How Agentic AI finds the hidden, cross-system causes of machine downtime

TL;DR

• The real cause of a stoppage often sits upstream, in an operator process, not in the machine that stopped.
• Your systems of record keep quality, maintenance and downtime in separate silos, so no single dashboard can show the link between them.
• Agentic AI, like Roima's Aura, correlates data across those silos and returns the connection with the evidence attached.
• The result is root cause analysis that surfaces what your team never had the time, or the reason, to compare.

A production line runs the same products on the night shift that it runs during the day, yet the night shift takes 20 per cent longer to change over.

Why is that?

The first instinct is to look at the machine. But the machine is rarely the answer.

The cause usually sits a few steps upstream, in something a person did, skipped or recorded, on a different part of the line and in a different system.

This is just one of the questions that go unanswered on most factory floors, not because the data is missing, but because it is scattered. At Roima we built Aura, an agentic AI toolset, to chase exactly this kind of cross-system question and bring back an answer you can act on.

Walk the floor and the symptom is easy to name.

• A machine stopped
• A changeover ran long
• A batch failed a check.

What is hard to name is the reason, because the reason and the symptom are often two different events in two different places.

Rafael Amaral, CTO and co-founder at TilliT, now part of Roima, describes a pattern he sees again and again: Quality checks all pass, despite everyone knowing that a faulty machine is feeding a bad input into the product.

The check and the cause never meet on the same screen, so the problem persists in plain sight.

The night shift changeover is the same shape of problem. If the equipment is identical to the day shift, the extra 20 per cent is not coming from the equipment. It is coming from how the work is being done, which lives in process and operator data, not in the machine's downtime record. Once you accept that the cause is usually upstream, the real question becomes why nobody can see it.

A system of record is a piece of software that holds one domain's data. The MES tracks production and downtime. The quality system holds the checks. The maintenance log holds the repairs.

Each one is good at its own job, and each one is a silo.

And the teams are siloed too.

As Rafael puts it, the maintenance person worries about whether the machine is running, the quality person worries about the process variable, and neither can see over the fence to the other side. Physically it is one factory making one product. In data terms it is four or five disconnected worlds.

This is why a normal dashboard cannot find an upstream cause – a dashboard reads only one silo.

It can show you your worst downtime reasons, or your quality trend, but it cannot ask whether the two are related, because it was never built to hold both.

So the real value in error management often sits in the join between systems that were never joined. In other words, the golden data is there, but – as Rafael explains – most companies have no shovel to dig it out.

This is where the agentic AI comes into play – specifically the agent swarm.

LISTEN TO THE PODCAST: Agentic AI in Operations Management, from insight to intervention

An agent swarm is a group of AI agents, each assigned its own goal or KPI, investigating the same factory data from different angles. That structure matters here, because a single analyst looking at a single report cannot hold the whole factory in view at once. A swarm can.

What makes Roima’s agentic swarm solution, Aura, different from an earlier generation of factory chatbots is that it writes its own code.

The swarm sets in motion many things at the same time. Multiple agents checking and cross checking data across your production. And once data is gathered, the swarm cross correlates the numbers to arrive a actual issue sources that needs to be solved.

The shear number of correlation analysis and the data agnostic approach only an AI swarm can reliably (and repeatedly) perform, do you get those factory insights that up until now has been invisibly slowing down your production pace.

One analyst Rafael works with had always studied the downtime reason report on its own. With Aura, that report gets correlated with what is happening in cleaning, inspection and lubrication, and in root cause analysis.

“It is hard to do that as a human, especially when you do not know you have to. Show a plant manager the finished link and the reaction is often, Oh my God, I've never seen my factory with this lens,” Rafael says.

Finding the link is only useful if someone acts on it, so the last step matters as much as the analysis. It’s in the habits of acting:

On Monday morning you open a board of insights. Each one carries a short summary written for a manager, a clear recommendation, and all the evidence behind it, down to the queries it ran and how confident it is.

This is what lets a manager audit a conclusion in minutes rather than take it on trust, and it is what turns a surprising correlation into a decision a team will stand behind.

The insight goes into the daily huddle or the weekly meeting, and the outcome is a changed process, a maintenance job, or a new step in a standard operating procedure.

Needless to say, if the insight is stranded in the dashboard without anyone assigned to solving it, well, the issue will not be solved.

None of this invents new information. The signal and data was always in your systems. What was missing was a way to read across them fast enough and widely enough to catch the connection.

That is the shift agentic AI brings to the shop floor, and it is why Rafael calls it untapped value by definition. You read the finding, it makes complete sense, and yet it had never once come to you.

And we’re not saying this is a golden nugget. Unfortunately.

Your machines will keep bumping into errors. They will keep stopping for new and odd reasons. The difference with agentic AI is whether you keep blaming the machine only, or see the cause sitting upstream.

Agentic AI and the insights they provide is value up for the taking. And, as Rafael says, if you are not going after that value, your competitors will.

Frequently asked questions

Q: What is cross-system root cause analysis in operations management?

It is the practice of finding the cause of a problem by comparing data from more than one system, for example linking a machine's downtime record to the operator process and quality data around it, rather than looking at each in isolation.

Q: How does Aura find a correlation a human would miss?

It writes and runs its own queries across the different data sets in parallel, testing connections a person would rarely think to compare, and returns the ones that hold up with the supporting evidence attached.

Q: Does agentic AI in root cause analysis work on any production plant?

Yes. The analysis is driven by goals and KPIs such as OEE, waste and pass or fail rates, so it adapts to the data a given plant holds rather than to a fixed template.

Q: How is the factory data kept secure?

Because the AI is given real analytical power, it runs inside a blocked and protected environment. The data is brought into that environment, the analysis runs there, and the results come out.