When the machine talks to the AI: Who is in charge?
AI can make sense of the data from your machines. But what happens if the data is wrong? How to stay in control when AI comes into production.
Picture a production hall where the machines send data around the clock: temperature, vibration, pressure and operating hours. An AI assistant reads along and helps the operations manager understand what is happening. One day a machine reports heavy vibration, and the AI recommends stopping it immediately.
That sounds like exactly what AI should be able to do. But what if the numbers were wrong? A faulty sensor, an error in an update or, at worst, someone deliberately sending false data. The AI cannot tell the difference. It reads what it is given and answers as best it can.
This is not an argument against AI in production. It is an argument for thinking about who is in charge when the machine talks to the AI.
Data is not true just because it comes from a machine
Many companies trust data because it comes from their own systems. That is understandable, but it is not enough. An AI treats everything it reads as something to act on, including a comment in a field or a value far outside the normal range.
So data has to be checked before the AI gets it. Does it come from the machine it claims to come from? Are the values within what is physically possible? Does it match the other sensors? These are simple checks, but they catch many errors before they turn into decisions.
Give the AI what it needs, and no more
The more data an AI gets, the more places something can go wrong. When an employee asks a question, the AI should only have access to the data needed for the answer, and only the data that employee is allowed to see.
That sounds technical, but it is really a leadership decision: Who may see what, and what may the AI use? It has to be decided before the solution is built, not afterwards.
The AI recommends. People decide.
The most important principle is simple: the AI may suggest stopping a machine, but it must not press the button itself. A recommendation must be explainable by the data it is based on, and it must be checked against the fixed limits and safety rules you already have.
After that, an employee with the responsibility makes the decision. The AI is good at creating an overview and pointing out the unusual. People and the existing safety systems are good at taking responsibility for what happens next.
The AI should make your people smarter, not take decisions away from them.
Plan for the day the AI does not work
All systems go down now and then. The network drops, a vendor changes something, or the AI gives an answer that makes no sense. So production must be able to continue without the AI, with the alarms and safety rules you have today.
It must also be easy to switch the AI off without stopping operations, and everything must be logged: which data the AI saw, what it recommended and who approved it. Then you can always find out what happened, and learn from it.
Four questions for management
- Do we know which data our AI solutions read, and where it comes from?
- Is it always a person who approves the decisions that matter?
- Can production keep running if the AI is down for a week?
- Can we change vendor or model without rebuilding everything?
How we approach it at BCT
When we build AI for industry, we start with a Shadow Walk, where we follow the work as it actually happens. That way we know which decisions the AI should help with, and which should stay with your people. We build the solution so data is checked, people have the final say, and operations can continue without the AI. We do it at a fixed price, and you own the code, so you stay in control, including over your vendors.
Want to hear more about using AI safely in your production? Contact us at info@blackcapitaltechnology.com or (+45) 39 39 42 43.
Want to see what it could look like for you?
We are happy to have a no-obligation talk about your day-to-day work and where AI and software can make a difference.
