AI in Sales: SMEs are not sleeping on it

Meinung & Provokation · 12. Oktober 2026 · Anthony Filipiak

AI in sales needs less hype and more ROI. Read why SMEs are not foolish for waiting – but learning is now mandatory for 2026.

43 percent of AI-using companies in Germany cannot quantify AI's contribution to operating profit (McKinsey, Germany analysis of "The State of AI in 2026: On the Road to ROI," published in 2026). This is the number I discuss with managing directors longer than any demo video from OpenAI, Microsoft Copilot, or Salesforce Einstein. Because AI in sales is not valuable when an SDR writes ten prettier emails, but when pipeline, close rate, or deal velocity measurably shift. That's where it gets thin. And that's precisely why the thesis "German SMEs are sleeping on AI – and that's okay" is not wrong enough to dismiss, but also not right enough to rest on.

My bold thesis – SMEs are not sleeping on AI, they are sleeping on the benchmark. Well, almost. They test, pay, click, marvel, integrate Copilot into Outlook, and let vendors tell them that everything will be different now; but when I ask after nine months which process was accelerated, served more cheaply, or escalated earlier as a result, many conference rooms fall silent.

And yes, that's partly okay. Anyone who still believes in 2026 that every SME must immediately unleash agents on ERP, CRM, MES, and PLM has probably never spoken to a CFO from a mechanical engineering company whose SAP system has been operating with three special logic rules for service prices since 2014. Caution is not backwardness. But caution without a learning plan is just another word for loss of control.

AI in Sales – Why Most Are Wrong

Most discussions about AI in SMEs start in the wrong place. They start with the model. GPT-5, Claude, Gemini, Llama, some agent framework with a shiny interface. I consider this a distraction. Not because models don't matter. Not entirely true. Models are important, but they are rarely the bottleneck I see in conversations with managing directors and sales managers.

The bottleneck lies in the company's operating system – ICP unclear, CRM data patchy, reasons for offers not coded, lost deals as free text in the "Other" field, nobody clearly knows which triggers truly lead to purchases, and then AI in sales is suddenly supposed to work wonders. At Phoenix Contact, one can talk about automation because processes have a different level of maturity. For a 180-person supplier from East Westphalia, whose sales manager still works with three Excel files, half a HubSpot license, and Sabine from internal sales' memory, the world looks different.

"That doesn't work for us," Thomas, sales manager of a component manufacturer from Pforzheim, recently told me. "Our customers don't buy because of an email." I hear that sentence often. It sounds experienced. Sometimes it's just convenient. Because, of course, nobody buys a packaging machine from Krones or an axle from Wittenstein because of an email. But a sales conversation doesn't arise from incense either. It arises from timing, relevance, repetition, and a sales apparatus that recognizes when a target account is in motion.

The misunderstanding runs deeper. Many SMEs confuse AI usage with AI capability. Usage means – someone writes a job ad with ChatGPT, internal sales has an email shortened, marketing generates LinkedIn posts that sound like everything else on LinkedIn (i.e., like wet cardboard). Capability means – the company can select a process, formulate a hypothesis, connect data, limit risks, measure results, and then decide whether to scale or bury it. The second is rare. The first is growing fast.

The LinkedIn version of the debate is too cheap for me

On LinkedIn, there are two camps. Camp one shouts – Germany is lost because SMEs are missing out on AI. Camp two shouts – we are cautious because we are world champions of quality. Both make it too easy for themselves. In March 2026, I spoke with Andrea, Head of Sales at a hidden champion in Bielefeld, about AI in sales. Her sentence stuck with me – "We tested six tools and didn't change a single process." That's precisely the point. Not a lack of tools. Process cowardice.

I say this deliberately harshly because I've been hearing the watered-down version for years. First, it was CRM. Then Marketing Automation. Then Intent Data. Now AI agents. Again and again, technology was bought before management had decided which sales truth it actually wanted to endure. If the best customer is no longer the loudest customer, but the most profitable customer after service costs, then sales must prioritize differently. If a segment brings many leads but every second order ends in a complaint, marketing must target differently. AI exposes such contradictions. Some don't want that at all.

AI in Sales – The Uncomfortable Truth Behind the Numbers

The AI Index for SMEs 2026 by the German Association for Small and Medium-sized Businesses (DMB) and Salesforce states that 51.2 percent of SMEs used or tested AI in 2025. In 2024, it was 33.1 percent. That's a jump, not a shuffle. 16.6 percent already used AI agents, 37 percent planned implementation or expansion for 2026, and 31 percent had no concrete AI plans. Anyone who only concludes "SMEs are sleeping" is reading carelessly.

But anyone who concludes "everything's fine" is reading dangerously. Because the McKinsey figure of 43 percent unquantifiable results contribution hits precisely the sore spot. A company can use AI in four business areas, according to McKinsey it was an average of 4.3 in Germany compared to 3.5 globally, and still not know if operating profit is increasing. That's like putting a new DMG Mori machine in the hall and after a year not knowing if unit costs, scrap, or delivery time have improved. No managing director would accept that. With software, it suddenly becomes romanticized.

The Qonto figure is particularly revealing. 34.46 Euros average monthly AI expenditure. That doesn't get you a sales process overhaul. That gets you a few single licenses, a credit card booking, and the good feeling of not being completely left behind. The smell of transformation? More like the smell of a freshly printed tool receipt from accounting.

I don't want to make these platform data bigger than they are. Qonto doesn't see the complete German corporate landscape. McKinsey surveyed 93 people in Germany, embedded in a global survey of 1,719 executives and professionals from 97 countries. Salesforce and Alibaba have their own market interests. Granted. But if several imperfect windows point in the same direction, one should at least look out.

We are not afraid of AI. We are afraid that someone will connect an agent to a bad master data system and then everyone will pretend the error was surprising.

— Matthias, COO of a mechanical engineering supplier, Stuttgart

Matthias is right. I know companies where a wrong spare part suggestion doesn't just trigger an embarrassing email, but causes downtime for the customer. I know sales organizations where an automated discount suggestion immediately destroys margin because historical deals were never adjusted for special effects. At Festo, Schaeffler, or Siemens, AI is discussed with different safeguards than at a 70-person company that maintains its offer templates in Word and network drives. That's precisely why the blanket 'sleeping on it' narrative is too crude.

But what if waiting is actually smart?

The strongest counter-argument to my harshness is – AI is not an end in itself. And honestly? I agree. If a managing director in April 2026 has to choose between energy prices, financing, skilled labor shortages, cybersecurity insurance, and an ERP upgrade, then an AI pilot in sales might not be priority one. Maybe not even priority five. A family business in metal processing in Tuttlingen will not improve by chasing every trend faster than its own customers pay.

Caution is often rational in an industrial context. A wrong quality release costs more than a bad social media post. A wrong maintenance instruction can stop a plant. An inaccurate export control statement can cause trouble that doesn't disappear in a pretty SaaS dashboard. And an AI agent that creates orders in SAP S/4HANA requires different governance than a text assistant for newsletters. Those who are slow here are not automatically backward. Sometimes they are mature.

However, mature caution must not turn into comfortable rigidity. That is the fine line. When a managing director says – "We'll wait until the tools are more mature," I ask – what data are you preparing until then? Which three processes do you want to measure first? Who in sales owns the hypotheses? Silence. Then it's not waiting. Then it's delegation to the future.

What I see in practice

What we at Amplifa specifically see: In the last 12 months, with B2B clients in mechanical engineering, technical services, and industrial software, we have observed that AI projects in sales only become reliable when they are linked to a commercial event – new market, shrinking inbound pipeline, post-trade show follow-up, dealer activation, or reactivation of lost target accounts. In 17 analyzed go-to-market setups, the biggest leverage was not in better texts, but in the combination of a tighter ICP, trigger detection, and consistent follow-up over 6 to 10 weeks. Teams that only wanted "more personalized emails" rarely got more than 10 to 18 percent better response rates. Teams that re-sorted target accounts, timing, and offer logic doubled qualified first appointments in several cases – without a new sales employee.

That sounds less sexy than "autonomous agent." But it's money. Markus, managing director of an automation service provider from Augsburg, told me in June 2026 – "I thought AI would write us better initial outreach. In the end, it showed us that we've been prioritizing the wrong companies for years." His team had previously worked on industry lists, clean, diligent, German. After ICP sharpening, target accounts were evaluated based on machinery, expansion signals, open service roles, investment announcements, and technical compatibility. The result after nine months – three times as many appointments with suitable accounts, same headcount, less discussion about lead quality between marketing and sales.

The point is not that Amplifa presses magic buttons. They don't exist. The point is that AI in sales works when it is unleashed on a messy, concrete sales question. Which companies look like our best customers before they buy? Which existing customers have expansion signals but no active account plan? Which lost opportunities were timing issues and not real noes? Which dealers bring revenue but no market data back? Such questions smell of work. Of CRM export, of sales meetings, of a whiteboard where someone writes "Why do we win?" with a dry marker.

Amplifa ICP Playbook A practical playbook to define target customers not by gut feeling, but by market, trigger, and revenue logic.

I'm not linking the ICP Playbook because a PDF will save the world. It won't. I'm linking it because most AI debates in sales without a clean ICP end up like a trade show without a follow-up list – expensive, loud, later embarrassing. If you don't know who you want to win, you're just automating your own imprecision.

AI Agents are Where Things Get Serious

The YouGov survey for Alibaba.com and bevh reported in 2026 that almost two-thirds of German SMEs use generative AI at least occasionally. However, only 24 percent used autonomously operating AI agents; for companies with ten or more employees, it was around 45 percent, with 21 percent using them regularly. This is the break. A chatbot writes. An agent acts. Between them lies an abyss of rights, data, liability, process design, and trust.

In sales, this abyss is often underestimated. An agent that merely summarizes website texts is harmless. An agent that evaluates target accounts, fills CRM fields, starts sequences, prioritizes contacts, or prepares offer drafts changes the division of labor. Who approves? Who checks? Who is liable if a key account is contacted twice because the duplicate logic is abysmal? Who stops the agent if the response rate is high, but only for accounts that will never buy?

I see two types of SMEs here. Some buy agents like they used to buy newsletter tools. Implement, train, done. Others build a small control architecture – clear use cases, sandbox, rights concept, KPIs, manual approval, audit log, kill switch. Guess who's still relaxed after six months. At a client from the technical wholesale sector, the server room in the old administration building still smelled of warm dust, but the agent governance was cleaner than at some software companies in Berlin-Mitte.

1. Choose a process that affects revenue or margin – not the most convenient demo case. Examples include target customer prioritization, reactivation, offer preparation, or dealer management.
2. Define a hard metric before the pilot – qualified appointments, offer rate, lead time, forecast accuracy, or contribution margin per account.
3. Separate assistance from autonomy – a text suggestion is not the same as a CRM update or a triggered sequence.
4. Build a control group – otherwise, you'll celebrate seasonality, post-trade show effects, or luck as AI success.
5. Set stop criteria – if no signal comes after eight weeks, don't tinker further, but make a decision.
6. Document learning curves – what data is missing, which fields are garbage, which approvals block, which step should not be automated.

Why Pure Inbound Strategies with AI in Sales Become Dangerous

Now it gets uncomfortable for many managing directors who have relied on inbound in recent years. Anyone who still relies on a pure inbound strategy in 2026 will have no pipeline left in five years. Yes, I know that sounds exaggerated. But look at B2B markets with long cycles, technical products, and shrinking search volumes. If everyone produces content, everyone does webinars, everyone hides whitepapers behind forms, and everyone scales texts with the same AI tools, visibility is produced more cheaply and acquired more expensively.

For a manufacturer of testing technology from North Rhine-Westphalia, organic lead inflow was still stable in the first half of 2026. On paper. In the pipeline, it looked different. More students, more competitors, more unclear consulting inquiries, fewer purchase-ready projects. Sales heard the short ping from the CRM in the office and knew after two weeks which leads would be time-consuming. AI in sales didn't help to create more content there. It helped to connect existing data with target market information and proactively approach accounts before they actively searched.

This is the market disruption that is too little discussed. AI makes content cheap. This weakens content as a differentiator. AI makes research scalable. This makes outbound more precise again, if it's not run as spam. AI makes CRM data more evaluable. This makes lazy excuses in the forecast more visible. Anyone who ignores this shift is not sleeping on AI. They are sleeping on the new sales economy.

Amplifa Pipeline Audit We check where pipeline originates, where it leaks, and which AI levers in sales can actually become measurable.

German SMEs are not slow. They are often imprecise.

Slowness is not my main accusation. Imprecision is. I meet SMEs that make decisions astonishingly quickly compared to corporations. A managing director, two department heads, a Friday meeting, Monday implementation. No steering committee with 19 people and three consulting logos on the slide. But these same companies can remain imprecise for years when it comes to target customers, reasons for loss, margin logic, and segmentation. That's the contradiction.

Trumpf, Kärcher, Brose, Webasto, Schaeffler – these names stand for industrial substance, not for trivialities. Many hidden champions are built similarly, just smaller and quieter. Precisely these companies have an advantage that startups often lack – real process data, genuine customer relationships, deep domain knowledge, installed base, service history. If they combine this substance with AI, no LinkedIn hype is created. A lead is created that is difficult to copy.

But if they wait until SAP, Microsoft, Salesforce, or Siemens deliver the perfect button, they outsource process intelligence. Then the value no longer lies in their own sales system, but in the platform logic of a provider. I am not a sovereignty romantic. I use US software every day. But I find it naive when companies blindly dump their customer data, decision rules, and sales logic into foreign systems and then believe they have control because the contract is in German.

What Must Happen Now

SMEs don't have to understand every foundation model. They don't. A managing director of a tool manufacturer in Villingen-Schwenningen doesn't have to be able to explain why a transformer uses attention. But he must know which five processes determine margin and growth. And he must demand that AI pilots are measured by these processes, not by the number of texts generated.

My concrete appeal to managing directors and sales managers is simple, but not comfortable – do fewer AI projects and sharper ones. One pilot for target customer prioritization. One for offer acceleration. One for account reactivation. Not twelve experiments in the fog. Not a Copilot rollout as a substitute for strategy. Not an agent that automates operational errors because nobody touched the process beforehand.

1. By the end of November 2026, write down your ten most important revenue processes – from lead source to payment receipt, not just the marketing funnel.
2. Mark three processes with high repetition and high awareness of error costs – that's where AI is more worthwhile than for prestige topics.
3. Define an economic hypothesis for each process – for example, 20 percent shorter offer time or 15 percent more qualified appointments in a segment.
4. Check data access and data quality – CRM, ERP, service, website, trade fair contacts, dealer reports.
5. Start with assistance before allowing autonomy – especially for prices, offers, compliance, and customer communication.
6. Compare against a control group – otherwise, you'll believe your own slide.
7. Scale only if a real process gain is visible – response rates alone are not a business model.

Amplifa AI Sales Sprint A focused sprint for B2B teams that want to align AI in sales not just for testing, but with pipeline, ICP, and measurable sales processes.

FAQ – Is the Mittelstand really sleeping on AI?

Is AI in sales already mandatory for SMEs?

Mandatory is the wrong word. Learning obligation fits better. No SME has to put an autonomous agent on its customer communication tomorrow. But every company with complex B2B sales should know in 2026 where AI changes research, segmentation, follow-up, forecasting, or quoting processes. Those who don't learn at all will pay later with frantic efforts.

Why is ChatGPT not enough in sales?

ChatGPT can provide texts, analyses, and structure. That helps. But sales success doesn't come from text alone. It comes from target account logic, timing, data quality, clean CRM, suitable offers, and discipline in follow-up. A good prompt doesn't cure a bad market fit. A sales manager from Hanover told me in August 2026 – "Our emails got better, but our pipeline didn't." That's where tool euphoria ends.

Are AI agents too risky for mechanical engineering and industry?

They are risky if introduced without a rights concept, approval process, and measurement logic. They are useful if they start with clear limitations – for example, in research, document search, account summaries, or offer preparation. For incorrect CNC parameters, quality releases, export information, or pricing decisions, human control is needed. Period.

How do you measure ROI for AI in sales?

Not by usage hours. Not by the number of emails generated. I would start with four key metrics – qualified first appointments per target segment, offer rate, sales cycle, and contribution margin per won account. Depending on the business, forecast accuracy, reactivation rate, or offer lead time can be added. The baseline before the pilot is important. Without a baseline, every improvement becomes a narrative.

The Provocative Thesis Remains – But Differently

"German SMEs are sleeping on AI – and that's okay" is too comfortable for me. The sentence hits a nerve, but not the diagnosis. SMEs are not sleeping on the demo. They often miss the translation into measurable process benefits. And yes, sometimes it's okay not to be the first. Anyone who doesn't have clean data, process ownership, and a business case shouldn't just roll out because a vendor in Munich has rented a stage and writes "Agentic AI" on an LED wall.

But waiting becomes wrong as soon as it happens without preparation. If 51.2 percent are testing or using and 31 percent still have no plan, a calm market doesn't emerge. A learning gap emerges. Some companies are currently building process knowledge. Others are collecting tool logins. Still others do nothing and call it risk management. The difference won't be immediately visible. It will show up later, when a competitor writes offers faster, manages dealers more targeted, systematically retrieves lost accounts, and knows their best target customers before they fill out a form.

I don't believe in AI as a panacea. I believe in sales work with better instruments. That's less loud, but harder. The managing directors who impress me don't ask "Which AI do we need?". They ask – "Where are we losing money, even though we should know it?" After that, it gets interesting.

Perhaps that's the real provocation for 2026 – not that SMEs are sleeping on AI. But that many AI providers are sleeping on SMEs because they sell demos where process responsibility is needed. And in some meeting rooms, you can smell that faster than you can see it on a slide.