AIoT: AI That Makes Decisions on the Device
From connected kitchen appliances to robots on the assembly line: we bring AI to where your product operates – even when no cloud is within reach.
Why AI Belongs on the Device, Not Just in the Cloud
Many decisions have to be made where the cloud cannot reach: in a dead zone, on a shop floor that is deliberately disconnected from the network, or at volumes where data transmission simply does not pay off. Yet the decision still has to be made – within milliseconds, reliably and on hardware with limited computing power. On top of that, there is a familiar pattern: what impresses as a demo in the lab often fails to make the leap into series production.
With AIoT, we bring AI directly onto your device or onto an edge computer on site. This ranges from compact models on microcontrollers and physical AI in robotics and autonomous systems to automated quality inspection on the line. We develop for series production from the very start, not for the demo. And we keep an eye on what actually pays off at your volumes.
Less Technology. Better Business.
Where AIoT Comes In for You
AIoT covers three very different starting points. Find the entry point that matches your role and your task:
| For Your Product | For Your Manufacturing | For Your Data |
|---|---|---|
Trigger:
• New operating concepts, assistance functions or more autonomy are needed. • Cloud dependency, latency or data protection are holding back previous AI approaches. • The product must also work reliably and securely offline. | Trigger:
• Quality problems are detected too late. • There are not enough skilled workers at the inspection stations. • Scrap and rework are on the rise. | Trigger:
• Production generates thousands of data points every day, but nobody can say for sure why the line performed worse last week than the week before. • There are dashboards, reports and Excel analyses – but no reliable recommendations for action. • The same causes are discussed over and over again. |
Solution:
AI directly in the product or on an edge device – from microcontrollers to controllers, ready for series production and updatable throughout the product lifetime. | Solution:
Use AI where production happens – systems that perceive and act physically, from robotics to in-cycle quality inspection. | Solution:
Turn data into better decisions: combine production know-how with data analytics and AI. Instead of just visualizing KPIs, we uncover influencing factors, patterns and levers for improvement. |
What AIoT Is – and How It Relates to Edge AI and Physical AI
The term is new, combines AI and IoT and is used inconsistently in the market. This is how we define it: AIoT connects networked devices with artificial intelligence. The difference from classic IoT lies in the role of the device. An IoT device captures and reports – it collects measurements and passes them on. An AIoT device interprets and decides – it assesses the situation and derives an action from it. The decision is therefore made on the device, no longer in a system behind it.
Edge AI and physical AI define two different directions within it:
| Term | What it defines | Example |
|---|---|---|
| AIoT | The product decides for itself. | A robotic lawn mower detects an obstacle and chooses a new path. |
| Edge AI | Where the processing happens: locally instead of in the cloud. | The model runs on the chip of the edge computer in the factory hall. |
| Physical AI | What the AI does: perceive, plan, act physically. | Gripping, driving, avoiding obstacles. |
“Edge” is often used for any device that is not located in a data center. To be more precise, three levels need to be distinguished:
| Level | Where the processing happens | Role |
|---|---|---|
| Device | In the product itself: microcontroller, controller or SoC. | Decides on its own. |
| Edge | On a dedicated on-prem device on site – a gateway or server in the control cabinet, in the hall, plant or vehicle. | Aggregates several machines and is the only handover point to the cloud. |
| Cloud | In a provider's data center, accessible only via the network. | Only sees what the edge passes on: fleet view, retraining, long-term analysis. |
Accordingly, edge AI means that the model runs on one of the two upper levels – on the device itself or on the edge. Which of the two is right depends on whether the decision has to be made in the individual product or whether several machines need to be considered together.
Edge AI and physical AI are independent of each other. An automated guided vehicle acts physically but receives its routes from a fleet management system in the data center. A camera in quality inspection decides locally between good and bad parts without moving an inch.
Things get demanding where both come together: a device that acts physically and is on its own has to decide within a few milliseconds on limited hardware – without asking back and without human approval. A wrong decision cannot be corrected with a second attempt, because the movement has already happened. The requirement thus shifts from model quality to the behavior of the overall system.
That is why we define two things for every AI function: how the device recognizes that its assessment is uncertain (confidence, plausibility limits, input data outside the trained range) – and what it does then: stop, continue more slowly, fall back on simple rule-based behavior or hand over to a human.
Where the Term Physical AI Comes From
In research, this field has been called embodied AI for years. “Physical AI” is the industrial term that has been shaped mainly by chip and platform providers since 2025. The term is new, the subject behind it is not. What has shifted is something else: where robots used to execute programmed motion sequences, today they work with trained models that can also handle situations nobody described in advance. This removes the limit that automation used to run up against: small batch sizes, changing variants and parts without an exact position become economically viable. The price for this is that the system's behavior can no longer be fully specified in advance.
When No Cloud Is Available, None Is Allowed or It Doesn't Pay Off
“This has to work without the cloud, too” – we hear this sentence in many customer conversations. Behind it are three very different motivations.
-
No connection available. A robotic lawn mower at the far end of the garden, a machine under a reinforced concrete ceiling, a vehicle in a dead zone: here, losing the connection is the norm, not a malfunction. It does not change the task – the machine still has to sort out the defective part, the vehicle still has to brake.
-
No connection allowed. In many German plants, the shop floor is deliberately separated from the internet – for operational safety and to protect know-how. This separation will not be lifted just because someone wants to use AI. Cloud-bound AI functions therefore rarely fail because of the technology, but because of IT security approval.
-
Connection possible, but not economical. This is the rule with high-frequency data – vibration and current measurements in the kilohertz range, camera images at production cycle speed, sensor streams from dozens of machines. Anyone who permanently pushes such raw data into the cloud pays three times: for transmission, storage and analysis – continuously, per machine and per device. Yet the vast majority of this data is unremarkable normal operation.
In all three cases, the computing stays local – on the device or on an edge computer on site. Only the result goes out: the detected anomaly, the condensed metric, the inspection verdict. That is the difference between a permanent stream of raw data and a few bytes per event. The benefits:
- no data leakage,
- no latency over the network,
- no ongoing costs for transmission, storage and inference,
- no dependency on a provider's availability – and the function keeps working when the connection is lost.
This is certainly more demanding than the cloud route: limited memory, limited computing power, an energy budget that has to be met, and an update path that lasts for years. This very work is our core business – we have been developing embedded software for decades and connected products for more than ten years. Where a connection exists and makes sense, we use it: for fleet view, retraining and central analysis, via the edge as the only controlled gateway. But your product's core function does not depend on it.
These Companies Already Trust Our Expertise
Three Fields We Work In
AI on the Device (Edge AI)
The model runs in the product itself – on a microcontroller or controller – or on a dedicated edge device on site that combines the data of several machines. Both are edge AI: measurements are analyzed where they are generated. At most, the result goes out, via the edge as the only handover point. Four requirements typically lead down this path: a response within a few milliseconds, data transmission that is too expensive, data that must not leave the site, or a device that simply has to work without a connection.
We bring models to the existing hardware – from microcontrollers to embedded Linux systems with AI accelerators (a chip that computes neural networks many times faster and more efficiently than the main processor). If the hardware has not yet been decided, we support you in selecting it and measure in advance what the target system can actually deliver. And the models remain updatable throughout the product lifetime.
Systems That Act Physically (Physical AI)
Perceive, plan, act – in an environment that does not follow specifications: an assistance robot in assembly, a vehicle in traffic, a machine in the field or in the barn. What sets these systems apart from classic automation is that they have to deal with situations for which there is no rule. This includes motion planning, collision avoidance and sensor fusion – computing a shared picture of the environment from camera, lidar, radar and other sensors. Testing takes place in simulation and against scenarios. When humans and machines share the same space, functional safety comes into play. The automotive world also belongs here, all the way to V2X, where decisions are made in milliseconds.
Automated Quality Inspection
Inspection at production cycle speed, every single part instead of random samples – AI-supported with cameras and sensors directly on the line or in the device. The benefits pay off quickly: less scrap, fewer inspection stations, fewer complaints. The difficulty lies elsewhere – an automated inspection must be ready for acceptance. You can find out what this involves in the FAQ.
Classic applications such as Predictive Maintenance, Condition Monitoring and Smart Maintenance build on the same foundations.
Our Case Studies and Projects
The best way to see what AIoT looks like in practice is through completed projects. Our examples show the wide range of possible applications.
-
About the Weigerstorfer ReferenceWeigerstorfer GmbH: Project Management Platform for Building Services EngineeringAI & MLData PlatformsIT ModernizationAbout the Weigerstorfer ReferenceUnified Project Management Platform for Office and Construction Site Use in Building Services Engineering
About the Weigerstorfer ReferenceProcess digitization, mobile-first app, GAEB integration, AI-powered data migration (Azure)
About the Weigerstorfer Reference250,000 orders migrated in four months, paperless maintenance, overview of 3,000 maintenance contracts
-
About the ifm Referenceifm: Azure AI strengthens global procurement capabilitiesAI & MLCloudIndustry 4.0About the ifm ReferenceStreamline global supplier evaluation in procurement instead of relying on individual search strategies that require a lot of manual work
About the ifm ReferenceA chat-based AI assistant built on Azure OpenAI and AI Search that aggregates ERP data, website information, and internal reviews
About the ifm ReferenceMVP up and running in 11 weeks; new employees can get started without a lengthy onboarding process; Azure AI Platform as the foundation for additional use cases
-
See robotics referencePhysical AI: Retaining Physically Impaired Workers in ManufacturingAI & MLEmbedded & RoboticsIndustry 4.0See robotics referenceCustomized assistance robots for people with physical disabilities in production
See robotics referenceIntegration of AI for automated adaptation of robots to people's capabilities
See robotics referenceEffective empowerment of people with physical disabilities
-
About the Monitoring ReferenceDräger: Improved workplace safety in industrial facilities thanks to the cloudCloudIoTCybersecurityAbout the Monitoring ReferenceCloud Platform for Live Monitoring and Management of Mobile Gas Detectors
About the Monitoring ReferenceCloud-native microservices, IoT gateway, scalable Kubernetes backend (Azure)
About the Monitoring ReferenceReal-time alerts worldwide in under 10 seconds, 140,000+ devices
-
To the BMW referenceVirtualization of software testing for control units in the AWS cloudCloudMobilityEmbedded & RoboticsTo the BMW referenceDigital twin for control unit development & testing in cars
To the BMW referenceVirtualization of control units & AWS cloud integration
To the BMW referenceSetup of virtual cars & control units without expensive hardware setups, worldwide & distributed testing
-
To the STARTRAIFF referenceSTARTRAIFF: AI-powered analysis supports insurance salesAI & MLCloudData PlatformsTo the STARTRAIFF referenceAggregation of internal customer data & external data in a single web application
To the STARTRAIFF referenceData bundling & analysis with Amazon Bedrock
To the STARTRAIFF referenceIntuitive user interface for sales, 88% reduced preparation time before customer visits
-
To the MAN referenceMAN - ATLAS L4. Control Center for the autonomous truckMobilityCloudData PlatformsTo the MAN referenceControl center for the technical monitoring of driverless trucks
To the MAN referenceUX design, product strategy, data structure, vehicle data visualization
To the MAN referenceMonitoring, remote support, mission management, reports for commercial autonomous transport solutions
-
To the NETZSCH referenceNETZSCH: Development of an IoT platformCloudIoTData PlatformsTo the NETZSCH referenceUnified IoT platform for 3 business units, harmonization of existing IoT solutions
To the NETZSCH referenceIoT device connectivity, visualization software for data analysis, cloud infrastructure, operations
To the NETZSCH referenceQuick testing in the cloud infrastructure, fast integration of use cases such as predictive maintenance, process optimizations, etc.
-
To the TKE referenceTK Elevator: Health Check Connectivity for the IoT gateway of elevatorsCybersecurityIoTEmbedded & RoboticsTo the TKE referenceIoT gateway (MAX Box) for data connection between elevator & IoT platform
To the TKE referenceExamination of code quality, architecture, operations & organization
To the TKE referenceOptimization of IoT gateway connectivity & digitalization of elevators
-
To inCTRL referenceinCTRL Solutions: Modernization of the IoT platform for water treatment plantsCloudIoTIT ModernizationTo inCTRL referenceIoT & software modernization, integration of new functions
To inCTRL referenceData warehouse setup, integration of microservices, automated quality assurance, Continuous Integration & Continuous Deployment (CI/CD)
To inCTRL referenceImproved resilience, maintainability & further development capability of the platform
-
To the Mixed Reality referenceUniversity Medicine Frankfurt - Mixed reality for the training of doctorsDigital & UX DesignAppsEmbedded & RoboticsTo the Mixed Reality referenceKnee training simulator for technology-supported medical education
To the Mixed Reality referenceMixed reality application combined with a haptic knee model, interactive training environment
To the Mixed Reality referenceDiagnostics on the human knee without simulated patients, simulation of various pathologies, real-time visualization of motion sequences
-
To the Health.exe referenceHealth.exe: AI-supported platform creates training plans for patientsAI & MLCloudData PlatformsTo the Health.exe referenceAI-supported service for orthopedic & sports medicine practices
To the Health.exe referenceCloud-based web application for doctors for the automated, evidence-based creation of individually tailored patient training plans
To the Health.exe referenceNew revenue source without fixed costs, higher patient retention, AI-supported & guideline-based
-
See MAN referenceMAN: Efficient threat analysis for control unitsCybersecurityIoTEmbedded & RoboticsSee MAN referenceProtection of digitalized trucks against virtual attacks
See MAN referenceRisk analysis based on 4x6 methodology, ThreatSea, ISO21434
See MAN referenceQuick identification of relevant threats for immediately effective security measures
-
See Siemens referenceSiemens: AI demand prediction platform for industrial production planningAI & MLCloudData PlatformsSee Siemens referenceMachine learning for time series forecasting
See Siemens referenceAutoML for automated adaptation of models to different data
See Siemens referenceUnified, scalable solution, optimized inventory costs, efficiency gains
-
See TÜV NORD referenceTÜV NORD GPT: Development of AI assistanceAI & MLAppsWeb & PortalsSee TÜV NORD referenceSecure operation of AI in the European MS Azure cloud environment
See TÜV NORD referenceFrontend & backend via MS Azure App, "Chat with your PDF" for TÜV employees
See TÜV NORD referenceQuick implementation of new technologies (AI), strengthening knowledge management
-
See Miele referenceMiele domestic appliances are networked worldwideCloudIoTEmbedded & RoboticsSee Miele referenceFurther development of the IoT platform for connected home appliances
See Miele referenceContainer-based architecture, open standards, modular design
See Miele referenceQuick availability & scalability of digital services, high added value for users
-
See BMW referenceMixed reality: driving a real car in a virtual worldDigital & UX DesignAppsEmbedded & RoboticsSee BMW referenceExperience a futuristic outdoor environment in the headset while inside a real vehicle interior
See BMW referenceReal-time 3D application, Unreal development, XR technology, game design, smart devices
See BMW referenceHigh-performance processing of vehicle signals for precise motion mapping for an immersive driving experience without motion sickness
-
See KUKA referenceKUKA: Web interface for a new Human Machine InterfaceDigital & UX DesignEmbedded & RoboticsWeb & PortalsSee KUKA referenceEasy & intuitive web UI for product control
See KUKA referenceDevelopment of HMI components & modules using Scrum
See KUKA referenceHigh user-friendliness, reliable operation, fast scalability & extensibility
-
See STIHL referenceSTIHL: Control iMOW robotic mower via appCloudIoTAppsSee STIHL referenceControl and configuration of the robotic mower via smartphone
See STIHL referenceDevelopment of app, web, cloud platform and direct Bluetooth communication
See STIHL referenceDigital benefits for users, app controllability, remote software updates
-
See ifm services referenceifm services: Remote maintenance of systems and machinesCloudIoTEmbedded & RoboticsSee ifm services referenceFully integrated remote access in the IoT platform
See ifm services referenceFull stack cloud application, RUST-based clients, UX design
See ifm services referenceAnalysis of sensor data from production as a basis for sustainable decisions for customers
-
See referencePlanning systems: Optimizing the capacity utilization of pressing plantsAI & MLData PlatformsIndustry 4.0See referenceCentralized planning of component manufacturing for cost- & resource-optimized production capacity worldwide
See referenceConversion from local processing with fat clients to a client-server application, migration to the cloud
See referenceData-based planning & calculation of different manufacturing scenarios & site-specific production costs
-
See referenceGlobal workforce planning systemAI & MLCloudData PlatformsSee referenceCentralized web-based IT system to replace individual isolated solutions
See referenceEvent sourcing for planning & analytics, domain-driven design, cloud migration
See referenceEasy updates, expansion, maintenance, optimized security
-
See BMW Group referenceBMW Group: Remote software upgrade for vehiclesCloudMobilityCybersecuritySee BMW Group referenceSoftware upgrades without the need to visit a service center
See BMW Group referenceBackend system for over-the-air communication with the vehicle, 24/7 support
See BMW Group referenceIT security, more comfort, on-demand provision of new features
-
See SMA referenceSMA: Development of a Web UI for ennexOS platformDigital & UX DesignIoTWeb & PortalsSee SMA referenceCreation of a unified customer experience across all products, smooth generational transition for customers, secure, agile operation
See SMA referenceWebUI for the digitalization & automation of energy management processes, open-source solution for energy flow visualization
See SMA referenceEnergy flow & cost optimization, operational reliability, customer-friendliness
-
See Weidmüller referenceWeidmüller: Progression of the Industrial Service PlatformCloudIoTWeb & PortalsSee Weidmüller referenceCreation of a centralized, intuitive, expandable portal as the foundation for industrial applications (remote access, data visualization, ML)
See Weidmüller referenceExploration, setup & further development of the base platform for industrial services
See Weidmüller referenceInnovative portal for end-to-end solutions, MVP in just 7 months
Even unstructured, even incomplete. Within a few days, you will know which AI use cases your data supports today, where the gaps are and what pays off fastest.
Why Product Manufacturers and Shop Floor Operators Work with MaibornWolff
Production ramp-up, bill of materials costs, certification, service cases, shift operation: we speak the language of your development and your production teams.
From hardware selection to the cloud in one team. We measure on your target hardware before you commit – no handover gaps between AI boutique, embedded specialist and cloud service provider.
Built for continuous operation: more than 30 business-critical IoT platforms, millions of devices in the field.
Mixed teams, ongoing knowledge transfer, and the source code we create is always handed over. Models and data belong to you.
In numbers: more than 10 years of IoT · 30+ IoT platforms · 800+ employees · IT services since 1989 · certified to ISO 27001 and TISAX.
How an AIoT Project Gets Started
We work in verifiable steps. Each one ends with a result on which you can base your decision.
Duration: 3–4 person-days
Your ideas are assessed for benefit and feasibility. You receive a prioritized list and a clear next step.
Duration: 2–3 weeks
If data is available: we assess which use cases it supports today. The result is a set of prioritized recommendations for action.
Duration: 3–6 weeks
If the device is the question: a model runs on your target hardware. We measure accuracy, latency, memory and energy requirements.
Duration: from 3 months
Integration into your product or plant – with an acceptance concept, update path and monitoring from the start.
After that, we support rollout and operation for as long as you want.
In a free initial consultation, we clarify whether and how it can be implemented on your hardware.
Frequently Asked Questions About AIoT
What Is AIoT?
AIoT connects networked devices with artificial intelligence. The difference from classic IoT lies in the role of the device: an IoT device captures measurements and reports them. An AIoT device assesses the situation and derives an action – a kitchen appliance recognizes from the camera image when the food is done and reduces the heat, a machine hears from the sound that a tool is getting blunt. The decision is made on the device, not in a system behind it.
What Is the Difference Between AIoT, Edge AI and Physical AI?
At the top level, AIoT means that the product makes decisions itself with the help of AI. Edge AI defines where the processing happens: locally instead of in the cloud, in the product or on an edge device on site. Physical AI defines that the AI perceives, plans and acts physically – in robotics, autonomous vehicles, assistance systems. Edge AI and physical AI are independent of each other; things get demanding where both come together.
Does AI on the Device Need an Internet Connection?
No. A model on the device works completely locally – without the cloud, without a network, without data leakage. This is crucial for devices that are offline at times and for production environments that are deliberately kept separate. A connection is used where it adds value (fleet view, retraining) – but the core function does not depend on it.
Is Our Existing Hardware Sufficient for AI on the Device?
The feasibility phase answers this on your specific hardware. In general, the requirements have dropped significantly: classic anomaly and pattern detection runs on microcontrollers, compact language models on more powerful controllers. Accuracy, latency, memory and energy requirements are measured on your target system before you commit.
How Do We Keep an AI Model Up to Date Over the Product Lifetime?
The same way as firmware: signed updates with rollback and documentation showing which device has been running which model version since when. Models age because materials, environments and failure patterns change – so the update path belongs in the architecture from the very beginning. The Cyber Resilience Act requires secure update paths for connected products over the entire support period anyway. A model on the device is part of the software and follows the same path.
How Do We Approve an AI Function That Does Not Always Give the Same Answer?
Through defined acceptance criteria, guardrails and tests against scenarios instead of individual cases, complemented by monitoring in the field. Classic test procedures require reproducibility and no longer apply here. For every function, we define how the device recognizes uncertainty and what it does then. If an AI function is safety-relevant, for example as a safety component of a machine, additional requirements for traceability and human oversight apply. Which rules apply depends on the use case.
What Does AI on the Device Cost Compared to the Cloud?
The cost structure is different. The cloud route incurs ongoing costs that grow with your success – transmission, storage, every inference. A model on the device costs once, in development and the bill of materials; after that, the computing power is paid for. With high volumes or frequent queries, the calculation usually tips clearly in favor of the device. However, the AI accelerator on the circuit board is a series production decision that has to be made early.
What Is the Difference Between Edge and Cloud – and Why Does IT Security Draw the Line There?
The cloud is a provider's data center, accessible only via the network. The edge is not just any device outside of it, but a dedicated instance at your site that aggregates the data of several machines and is the only handover point to the cloud: whatever is not let through there does not leave the site. That is exactly why IT security and plant security draw the line here – one hardened, monitored instance instead of hundreds of machines with their own internet connection. This needs to be distinguished from the third level, the device itself, on which a model computes directly in the product and works without any connection at all.
How Do We Protect Our Model on a Device Located at the Customer's Site?
A trained model contains your process know-how, and on a delivered device it is physically accessible. Protection comes from the architecture: secure memory areas, encrypted model storage and secure boot. This belongs in the concept phase.
Who Owns the Data and the Trained Models?
To you – data, models and the source code we create. This is not up for negotiation. If the model runs on the device, the raw data stays there anyway. During training, data is brought together; where this happens, who has access and what happens to it after the project ends is agreed in writing beforehand.