Agentic AI
Agentic AI in production: automation within limits you define
When the same fault keeps tying up several systems and several people, an agentic system can take over the recurring steps. SABO IT develops these AI agents for manufacturing, tailored to your use case and your system landscape: above PLC and SCADA, within limits you define.
Where an agentic system takes over the work
An agentic system works independently toward a goal within defined limits. It acts on a trigger, checks the connected systems, and carries out approved steps. Anything outside those limits goes to your experts.
Request a workshop
From alert to approved action: how an agent investigates a case and when it hands over to a person.
Where it runs
Above PLC and SCADA
Integration happens through defined interfaces, not by changing your core systems.
Operating model
On-premises if you want it
You decide where the process data is stored and which permissions apply to each component.
Limits
You define them
Your experts approve any step that goes beyond the limits you set.
Evidence
A log for every step
Each step is written to a decision log — on record, not from memory.
What you end up with
Recurring steps run without the back-and-forth
The path from alert to information is the same every time. The system follows it as soon as the trigger comes in and puts the result where your people are already looking.
Every step is documented
Which source was read, which tool was called, which action was carried out — it’s all in the log. Not in the memory of whoever was on shift.
The limits stay yours
You decide which actions run without approval, how much uncertainty it takes before the system escalates, and whether it has write access at all. You can change this at any time without rebuilding the application.
Your existing systems stay the systems of record: there’s no second interface for the shift crew to learn, and the scope starts narrow and only grows once it has proven itself.
How an agentic system works
Three steps, and one point where the system hands over.
At the limit you define, the path splits.
01
Trigger
A trigger starts the process, such as an alert from the plant.
02
Gather and plan
The system gathers information from previous cases, manuals, and the current machine state, then plans the next steps.
03
Execute or hand over
Within the limits you define, the system carries out the planned steps and logs them. Everything else goes to your experts.
Where AI agents pay off in manufacturing
Five areas where the same work keeps coming back.
Tracking down faults on machines and lines
Typical starting point
Beverage filling, injection molding
Investigating quality deviations
Typical starting point
SMT assembly, packaging
Prioritizing maintenance requests
Typical starting point
Robot and CNC cells, conveyor systems
Handling cross-system exceptions
Typical starting point
MES, ERP, and SCADA integration
Running service diagnostics
Typical starting point
Field devices, remote monitoring
How we limit the risks of AI in production
Autonomy
What SABO IT builds in
Tiered actions, approval steps, safety interlocks
What you decide
Which actions run without approval
Reasoning
What SABO IT safeguards
Cross-check against the plant state, escalation when unclear
What you decide
How much uncertainty it takes before the system escalates
Control
What SABO IT safeguards
Above PLC and SCADA, no interference with the safety function, not real-time capable. Read access via defined interfaces, write access only after approval in the project
What you decide
Whether there is write access, and how far it goes
Data and permissions
What SABO IT builds in
Cross-checking multiple sources before every action, role and permission model, least privilege per component, on-premises deployment
What you decide
Sources, operating model, permissions
Traceability
What SABO IT safeguards
Decision logs, result checks
What you decide
How long logs are retained
Concept example
Agentic troubleshooting on a filling line
Not a customer installation, but a fully worked-out concept: how an agent would work on a beverage filling line alongside the existing control systems and knowledge sources.
Today, alarms and drops in throughput are investigated manually. Whoever is on duty reads telemetry and searches the manuals for the last similar case. With an agent, the system opens a case on its own when the trigger comes in, reads the same sources, and carries out whatever is on the approved list. Anything else, it proposes.
If work on the hardware is needed or the situation stays unclear, the case goes to the technician, along with everything the agent has checked so far.
Beverage filling line — at a glance
Primary goal
Fewer manual steps to a diagnosis
Scope
Alarms, quality, throughput
Integration
PLC, MES/SCADA, manuals
Trigger
Alarm or deviation reported by the line
Investigation
Telemetry and previous cases, in one pass
Action
Approved steps run on their own, everything else comes as a proposal
Learning
Outcome and decision go back into the knowledge base
The system doesn’t invent new machine behavior
The system works from a list of safe commands that you define: controlled resets and parameter changes within permitted limits. Work on the hardware stays with the technician, who gets clearer information on likely causes.
From one line to a reusable architecture
Answering four questions makes the setup transferable to the next machine: which problems matter, what information is available, which triggers apply, and which actions are permitted.
Questions and answers
What is agentic AI, and where does it make sense in industry?
Agentic AI refers to AI systems that work independently toward a goal within defined limits. Such a system acts on a trigger, checks the connected systems, and carries out approved steps. It makes sense where recurring faults are investigated across several systems. SABO IT develops these systems in customer projects.
Will an agentic system replace our maintenance team?
No. An agentic system from SABO IT takes over recurring steps and documents them. Everything outside the defined limits stays with your experts, and so does any work on the hardware.
Does the AI interfere with machine control?
Read access, yes; write access only if you approve it. Agentic systems from SABO IT operate above PLC and SCADA, do not interfere with safety functions, and are not real-time systems; cycle times stay in the controller.
Does our process data stay inside our own network?
You decide. Agentic systems from SABO IT can run on-premises, with a role and permission model, and least privilege for each component.
The first step is a workshop
Bring a case that keeps coming up at your company. In the workshop, we clarify the use case, the systems to be connected, the actions the system may carry out on its own, and how you will measure success. We start with a narrow scope and expand only once it has proven itself, step by step and together with the people who will use it.
Get in touch
