According to McKinsey , one of the world’s leading strategy consulting firms, 65% of organisations were already using generative AI in at least one business function by 2024—twice as many as the year before. And that was before autonomous agents entered the picture. Today, the question is no longer whether to implement AI agents, but how to do it effectively.
For some, an “agent” is little more than a chatbot with memory. For others, it is a system that can detect a production issue, assess its impact on the supply chain, and trigger a work order in less than three seconds. This article is for those who want to be in the second group.
1_ What you need before you start
Three non-negotiable requirements. If any of them fail, the agent will fail with them.
1_Real-time data
An agent cannot act on information that arrives hours late. You need data sources that can be queried in real time: sensors, APIs, ERP systems, SCADA platforms. If your freshest data still comes with significant delay, start there.
2_A use case with a clearly defined decision
Not “optimise production”, but: “when the energy consumption of line 3 exceeds threshold X for 15 minutes, adjust the speed of the conveyor belts.” Absolute specificity.
3_Clearly defined error tolerance
Every agent will make mistakes. Define in advance which errors are acceptable, which should be escalated to a human, and which should prevent any action from being taken.
2_ The 5 steps to implementing AI agents
Define the use case precisely
Look for processes where three conditions are met: data is available, decisions follow recognisable patterns, and the cost of inaction can be measured. Start small, aim for high impact, and keep the risk low.
Audit your data sources
In industrial environments, it is common to find PLCs, SCADA systems, MES platforms and ERPs that do not communicate with one another. Data integration is rarely the most visible challenge, but it is often the one that causes the most projects to fail.
Choose the architecture
Rule-based — Deterministic, auditable and ideal for getting started. They do not learn, but they work.
With predictive models — They anticipate failures before they happen.
With LLMs — Flexible reasoning for unstructured contexts; more powerful, but harder to audit.
Multi-agent — Several specialised agents working together. This adds complexity, but makes it easier to scale.
Integrate it with your systems
An agent that makes decisions but cannot execute them is just a more expensive alerting system. You need APIs, webhooks or native connectors — and, in industrial environments, adapters for OPC-UA or MQTT.
Practical tip: deploy it in observation mode for the first 2-4 weeks and compare its decisions with human decisions before giving it real control.
Define metrics and oversight
Three metrics are essential: decision accuracy, the frequency of escalation to humans (too many escalations usually indicate poor design; none at all may indicate a lack of oversight), and response time. If you do not measure speed, you cannot demonstrate its value.
3. OT Security and IT/OT Convergence
OT Security: The First Question Every Plant Manager Asks
“Are you really going to let an AI agent interact with the PLC?” It is the first question that comes up in almost every meeting—and rightly so.
Industrial environments present a very different risk profile from corporate IT. A failed server can simply be restarted; a stopped production line or an incorrectly commanded actuator can lead to physical damage, costly downtime, or safety incidents. That is why deploying AI agents in OT environments requires a security-first approach.
OT Network Segmentation: A Non-Negotiable Requirement
The OT (Operational Technology) network—where PLCs, sensors, and control systems operate—must be physically or logically isolated from the corporate IT network and from any internet-connected AI agent components.
The recommended architecture follows the Purdue Model or ISA/IEC 62443, where AI agents operate at the supervisory layer (Levels 3–3.5) and never communicate directly with the control layer (Levels 1–2), except through secure, audited gateways.
Industrial protocols such as OPC UA provide certificate-based authentication and native encryption, while MQTT can be secured using TLS and client authentication. Under no circumstances should these protocols be exposed on the OT network without encryption.
Observation Mode: The Safest Way to Introduce AI
This guide recommends deploying an AI agent in observation mode for two to four weeks before enabling any automated actions. In OT environments, this is not simply a best practice—it is a fundamental safety requirement.
During this period, the agent has access to live operational data but does not execute any control actions. The engineering team can validate that its recommendations are appropriate before granting it permission to interact with production systems.
Once automation is enabled, strict operating boundaries must be enforced. The agent should only be allowed to adjust parameters within predefined limits, must never issue emergency stop commands, and any situation outside the approved operating range should automatically be escalated to a human operator.
This approach—observe, recommend, act within defined boundaries, and escalate when uncertain—provides the level of control needed to deploy AI agents safely in industrial environments.
What to Require from Any AI Agent Vendor for Industrial Environments
- An architecture with an isolated OT network and a secure, audited gateway between IT and OT.
- Immutable logs recording every decision and action performed by the agent.
- Automatic rollback or fail-safe mechanisms if connectivity with the supervisory system is lost.
- Clearly defined contractual operating boundaries before deployment begins.
4. EU AI Act and Compliance
EU AI Act: What Every Industrial Decision-Maker Needs to Know in 2026
The European Union Artificial Intelligence Act (EU AI Act) entered into force in August 2024 and is being implemented progressively. For industrial organisations, the key question is: what risk category does my AI agent fall into?
Risk Classification for Industrial AI Agents
Most AI agents used in industrial automation are likely to fall under the high-risk category (Annex III), particularly when they interact with critical infrastructure, workforce management, or process safety. In these cases, organisations must meet a number of requirements before deploying the system:
- Technical documentation: a complete description of the system architecture, training data, and known limitations.
- Decision traceability: the system must record the data it used, the decisions it made, and the reasoning behind them, allowing an auditor to reconstruct its behaviour.
- Human oversight: the agent must be designed so that a human operator can intervene, override, or stop the system at any time.
- Risk management: organisations must perform a formal risk assessment before deployment and review it regularly throughout the system’s lifecycle.
AI agents based exclusively on deterministic rules generally have a lower compliance burden than those powered by Large Language Models (LLMs), where explaining how a particular decision was reached is inherently more challenging.
Why Traceability Is a Competitive Advantage, Not Just a Compliance Requirement
A system that records every decision together with its context—input data, applied thresholds, actions taken, and resulting outcomes—not only helps organisations meet regulatory requirements, but also creates a valuable historical record for continuously improving the agent’s performance. It also provides clear evidence of the system’s value to management and during external audits.
In practice, organisations that build traceability into their AI systems from day one are the ones that can demonstrate a measurable return on investment (ROI) using their own operational data within just a few months.
At WonderBits, we design logging and explainability as core architectural components from the outset, rather than treating them as features to be added later.
5_ Use cases by sector
- Manufacturing.
Failure prediction and automatic generation of work orders. Typical 30-50% reduction in unplanned downtime. - Energy.
Real-time consumption balancing based on electricity tariffs, with integration of on-site renewable generation where available. - Logistics.
Detection of inventory deviations, automatic orders within approved limits, and delivery time adjustments based on supplier data. - Quality.
Defect classification using computer vision and in-line parameter adjustment before the defect occurs.
6_ Mistakes to avoid
- Starting with the most complex use case.
Successful projects reach production in 6-8 weeks. Failed ones try to solve everything at once. - Underestimating data integration.
The agent architecture is the fun part. Connecting it to existing systems takes up 60% of the real work. - Not assigning ownership.
An agent in production needs an owner who reviews its metrics and decides when it needs to be adjusted. Without that role, it ends up being ignored or feared.
7_ How much does it cost?
With already connected and normalised data: between €15,000 and €50,000 for a first agent in production. If integration needs to be solved first, that cost can multiply. In industrial environments, ROI is typically justified within 6-18 months through reduced downtime, energy consumption or rejection rates. Without metrics from day one, there is no way to prove it.
8_ Glossary: Key Terms in Industrial Automation with AI
- PLC (Programmable Logic Controller)
A programmable industrial controller that executes real-time control logic for machines and industrial processes. It is often considered the “brain” of a production line. - SCADA (Supervisory Control and Data Acquisition)
A supervisory control and data acquisition system used to monitor and control industrial processes, often across geographically distributed facilities, from a central control station. - MES (Manufacturing Execution System)
A system that manages, monitors, and optimises manufacturing operations in real time. It acts as the link between enterprise-level ERP systems and shop-floor control systems. - ERP (Enterprise Resource Planning)
An enterprise resource planning system that integrates core business processes—including finance, procurement, production, logistics, and human resources—into a single platform. - OPC UA (Open Platform Communications Unified Architecture)
A standard industrial communication protocol designed for the secure and reliable exchange of data between heterogeneous systems such as PLCs, SCADA platforms, and MES solutions. It provides built-in encryption and certificate-based authentication. - MQTT (Message Queuing Telemetry Transport)
A lightweight messaging protocol designed for devices with limited computing resources and unreliable network connections. It is widely used in Industrial IoT (IIoT) applications and sensor telemetry.
Conclusion
You don’t need everything in place to get started. You need to find the first use case where the value is clear, the data exists, and the decision has defined criteria. Everything else gets built from there.
That includes security: deploying in observation mode before handing over real control isn’t a limitation — it’s the right way to build trust, both with the plant floor team and the IT manager. And it includes compliance: the EU AI Act isn’t a late-stage obstacle, it’s a design variable that belongs on the table from the very first sprint.
Sergio Álvarez
At WonderBits, we work with industrial companies from the initial diagnosis — data, integration, architecture — through to production deployment. If you have operational data and you’re ready to make the move to intelligent automation, we can help you map the way forward.
Did you find this useful? Share it with your team. And if you want the strategic context, read our article on what AI agents are and how they are redefining industry.