforms, the development is more pragmatic. Datasolut describes Microsoft Fabric with Copilot as an obvious choice for many Microsoft-affine SMEs because engineering effort and entry barriers are lower. Databricks with Mosaic AI and Agent Bricks is more suitable for companies that see AI as a core competence and want to orchestrate their own agents. This is a strategic switch. Anyone with 140 employees, two IT admins, and an overloaded controlling department should not pretend to be OpenAI with forklifts. But anyone with 480 employees, a data team, and a digital product promise can build more strongly.

German SMEs only have five years left. AI in mechanical engineering is no longer a future topic, but decides who remains marketable.

— Etienne Fieg, Co-Founder of lytra

I would phrase that even more harshly. Anyone in mechanical engineering who doesn't have a productive AI agent running in service, back office, or sales support by 2027 won't disappear immediately. But they will learn slower than the competition. And learning slower in markets with price pressure, a shortage of skilled workers, and a global service promise is almost the same as shrinking. Schaeffler, Festo, Kärcher, Webasto, or Brose have different resources than the typical SME. Nevertheless, they set standards for response times, digital interfaces, and data quality, against which customers will eventually measure smaller providers as well.

A second look: Why many Agentic AI projects fizzle out

I often hear the counter-position from CTOs. “First of all, we don't have our data under control,” Jens, CTO of an electronics supplier from Nuremberg, told me. He's right. But this sentence sometimes becomes an excuse not to make a decision for years. Data quality doesn't improve in a vacuum. It improves when a specific process costs money and an agent brutally shows during the first test which fields are missing, which article master data are duplicated, and which CRM notes only consist of “see email.”

Nevertheless, there are real pitfalls. The first is tech-first. Companies start with LangChain, AutoGen, Copilot Studio, or some agent builder without first building a process map. Then a demo is created that gets applause from the board but helps no one in everyday life. The second is over-automation. A managing director wants end-to-end autonomy directly because the slide then looks better. Wrong. In critical processes, people should be in the approval loop first. After that, standard cases can be released. Step by step. Not out of fear, but because trust is a production factor.

The third pitfall is governance as an afterthought. Especially in DACH SMEs, this is deadly slow. Works council, data protection, and IT security are only invited when the MVP is almost finished. Then comes the question: What personal data does the agent actually process? What logs are stored? Is performance monitored? Is data located in the EU? Who can stop the agent? Suddenly, the meeting room smells of cold pizza and panic. I've seen it multiple times. Not just at Amplifa, but across conversations with digital leaders in Munich, Stuttgart, Hanover, and Linz.

Agentic AI Governance: SMEs need guardrails, not bureaucracy

Many managing directors hear governance and immediately see folders, committees, and consultant fees. Understandable. But bad governance is expensive; good governance is a brake with ABS. It allows speed because it's clear when to stop. I would start every Agentic AI program in SMEs with four governance artifacts: use case register, risk classification according to EU AI Act, agent job description, and approval matrix. That sounds like paper. In reality, it's an operating system.

The EU AI Act is not just about high-risk systems in HR, credit, or safety-critical quality. Even low-risk agents need competence, documentation, and control when they intervene in business processes. A customer service agent at Kärcher who makes response suggestions is different from an HR agent who pre-sorts applications. A quality agent at an automotive supplier near Wolfsburg has different risks than a sales agent who analyzes company websites. Anyone who doesn't cleanly separate these differences over-regulates harmless cases and underestimates dangerous ones.

My advice is blunt: No agents in HR selection, salary decisions, or safety-critical quality as a first project. Period. The first agent belongs where high manual effort, clear data, and limited damage come together. Service triage. Back-office entry. Sales research. Knowledge search with sources. There, the organization learns without immediately throwing itself into the toughest regulations. Anyone who starts with the most difficult process confuses courage with vanity.

Platform or in-house development: The wrong question of pride

I have an allergy to technical questions of pride. “Do we build it ourselves?” is often asked in SMEs before it's clear what should be built at all. Platforms like Microsoft Fabric, Databricks, SAP BTP, or specialized providers like lytra don't solve the same problem. Fabric with Copilot is often the faster entry point for Microsoft-affine organizations, especially if data is already in Power BI, SharePoint, and Azure. Databricks with Mosaic AI and Agent Bricks is stronger if data engineering, ML-Ops, and custom agent logic are strategically relevant. SAP BTP is obvious if core processes are deeply embedded in the SAP ecosystem. lytra is exciting if mechanical engineering service is not just a process, but a business model.

In-house development is worthwhile if the process creates differentiation. Not if you're just extracting an invoice from a PDF. There are enough tools for that. But if an agent combines product knowledge, pricing logic, service history, and customer value into a sales or service proposal, in-house development or at least strong customization can make sense. The difference is IP. At Wittenstein or a similarly knowledge-intensive provider, value lies not only in the product but in the decision of which solution fits which customer problem. There, you don't want to throw everything into a generic box.

For companies with 50 to 500 employees, I usually see a hybrid. Standard platform for data access, security, and authentication. Individual agent logic for processes that influence revenue or customer loyalty. No CTO should occupy their scarce team with rebuilding generic chat interfaces. But no managing director should completely delegate the core logic of their service or sales model to a vendor. That's not a contradiction. That's mature.

How industries use Agentic AI differently

In mechanical engineering, service is the natural entry point. The reason is simple: knowledge is distributed, error patterns repeat, spare parts have margins, downtime costs money. A service agent can directly impact time-to-resolution, first-time-fix rate, and after-sales revenue here. At a plant manufacturer from Baden-Württemberg, whose name is not to be mentioned, the test room smelled of hydraulic oil and warm control cabinets during a workshop. The most exciting moment was not the demo. It was the sentence of a service technician: “If the agent shows me the last three similar cases, I save myself two calls.” That's exactly what productivity is.

In trade and technical wholesale structures, the back office is stronger. Orders come via email, PDF, EDI, remaining faxes (yes, still), and portals. Agents extract items, check customer numbers, identify deviations, and create processes. This is less sexy than an autonomous sales agent. But if 14 people type orders daily, the ROI is obvious. A company like Würth naturally has completely different economies of scale, but the patterns also apply to the 90-person specialist dealer in Kassel.

In B2B sales, Agentic AI is trickier because bad automation immediately smells like spam. Anyone who still believes in 2026 that an agent can simply send out 10,000 cold emails and that's a strategy has forgotten their market understanding in 2018. The sensible role lies before and in between: checking ICP fit, identifying triggers, researching contacts, updating CRM, preparing follow-ups, translating meeting notes into next steps. Markus, sales manager at a component manufacturer from Heilbronn, told me in April 2025: “Our problem is not that we know too few companies. Our problem is that we don't know which ones are currently ready to buy.” Exactly there, agents can help.

Practical example: Service agent with real numbers

Let's take a typical mechanical engineering SME with 260 employees, 42 million euros in revenue, 18 service technicians, 7 people in internal technical support, and customers in DACH, Benelux, and Italy. Not a fictional factory visit, but a condensed pattern from several conversations and project calculations, as we have seen them in 2024 and 2025. 18,000 service requests are received per year. Of these, 55 percent are recurring error patterns, spare parts questions, or operating issues. The average processing time in internal support is 18 minutes per ticket, significantly higher for complex cases. The room for improvement is not theoretical. It sits in the inbox every morning.

The first agent is given a narrow role: read ticket, identify customer and machine, retrieve history, display similar cases, make spare part suggestion, generate response draft with sources, escalate in case of uncertainty. No autonomous price approval. No safety-critical repair instructions without human intervention. After 12 weeks of MVP, 6,000 tickets are historically evaluated, 1,200 tickets are processed in test operation, and 300 are accompanied live with human-in-the-loop. The agent achieves a correct classification of 82 percent for standard cases, suggestions are adopted or slightly adjusted by internal support in 68 percent of cases. Not perfect. But usable.

The calculation then looks like this: If 9,900 standard tickets are affected per year and the average processing time drops from 18 to 10 minutes, that saves 79,200 minutes. That's 1,320 hours. At internal full costs of 62 euros per hour, this results in an efficiency potential of around 81,840 euros per year. In addition, there is service revenue. If the agent enables a suitable additional item or a faster quotation process in only 8 percent of relevant tickets in spare parts cases, and this generates 120,000 euros in additional contribution margin per year, the case is suddenly no longer just cost reduction. Then after-sales becomes a measurable revenue driver. That's why I find pure productivity calculations too small.

The pilot in this example costs 160,000 euros, including data connection, workflow design, model costs, tests, training, and governance. Ongoing costs are 36,000 euros per year. In the conservative case, the project pays for itself after 13 to 18 months. In the better case, in under 12 months. The difference rarely lies in the model price. It lies in data access, departmental usage, and the question of whether service is really allowed to sell or just put out fires.

Amplifa ICP Playbook Use the ICP Playbook to align Agentic AI use cases in sales and service first with the right customer segments.

FAQ: What managing directors need to know about Agentic AI

Is Agentic AI even realistic for 50 to 500 employees?

Yes, but not as a miniature corporate program. A company with 80 employees doesn't need an AI Transformation Tower. It needs a process with high pain, a professional owner, a platform decision, and an MVP that measures something in 8 to 12 weeks. For 50 to 500 employees, the advantage is often even greater because decisions are made faster. The disadvantage: every wrong construction site immediately eats up capacity.

Which Agentic AI use cases deliver ROI first?

In my experience, first service triage, order entry, sales research, knowledge search, and standard reporting. Not because these cases are the coolest. Because they are measurable. Time per process, error rate, throughput time, response time, quotation rate, ticket throughput. If a CFO from Stuttgart asks after four weeks whether the agent is working, you don't need a vision, but a baseline.

How much autonomy can an agent have at the beginning?

Little. First assistive, then semi-autonomous, then autonomous in clearly defined standard cases. I know that sounds cautious. It's not. It's faster because trust doesn't have to be repaired afterwards. An agent who sends three wrong customer emails can damage an entire program. An agent whose suggestions people check and evaluate learns with the organization.

What does a serious entry into Agentic AI cost?

For strategy and use case prioritization, managing directors should budget 20,000 to 50,000 euros. An MVP with one or two agents usually costs between 80,000 and 250,000 euros if real system integration is involved. Anyone who believes they can get productive ERP autonomy with a 9,000-euro prototype from a workshop will be disappointed. Perhaps they will get a demo. But no reliable operational capability.

Do we need a complete data strategy first?

No. But you need enough data strategy for the chosen process. That's a big difference. For a service agent, tickets, machine master data, manuals, and spare parts must be accessible, versioned, and authorized. For a sales agent, you need clean ICP criteria, account data, trigger sources, and CRM rules. Anyone waiting for the perfect data landscape will wait a long time. Anyone starting without data rules builds chaos with a nice interface.

Seven steps to an Agentic AI strategy in SMEs

1. Start with a use case inventory across service, back office, sales, production, and controlling. Collect 20 to 40 candidates, but select only one first process. A workshop in May 2025 with a B2B manufacturer from Ulm showed the same pattern again: the first ten ideas were too broad, the best idea lay in a boring quotation handover.
2. Evaluate each use case according to impact, effort, data access, risk, and process ownership. I would not start a use case where a department head does not explicitly take responsibility. If no one puts their name on it, the case is not politically mature.
3. Write a job description for the agent. Inputs, outputs, allowed actions, forbidden actions, escalation path, KPI, audit log. Treat the agent like a new operational role, not a plugin.
4. Clarify platform and data storage early. Microsoft Fabric, Databricks, SAP BTP, Copilot Studio, lytra, or in-house development are not religions. The choice depends on existing IT, engineering capability, EU data requirements, and the degree of differentiation of the process.
5. Involve data protection, works council, and IT security in week one. Not as brakes. As risk designers. Especially with personal data, performance-related issues, or safety-critical decisions, there should be no surprises.
6. Measure a baseline before the MVP. Processing time, error rate, throughput time, response time, quotation rate, service revenue. Without a baseline, the loudest opinion in the room wins in the end.
7. Plan scaling only after proof of use. An agent ignored by 80 percent of the team is not a scaling candidate. An agent whose suggestions are adopted in 65 to 75 percent of cases deserves a budget.

Amplifa Product Amplifa helps B2B teams productively set up AI-powered sales and account workflows instead of just building demos.

What we specifically see at Amplifa

What we specifically see at Amplifa: The best Agentic AI projects in sales don't start with outreach, but with market logic. In the last 12 months, we have seen a recurring pattern with customers from mechanical engineering, technical services, and B2B software: If the ICP is unclear, AI generates more activity and less truth. As soon as ICP criteria are operationalized – industry, triggers, installed technologies, growth signals, regional priority, exclusion criteria – the number of contacted accounts sometimes drops by 35 to 55 percent, while the appointment rate increases. A customer from the Stuttgart area booked three times as many qualified initial appointments after nine months, without hiring an additional SDR. Not because the emails were suddenly poetic. Because the agent stopped pursuing the wrong companies.

That's the lesson for Agentic AI as a whole. Autonomy without selection criteria is noise. A service agent needs criteria for standard cases and exceptions. A back-office agent needs criteria for secure extraction and manual review. A sales agent needs criteria for fit and timing. Otherwise, you're not automating work, but imprecision. And imprecision scales terribly.

ROI depends on adoption, not model costs

Many ROI calculations for Agentic AI stare at LLM costs. That's the wrong place. Model costs are relevant, but rarely the main lever. Integrations, process clarification, change, quality assurance, and operation are expensive. Non-use is even more expensive. An agent that works technically but is bypassed by clerks has a negative ROI. Then you pay for the platform, consulting, internal capacity, and lose trust. I prefer a lean agent that is used daily to a large architecture that shines in Confluence.

Adoption doesn't happen through training alone. It happens when the agent noticeably takes work away without embarrassing people. In one project, an employee from customer service in Cologne said: “I don't want the AI to write my answer. I want it to get me the three pieces of information I would otherwise have to search for.” This sentence is gold. Many people don't want an automatic voice to the outside world. They want less searching, less copying, less system switching. That should shape the design.

Therefore, employees belong in agent development. Not in a change theater with sticky notes, but concretely: Which cases are annoying? Which exceptions are dangerous? Which formulations would you never send to customers? Which ERP fields are reliable and which lie? Anyone who doesn't ask these questions builds for organizational charts instead of for work.

My forecast for 2026 and 2027

I don't believe that every SME will have an Agentic AI organization in 2026. The term sounds bigger than most businesses need anyway. But I do believe that good SMEs will operate two to five productive agents in clear processes by the end of 2026. Service, back office, sales support, reporting, knowledge search. Not as a laboratory. In operation. With logs, roles, KPIs, and a person who stands behind it.

The losers will not be the companies that buy the latest model too late. The losers will be the companies that do not make their processes decidable. Because Agentic AI forces organizations to do something they have long been able to avoid: explicitly state who can decide what and when. That is uncomfortable. It scratches at departmental boundaries, at old habits, at small shadow processes in Excel. But that's exactly where the profit lies.

My personal point of view is simple: Anyone who treats Agentic AI as an IT project will get an expensive demo. Anyone who treats it as a process and market project can see measurable benefits in 12 months. Not everywhere. Not without friction. But enough to make competitors nervous. And perhaps that is the best early indicator: If the first service technician says he no longer wants to give up the agent, the strategy has arrived in everyday life for the first time.