Industrial IoT and Edge AI in Manufacturing: 2026 Guide

When we first published this article, it listed 20 emerging IoT technologies, from eSIM to LiFi. Some matured, some did not, and none of them was the reason a plant manager reads about IoT. The reason is downtime, scrap, and the gap between what the machines know and what the ERP shows.

This is a rewrite for that reader. It covers the industrial IoT stack as it is built in 2026, the two use cases with the clearest returns (predictive maintenance and computer vision quality control), where edge AI fits, how to get data out of PLCs over OPC UA and MQTT and into MES and ERP, what OT security requires, and why most pilots stall.

Where manufacturing AI stands in 2026

Adoption is broad and shallow. 77% of manufacturers use AI in some form, and 47% use it in quality processes, up from 33% in 2025; defect detection accounts for 44% of quality use cases, and more than 70% plan AI visual inspection within 18 months (Metrology News, IIoT World). The AI quality inspection market is estimated at $18.6B in 2026 with a 19.1% CAGR (Evolvance), and predictive maintenance at $13.9B to $19.8B in 2026 depending on the analyst (MarketsandMarkets, MRFR).

The outcomes that justify the spend are operational: 30 to 50% less unplanned downtime and 25 to 40% lower maintenance cost for manufacturers using AI (Deloitte, cited by AI Magazine). The barrier list is also consistent: labeled-data shortage (35%), integration (30%) and explainability (20%), plus operator trust, and 77% of AI manufacturing pilots never get past the prototype (Xenonstack, BuildMVPFast). Everything below is aimed at that last number.

What changed since the 20-technology list

Three things.

• Edge AI became practical. Industrial PCs and modules with GPUs or NPUs run vision and time-series models on the line at millisecond latency, with fleet management and over-the-air updates that IT can operate.
• Standards won. OPC UA for machine data models and MQTT (usually with Sparkplug B) for publish-subscribe transport are now the default way to get data off the floor, and most PLC and MES vendors support them.
• The data platform moved. Historians still exist, but the system of analysis is a cloud lakehouse fed by the edge, and the models trained there are deployed back to the edge.

Blockchain for IoT, LiFi and neuromorphic chips did not become part of a normal plant’s stack, and they are not in this guide.

The industrial IoT stack in 2026

A reference stack has six layers. Reading from the machine upward:

1. Sensors and controllers. PLCs, drives, CNCs, vision cameras, and retrofitted sensors (vibration, current, temperature, acoustic) on machines that have no native data interface.
2. Edge gateway. An industrial PC or module that speaks OPC UA and Modbus to the controllers, normalizes tags, buffers data when the network drops, and runs inference locally. AWS IoT Greengrass and comparable runtimes manage the software on it.
3. Broker and namespace. An MQTT broker with a unified namespace: one topic tree and one payload convention (Sparkplug B), so every consumer reads the same structure. AWS IoT Core or an on-premises broker.
4. Data platform. Time-series storage for high-frequency data, a lakehouse for history and features, and a catalog. On AWS this is typically AWS IoT SiteWise or Timestream plus S3, Glue and Redshift or Databricks.
5. Applications. Predictive maintenance, vision inspection, energy monitoring, OEE, digital twins, and the agents that act on their outputs.
6. Enterprise integration. MES for work orders and quality holds, ERP for maintenance orders, inventory and cost, CMMS for maintenance execution. This is the layer that turns a prediction into an action.

Design decisions in layers 2 and 3 decide whether the rest is easy. Get the namespace and the tag model right on the first line and every later line is a copy.

Predictive maintenance

Predictive maintenance uses sensor data to estimate the condition of equipment and the probability of failure, so maintenance happens before the failure and not on a calendar.

Data. Vibration and motor current for rotating equipment; temperature, pressure and flow for process equipment; PLC alarms and cycle counts for everything. Most plants have to add sensors for vibration; the rest is usually already in the PLC.

Models. Anomaly detection on the healthy baseline is the starting point, because failure examples are rare. Remaining-useful-life and failure-mode classifiers come later, when the CMMS history is clean enough to label events. Physics features (bearing frequencies, harmonics) still beat raw deep learning on small datasets.

Where it runs. Feature extraction and anomaly scoring at the edge, so the alert arrives in seconds and bandwidth stays low. Model training in the cloud on the full history.

Integration. An alert that does not become a work order does nothing. The application creates the order in the CMMS or ERP with the evidence attached, and the technician’s finding closes the loop as a label for the next model version.

How to start. One asset class with a known failure cost and a maintenance team that wants it. Run the model in shadow mode for a full cycle, compare its alerts with what the technicians found, then let it create orders.

A digital twin extends this: given a failure probability, the twin answers when the cheapest intervention is, given the schedule. See Digital twins and AI simulation.

Computer vision quality control

Vision QC puts cameras on the line and a model behind them to detect defects, verify assembly, read codes and measure parts. It is the most common manufacturing AI use case, and the most common failed pilot.

Data. The labeled-data bottleneck is the top barrier for a reason. Defects are rare and varied. Plan a labeling workflow from day one, use anomaly detection on good parts to catch the unknown unknowns, and use synthetic and augmented data to cover rare classes. Line changes, lighting and camera drift will change the images; monitoring for that is part of the system.

Models. Detection and segmentation models trained on the plant’s own images, sometimes starting from a foundation model. Explainability is a practical need: operators trust a box around the defect more than a score.

Where it runs. On the edge, next to the camera, with inference in the tens of milliseconds so the reject mechanism can act. Models are exported to ONNX or TensorRT and deployed to industrial PCs or modules; the cloud handles training and fleet management.

Integration. A defect is a quality event in the MES, a hold on the batch if the rate crosses a threshold, and a data point for process engineering. Vision that only produces a dashboard gets ignored.

Human review. Uncertain cases route to a person, and their decisions feed the next training set. The false-reject rate is what operators judge the system on, so track it and publish it.

Our manufacturing practice covers vision QC deployment from camera selection to MES integration.

Edge AI: where the model runs and why

Run inference at the edge when latency, bandwidth or data sensitivity require it, which for vision and high-frequency sensor data is almost always. Train in the cloud, where the history and the GPUs are.

What an edge deployment needs, beyond the model:

• Hardware sized for the model: industrial PCs with GPUs or NPUs, or modules such as NVIDIA Jetson, in enclosures rated for the environment.
• Runtime and packaging: containerized inference with a defined interface to the camera or the broker.
• Fleet management: versioned models, staged rollouts, rollback, and health telemetry across dozens or hundreds of devices.
• Store-and-forward so a network outage loses no data and no inference.
• Monitoring for input drift (lighting, sensor degradation) as well as model performance.

Treat the edge fleet as software infrastructure with releases, not as a set of appliances. Erbis’s cloud and AI infrastructure practice runs this with the same DevOps discipline as any other deployment.

OPC UA, MQTT and getting data out of the PLC

The integration problem is usually harder than the modeling problem.

OPC UA is the standard interface for industrial equipment. It provides an information model (what the tags mean, in a hierarchy with types and units) plus secure transport. Companion specifications exist for machine types, so a CNC or an injection molding machine exposes a known structure. Use OPC UA to read from controllers and to keep the semantics.

MQTT is a lightweight publish-subscribe protocol. With the Sparkplug B specification on top, it carries structured payloads and device state (birth and death certificates) and supports many consumers without loading the PLC. Use MQTT to move data from the edge to the platform and between applications.

Modbus, PROFINET and EtherNet/IP are still everywhere on older equipment; the edge gateway translates them.

CAN bus is the equivalent standard in vehicles and mobile equipment. For WH Bence, a UK special-vehicle manufacturer, Erbis built an embeddable monitoring console that treated every sensor and subsystem as a node on the CAN bus, so one console could control lights, hydraulics, generators and communications across every vehicle the company builds. The lesson transfers: standardize on the bus, not on the device.

A unified namespace, one MQTT topic hierarchy that mirrors the plant (enterprise, site, area, line, cell, asset), is the single most useful design decision in an IIoT program. Every application publishes to and subscribes from it, so integrations stop being point to point.

Integrating with MES and ERP

ISA-95 still describes the layers: sensors and control (levels 0 to 2), MES (level 3), ERP (level 4). IIoT and AI do not replace the layers; they add data flowing up and decisions flowing down.

• MES receives quality events, machine states and OEE inputs, and issues holds and rework instructions.
• ERP (SAP, Oracle, Microsoft Dynamics and others) receives maintenance orders, consumption and cost, and sends production orders and material availability.
• CMMS executes maintenance and returns the findings that label the next model.

Closing the loop means every model output has a system-of-record action attached, and a person approves the actions that carry cost or safety risk. This is where the agentic pattern applies: an agent reads the vision and maintenance outputs, drafts the work order or the hold, and submits it after approval, with every step logged. Erbis’s AI automation and integration work is built around exactly these ERP and MES connections.

Digital twins and smart sensors

Two items from the original list survived because plants use them.

Smart sensors carry processing on board, so a vibration sensor can compute its own spectrum and publish an alarm over MQTT without a gateway. They simplify retrofits and reduce edge compute needs.

Digital twins use the same live data to run simulations of the line, the plant or the network, so a maintenance window or a schedule change can be tested before it is made. The companion guide, Digital twins and AI simulation for manufacturing and supply chain, covers them in depth.

Security for OT

Connecting the floor to the platform creates an attack surface that did not exist when the PLC was air-gapped.

• Segment networks according to the Purdue model, with an industrial DMZ between OT and IT; data leaves OT through the broker, not through direct connections.
• Follow IEC 62443 for zones, conduits and device requirements.
• Give every edge device an identity (certificates), rotate credentials, and sign models and software updates.
• Log access to controllers and broker topics; alert on unexpected writes.
• Keep the ability to run the line with the platform down.

Erbis holds ISO 27001 certification and designs OT connectivity with security review as part of the architecture, not as a step at the end.

Why 77% of pilots stall, and how to avoid it

The pilot that stays a pilot usually shares four features: it ran on exported data rather than live data, it never wrote to a system of record, no operating metric was named up front, and the operators were shown the result rather than involved in it.

The plan that gets to production is the reverse.

1. Pick one line, one asset class and one number: unplanned downtime hours, first-pass yield, scrap cost.
2. Connect live data through the namespace from the start, even for the pilot.
3. Run in shadow mode next to the current process, then switch.
4. Write the result into MES, ERP or CMMS, with human approval where it matters.
5. Give operators the explanation, the override and the credit.

A fixed-fee 4-week AI pilot is Erbis’s format for this: a defined asset or line, live data, a named metric and a go/no-go decision at the end.

What Erbis brings

Erbis has 200 engineers and has built industrial software since 2012 and production AI since 2017. We are an AWS Partner (Select Tier) and a Databricks Consulting Partner, which covers the two platforms where most manufacturing data ends up. We build the edge, integration and data layers as well as the models, because in manufacturing the model is the smaller part of the work.

FAQ

What is the difference between IoT and industrial IoT?

Industrial IoT applies connected sensors and analytics to machines, lines and plants, with requirements consumer IoT does not have: deterministic latency, ruggedized hardware, OT protocols such as OPC UA and Modbus, safety and OT security standards, and integration with MES and ERP.

Should we use OPC UA or MQTT?

Both. OPC UA to read structured data from controllers and preserve its meaning; MQTT with Sparkplug B to move that data to the platform and between applications through a unified namespace.

Do we need edge computing for predictive maintenance?

For high-frequency data such as vibration, yes: feature extraction and scoring at the edge keep latency and bandwidth manageable. Slow signals such as temperature can go straight to the cloud.

How much labeled data does vision QC need?

Less than teams expect if they start with anomaly detection on good parts, and more than they expect for rare defect classes. Plan a labeling workflow and synthetic data from the start, and budget for relabeling when the line changes.

If you have a line, an asset class and a number you want to move, talk to Erbis.