AI in Manufacturing: Almetra and the Shop Floor

KI & Automatisierung · 1. Juli 2026 · Ohiku Mose Guy

AI in manufacturing is becoming operational. Read what Almetra's €16M funding round means for DACH mid-sized companies and how to get started.

Anyone who still believes in 2026 that AI in manufacturing is an IT project will be overtaken by their own shop floor. I mean this seriously, because the next productivity gains won't come from prettier dashboards, but from decisions made at the pace of the line. Not monthly. Not after the next Lean workshop. But while a plant is running, material fluctuates, an operator improvises, and sales still promise delivery dates. Almetra's €16 million Series A is therefore not just a startup announcement from Berlin – it's a signal that Manufacturing Intelligence is breaking out of the pilot box and wants to enter everyday production.

Almetra, formerly Deltia, raised a Series A of €16 million in January 2026, led by blisce/ from New York and Paris, with NAP, Merantix Capital, Robin Capital, Underline Ventures, and Critical Ventures as additional investors [1]. For me, this is not exciting because of the sum. €16 million is not a fireworks display in the VC calendar. What's exciting is where the money is being used – on the shop floor, where many DACH mid-sized companies are still grappling with Excel exports, shift logs, MES gaps, and gut feelings. Kärcher, Trumpf, Phoenix Contact, Festo, Schaeffler – the big names have been showing for years how much structure can be found in production data. But the real question is: What happens at a tool manufacturer with 180 people in East Westphalia, at a plastics injection molder in Vorarlberg, or at a precision parts manufacturer near Winterthur?

Status Quo – AI in Manufacturing is More Advanced Than Many COOs Believe

The status quo is contradictory. In strategy slides, the Smart Factory has long been a reality. On the factory floor, it smells of coolant, an old label printer rattles next to the line, and the most important OEE deviation is handwritten on a clipboard. Both are true. According to the AlixPartners Swiss Disruption Landscape in 2026, technological disruption – including automation and AI in production processes – already affects 62 percent of Swiss manufacturing companies [4]. Regulations and ESG pressure even affect 68 percent [4]. This is not a side note for sustainability officers. This lands with the COO, the managing director, and eventually the sales manager when delivery promises no longer match actual capacity.

I often see the same curve in DACH mid-sized companies. First, there's talk of a shortage of skilled workers. Then about energy prices. Then about scrap. Then someone says: We should actually know which line is losing money. That's where the conversation shifts. Because many factories have machines with sensors, a PLC, sometimes an MES, sometimes an ERP with reasonably clean orders – but they don't have a reliable common truth about losses. A stamping line reports downtimes differently than assembly. Rework lives in a separate quality form. The reasons for micro-stops are estimated by the shift, if anyone even has time. Well, almost. In well-managed factories, there are, of course, key figures. But key figures are not the same as causes.

This makes Almetra's positioning interesting. Maximilian Fischer, Co-Founder and CEO of Almetra, puts it bluntly in the announcement: Factories produce everything around us, but often operate blindly [1]. This sentence resonates because it doesn't sound like AI. It sounds like production management at 6:10 AM, when the early shift report is an hour old and still doesn't explain why order 4711 is stuck on machine 4. Manufacturing Intelligence doesn't promise magic here. It promises a brutally practical thing – making losses visible, assigning them, prioritizing them.

For sales managers, this is not purely an operations story. If a factory doesn't know its real usable capacity, it either sells too cautiously or too boldly. Both cost money. Too cautiously means: competitors like DMG Mori or a Polish contract manufacturer with better planning take volume. Too boldly means: sales win the order, production loses the margin, the customer loses trust. I have seen in projects that a single misunderstood bottleneck distorts the entire quotation logic. The calculator works with standard times. The line lives in a different reality.

Trend 1 – Manufacturing Intelligence Becomes an Operational Layer Above MES and ERP

The first trend: AI-native Manufacturing Intelligence platforms do not replace ERP, MES, or SCADA. They lie on top of them. This sounds technically unspectacular, but it's why this category is now scaling. A mid-sized manufacturer with 50 to 500 employees won't just rebuild SAP S/4HANA, ProAlpha, Infor, an old BDE system, and three machine islands. Anyone who demands that loses the project before the first workshop. Almetra essentially says: We take existing production data, identify bottlenecks, quantify performance losses, and enable better decisions with minimal disruption to ongoing operations [1]. Exactly this connectivity is decisive.

In practice, this means: The platform must tolerate dirty data. Different cycle times. Missing downtime reasons. Operator comments with typos. Machines that speak via OPC UA and machines that are only accessible via a CSV file once per shift. Anyone who only paints Manufacturing Intelligence as a clean cloud architecture has not been in factories often enough. At an automotive supplier in Franconia, Thomas, Head of Industrial Engineering from Nuremberg, told me: “The plant is not old, but the data chain stopped in 2008.” That's exactly where the market is. Not at the lighthouse factories of Siemens Amberg, but at the 10,000 factories in between.

What we at Amplifa specifically see: In the last 12 months, we have observed a recurring pattern among customers in mechanical engineering, plastics processing, and technical B2B manufacturing – between 18 and 32 percent of sales-relevant capacity assumptions in CRM and quotation processes were not aligned with current shop floor data. This does not mean that a sales manager invents false figures. It means that the figures age. A line that still had 82 percent availability in March 2025 may only be running at 74 percent in November because a tool jams more often, because two experienced operators have changed, or because a new material mix shifts the cycle time. Sales only notice it when delivery dates slip.

The technical layer above MES and ERP is therefore not just an IT pattern. It becomes a layer of trust. If a COO sees that Line A loses 14 hours of capacity per week due to the same micro-stops, they can prioritize. If sales see that product family B is running on a bottleneck, they cannot blindly give discounts just to buy utilization. And if management sees that an investment in maintenance brings more than the next machine, Capex is discussed differently. Is that boring? Not quite. It's the difference between margin defense and growth on sand.

We give them certainty instead of guesswork. Most of our customers find significant optimization opportunities in the first few weeks – and that's exactly the speed the industry needs now.

— Maximilian Fischer, Co-Founder & CEO of Almetra, Berlin [1]

I like the part about the first few weeks in this quote. Not because I believe every provider who promises quick results. Honestly? I get suspicious of such statements at first. But in manufacturing, there is indeed a category of problems that become visible after short data integration: incorrect downtime reasons, underestimated setup losses, quality drift on a specific shift, overload on a secondary plant that never appears as a bottleneck in production planning. These things are not hidden because no one is looking. They are hidden because they are distributed across systems.

Trend 2 – Predictive Maintenance Leaves the Slide and Lands in the Weekly Schedule

The second trend is Predictive Maintenance, but please without trade show fog. Predictive maintenance was a promise for ten years with too many sensors and too little responsibility. Now the economics are changing. Energy is expensive, spare parts are not always available the next day, experienced maintenance technicians are retiring, and unplanned downtimes don't just affect production. They affect order acceptance, customer loyalty, and price discipline. At Webasto or Brose, the maintenance organization is large enough to run its own data programs. At a 220-person supplier near Heilbronn, it's different. There, a maintenance manager often decides between putting out fires and root cause analysis.

Manufacturing Intelligence makes Predictive Maintenance valuable when it doesn't just say: Bearing 3 sounds funny. It has to say: If this pattern continues, Line 2 will probably lose 9 to 12 hours of productive time in the next 14 days, specifically on orders for customer X and Y. Only then can maintenance be prioritized economically. Just marking anomalies is not enough. A factory needs a ranking by cost, risk, and delivery impact. In a conversation in December 2025, Jan, COO of a precision manufacturer from Pforzheim, told me: “We have enough sensors. What we lack is a decision that fits into the plan on Monday morning.” It's hard to put it better.

For sales managers, Predictive Maintenance indirectly becomes a pipeline question. If a large order with an eight-week delivery time can only be manufactured on a key plant, the technical condition of that plant is part of the deal qualification. Sounds exaggerated? Then ask a key account manager at an automotive supplier who promised an OEM a series launch and three days later has to explain an unplanned spindle damage. Sales talks about availability, but the machine decides. That's why shop floor data will increasingly flow into quotation and account strategies in the coming years. Not as a pretty export. As a risk indicator.

Here, ESG suddenly becomes concrete. Not in a PDF for the website, but in consumption per order, per line, per scrap batch. If an AI-powered platform recognizes that a certain material batch on plant B leads to more scrap and at the same time consumes more energy per good part, that's not a sustainability slogan. That's margin. That's delivery capability. That's a conversation with purchasing, production, and sales in the same room. At Kärcher or Phoenix Contact, such data chains are strategic programs. In mid-sized companies, they are often held together by a single person whom everyone asks because they have known for 17 years which machine is acting up. This person is valuable. But they are not a scalable system.

Trend 3 – AI Quality Control Will Shift from the Test Station to the Sales Process

The third trend is underestimated: AI-based quality control not only changes scrap. It changes what sales can credibly promise. Computer vision, acoustic inspection, process data analysis, and digital test plans will converge more closely. Almetra mentions product development as a use of capital in the announcement [1]; obvious modules include computer-vision-based quality control, predictive maintenance models, and digital twin analyses at the line level. Whether exactly these modules come in what order is known only to Almetra and its customers. But the direction is clear: quality data will be earlier, denser, and more operational.

Why is this a sales topic? Because quality is no longer negotiated downstream in many industries. Medical technology, automotive, electronics, mechanical engineering – customers want proof. Schaeffler, Bosch, Trumpf, Wittenstein, Festo: Anyone who supplies such companies knows audits, initial sample inspections, 8D reports, and the silent threat of being removed from the supplier list after the next error. A mid-sized manufacturer can not only find errors earlier with AI quality control. They can prove that process windows were stable. That's a different sales conversation. Less “trust us,” more “here's the trend of critical parameters over the batch.”

I am still cautious about AI quality control. Many projects fail not because of the model, but because of the definition of errors. What is a scratch? Which surface deviation is relevant? Who decides in borderline cases? A vision model can only learn what has been professionally and cleanly marked. At a plastic parts manufacturer in Lower Austria, the test cell smelled of warm granulate, while the quality manager showed me 40 parts where two customers evaluated the same surface differently. AI romanticism doesn't help there. What helps is a process that connects customer requirements, inspection strategy, and production data. Only then does AI become useful.

This table is intentionally not built as an oracle. Analyst forecasts are useful, but production reality is nasty. A forecast doesn't tell you if your Line 7 loses 23 minutes every Thursday after a tool change because the scanner misreads labels. But that's exactly the difference between a trend and profit. Manufacturing Intelligence must build the bridge – from macroeconomic pressure to a concrete measure in the factory. If this bridge is missing, AI in manufacturing remains a nice budget label.

Amplifa ICP Playbook For sales leaders in manufacturing: Sharpen your Ideal Customer Profile with production reality, capacity patterns, and reliable buying signals.

Why Almetra's Series A Matters for Managing Directors in DACH

One can read the Almetra round as startup news. Then you nod, remember Berlin, Merantix Capital, blisce/, and €16 million, and move on. That would be a mistake. For managing directors of mid-sized manufacturing companies, this funding is a market signal: a software category is emerging that aims to operationalize production knowledge faster than classic ERP and MES projects could. Not replace. Complement. And sometimes expose.

Why expose? Because many organizations manage with averages. Average utilization, average delivery time, average scrap. Averages are convenient and dangerous. If a product family brings 18 percent margin but only runs on an unstable plant, the average is a trap. If a customer brings a lot of revenue but their call-offs tear up the bottleneck line, revenue is not automatically good revenue. If a factory is green in the monthly report but two shifts regularly have to improvise, green is just a color. Manufacturing Intelligence makes these contradictions visible.

I find it particularly relevant that Almetra wants to expand into Europe [1]. DACH is not an easy market. Works councils, data protection, grown system landscapes, machine parks with a 30-year span, high quality requirements, tough investment committees. Anyone who succeeds here has built something that can withstand friction. A US tool with perfect cloud onboarding can fail due to a single missing machine approval. A European provider who anticipates fragmentation has an advantage. Not a guarantee. But an advantage.

What AI in Manufacturing Means for Mid-sized Companies

For mid-sized companies, AI in manufacturing first means: fewer excuses for flying blind. I say this deliberately harshly. Many 50- to 500-person companies are technically better than they appear organizationally. They have good machines, experienced people, stable customers, sometimes even very clean processes. But they don't consistently measure losses enough, and they don't connect these losses with business decisions. That's the break. A COO looks at OEE. A sales manager looks at forecast and order intake. Management looks at EBITDA and cash. The truth lies somewhere between setup time, complaints, and delivery date.

The first business effect is capacity clarity. No longer: We are roughly full. But: We actually have 11 percent hidden capacity on Line 3 if we eliminate two downtime causes and adjust the product mix differently. Markus, sales manager of a special machine supplier from Augsburg, put it this way in April 2025: “If I knew which orders really fit, I would sell differently.” That's exactly the point. Sales without production intelligence is a gamble in many manufacturing companies.

The second effect is price discipline. If you know that an order consumes a bottleneck plant, you don't give a discount just because the customer gets loud. If you know that a product variant generates above-average scrap, you calculate differently. If you can prove that your quality data is more stable than that of the competition, you don't just sell parts, but process reliability. Premium positioning, which AlixPartners describes as a strategy against disruption for Swiss manufacturers [4], requires exactly such evidence. Not brochures. Data.

The third effect is internationalization with less gut feeling. Many DACH manufacturers are considering nearshoring, additional plants in Eastern Europe, or supply chains with more redundancy. Without reliable production data, this becomes expensive. You then don't relocate processes, but assumptions. A Manufacturing Intelligence layer can help make lines, product families, and locations comparable. At Phoenix Contact or Festo, site comparison is an established control topic. In smaller companies, it often only emerges when the second site is already running and no one can explain why the same assembly causes 9 percent more rework in the Czech Republic.

The Technical Reality – Why Many AI Manufacturing Projects Fail

I'm writing as an engineer, so the unpleasant part has to be included. Projects don't just break because people resist change. They break because data models are naive. Because a machine signal is not unambiguous. Because an 18-second stop is sometimes ignored, sometimes counted. Because an order is finished in the ERP while rework is still ongoing. Because the camera sees different light in summer than in January. Because network segments are separated for good reason. Because security says: No cloud connection from the OT. And sometimes, because the most important operator doesn't feel like typing the same downtime reason into a mask for the third time.

Therefore, the provider with the best machine learning model doesn't automatically win. The provider who can stabilize the data path wins. From the machine to the edge gateway, from the edge gateway to the context model, from the context model to the decision. Context is the expensive word. A power peak means little if I don't know which tool, which material, which order, which shift, and which operating mode were active. A quality error means little if I can't link it to process parameters. This is where Manufacturing Intelligence separates itself from reporting.

At Amplifa, we don't build Almetra's product, and I won't evaluate a foreign architecture here that I haven't seen in the code. But I know this class of systems well enough: The hard part is not the demo with historical data. The hard part is production. New item number. Changed inspection characteristic. Machine downtime during maintenance. Employee change. ERP update. VPN certificate expired. An AI that delivers good insights in week three must still deliver usable signals in month nine. Otherwise, it becomes another screen that no one opens.

FAQ – Is Manufacturing Intelligence Just a New MES?

No. An MES plans, records, and controls production processes, very deeply depending on its maturity level. Manufacturing Intelligence typically sits across it and tries to derive patterns from machine, quality, order, and maintenance data. The boundary is not always clean. Some MES providers integrate AI modules, some Manufacturing Intelligence providers take over MES-related functions. For a managing director, the better question is: Which system will tell me in two weeks where we are losing money and which measure should be taken first?

FAQ – Do We Need Perfect Data for This?

No. But you need honest data. That's a difference. Perfect data is rare in a factory. Honest data means: known gaps, documented assumptions, clear definitions for downtime, scrap, rework, and good parts. If a provider pretends that AI will automatically fix all of this, I would leave the room. Or at least remain silent for a very long time.

FAQ – What Does Sales Specifically Get Out of It?

Sales gets a better answer to four questions: Which products can we grow profitably? Which customers block bottleneck capacity? Which delivery promises are realistic? Where can we sell quality as a differentiator? That sounds like operations. And it is. But in manufacturing, deal quality and production reality are more closely linked than many CRM processes admit.

Preparation – 7 Steps Before You Buy AI in Manufacturing

1. Name an economic bottleneck, not a technology. Example: Line 2 loses an estimated 10 hours per week, but no one knows the main cause. If you start with “we want AI,” you get slides. If you start with loss, you get measurability.
2. Clarify your data sources. ERP, MES, BDE, SCADA, quality database, maintenance tickets, Excel shift reports – write down what exists, who owns it, and how often it is updated. At a Festo supplier in Baden-Württemberg, a team found three different definitions for scrap in June 2025.
3. Define key figures at the plant level. OEE, scrap rate, setup time, and energy per good part are only useful if everyone accepts the same calculation. Otherwise, you'll later discuss mathematics, not measures.
4. Involve sales and controlling early. Manufacturing Intelligence becomes weak if it only shows technical losses. It must show which losses affect revenue, margin, delivery capability, or complaint risk.
5. Start with one line or product family. Not with the entire plant. Choose an area with high volume, visible pain, and responsible people who want to use results. A quiet pilot line without management pressure is a graveyard for good ideas.
6. Plan OT security before vendor onboarding. Network access, edge devices, cloud approvals, role models, audit logs – this is not paperwork. A single unresolved firewall issue can block a project for four weeks.
7. Determine who decides after the insight. If the AI shows bottleneck A, who changes the shift plan, the maintenance plan, the quotation calculation, or the product mix? Without decision-making power, Manufacturing Intelligence becomes diagnostics without therapy.

Amplifa Product Amplifa connects B2B sales processes with data-driven prioritization – so that pipeline, ICP, and operational reality don't diverge.

These seven steps sound down-to-earth. That's exactly why they work. I've seen too many projects that started with model architecture and died with responsibilities. Who should react if a system shows that a premium customer regularly triggers loss orders? Sales? Production? Management? Who tells the customer that their special request is no longer free? AI finds the conflict. It doesn't solve it automatically.

The Sales Angle – Why Pipeline Management Becomes Thin Without Shop Floor Data

Many readers probably expect OEE, maintenance, and quality in an article about Almetra. Fair. But I want to emphasize the sales angle more strongly, because it is too rarely discussed cleanly in DACH. Pipeline management in manufacturing companies is often decoupled from the factory. The CRM knows opportunities, probabilities, customer segments, perhaps contribution margins. The factory knows bottlenecks, setup logic, quality risks, personnel restrictions. Between these two worlds usually sits an Excel sheet or an experienced production planner who, under pressure, says “it'll work.”

Anyone who still relies on a pure inbound strategy in the B2B manufacturing environment in 2026 will have no pipeline in five years. Yes, blunt. But look at the reality: procurement processes are getting longer, technical requirements tighter, supplier evaluations more data-intensive. If your sales team doesn't know which customers fit the real production strength, they are selling against their own operations. An ICP for manufacturing companies must not only contain industry, revenue size, and region. It must contain production fit. Which parts run stably? Which variants have a learning curve? Which customer requirements match the inspection strategy and capacity?

From our implementations, we know: When sales teams align their target customer lists with operational constraints, the top accounts surprisingly often change. At a DACH manufacturer of technical assemblies, 27 percent of the prioritized target accounts shifted after we included product families, delivery time risk, and historical complaint patterns in the ICP evaluation. Previously, a large customer looked attractive because of revenue potential and logo. Afterwards, it was clear: the customer would have pulled exactly the variants that block the bottleneck inspection. This is not theory. This is pipeline hygiene.

Amplifa ICP Playbook for Manufacturers Use the playbook to evaluate target customers not only by market potential, but by production fit, margin, and delivery capability.

What Almetra Addresses Correctly – and Where the Market Is Still Open

Almetra addresses a sore point: factories know they are losing capacity, but not precisely enough where and why [1]. This phrasing is strong because it doesn't sell science fiction. It sells orientation. If customers, according to Fischer, find significant optimization opportunities in the first few weeks [1], then the value probably lies in quickly contextualizing losses. Not in the perfect digital twin from day one.

The market remains open nonetheless. Epicor Prism was also positioned as an AI-powered Manufacturing Solution for Europe, according to The Retail Data [2]. Classic ERP and MES providers will retrofit AI layers. Cloud hyperscalers will provide reference architectures. Specialists for computer vision, maintenance, and energy optimization will go deeper into niches. For customers, this is both good and annoying. Good, because choice emerges. Annoying, because every tool claims to be the central intelligence layer. My advice: Don't believe any architectural diagram that draws your existing systems too cleanly.

The winners will not only be able to do AI. They will be able to implement. They will be able to talk to a factory manager who has no time for platform poetry. They will be able to talk to IT security without appearing offended. They will build a loss statement with controlling. They will explain to sales why a new deal is operationally risky. And they will accept that some data will remain manual for now. Well, almost. Manual often remains longer than providers admit.

Risks – Data Protection, Works Council, OT Security, and Model Trust

No trend report would be honest if it downplayed the risks. In DACH, production data is sensitive. Not only because of data protection, but because of competitive knowledge. Much can be derived about cost structure and customers from cycle times, scrap, order mix, and downtimes. If a provider processes production data in the cloud, it must be clear where data is located, who has access, how models are trained, and whether customer data ends up in general training processes. Works councils rightly ask questions when shift data is evaluated at the operator level. The smell of hot metal in the hall is romantic. The labor law assessment of performance data is not.

OT security is the second hard point. A Manufacturing Intelligence platform needs data from production, but it must not endanger production. Segmentation, read-only access, edge processing, certificate management, patch processes – this is the engine room, not marketing. A COO should ask their provider: What happens if the connection is lost? Can lines continue to run? How are updates tested? What logs are available? What data leaves the factory? If the answers become vague, that's a warning sign.

Model trust is the third point. An AI can identify a bottleneck and still be wrong if the context is missing. Perhaps the line was slow because a new employee was being trained. Perhaps scrap was intentional because a testing process was tightened. Perhaps the order was a special case. Therefore, systems need feedback loops. Operators, maintenance technicians, quality managers, and planners must be able to make corrections. Otherwise, a machine is created that sounds smart and prioritizes stupidly.

Budget Logic – Why AI in Manufacturing Must Calculate Differently

AI in manufacturing will not scale through innovation budgets. Not permanently. It must pay for itself through hard value drivers: less downtime, less scrap, shorter setup times, more stable delivery dates, better quotation margins, lower energy consumption per good part. I would start every project with a baseline. Four weeks of data, one line, clear loss categories. Then measures. Then comparison. Not perfect, but reliable enough to justify a second plant.

An example: A CNC cell with four machines loses 12 hours per week due to unplanned stops and minor disruptions. The internal hourly rate is €95, and the bottleneck additionally prevents two customer orders per month with a contribution margin of €18,000 each. If a Manufacturing Intelligence layer reduces only a third of these losses, the business case is not subtle. But for that, the system must not only count stops. It must show which stops are influenceable and which measure has the highest effect. Otherwise, you optimize the loudest instead of the most expensive.

For managing directors, the best question is not: What does the software cost? The best question is: What loss are we currently accepting because we don't see it? This question hurts. It reveals that some factories have been living with a shadow tax for years – in the form of rework, special trips, schedule delays, discounts, and internal escalations. Manufacturing Intelligence promises to make this tax visible. You still have to pay it as long as no one acts.

Personal Forecast – The Next 2 to 3 Years

My forecast: By the end of 2028, Manufacturing Intelligence in DACH will no longer be perceived as an AI playground, but as normal operational infrastructure for demanding manufacturers. Not everywhere. Not at every 70-person company with three stable machines and a full order book. But at companies that have multiple lines, demanding customers, audit pressure, and real capacity conflicts. There, the question will not be whether production data is used. It will be why sales, planning, and maintenance still have different truths.

I expect three shifts. First, COOs will increasingly buy based on time-to-insight. Not on feature lists. Whoever shows reliable loss patterns in four weeks beats the provider with 80 features and a nine-month project plan. Second, sales will pull data from production into ICP, quotation prioritization, and account planning. This will generate resistance because it makes some favorite customers look worse. Third, ESG and energy data will lock into production optimization. Not out of idealism, but because customers and costs demand it.

Almetra's €16 million round is a marker for this, not an endpoint. It shows that investors believe in a European category: AI-native tools that directly generate value from shop floor data. Whether Almetra will be the dominant platform in five years, I don't know. No one knows. But I believe that the old separation between factory data and business decisions is breaking. Sales will move closer to the machine. The COO will move closer to the pipeline. And somewhere in a hall, an old label printer will continue to rattle, while a dashboard shows for the first time what that noise really costs.