AI can’t make a good decision about a production line it can’t see. That’s the part of the AI-in-manufacturing conversation that gets skipped most often: every predictive maintenance model, every anomaly detector, every AI copilot that promises to tell a plant manager what’s about to go wrong needs a constant stream of structured, real-time floor data to work from. Without an Andon system logging events and an MES structuring them, there’s nothing for AI to learn from — just a factory generating data that never gets organized in the first place.
That’s the honest starting point for any manufacturer evaluating AI right now: the AI layer is only as good as the data layer underneath it, and most plants haven’t built the data layer yet.
The AI hype vs. the shop-floor reality
Manufacturing media has spent the last two years full of AI predictions — autonomous factories, self-optimizing lines, copilots that talk you through a changeover. Most of that is still ahead of where the average plant actually is. What’s real, right now, is narrower and more useful:
- Predictive maintenance models that flag a machine likely to fail before it does, based on historical downtime and sensor patterns
- Anomaly detection that catches a quality or performance deviation faster than a human reviewing a shift report would
- Natural-language reporting that lets a supervisor ask “why was Line 3 down yesterday” and get an answer instead of pulling a spreadsheet
Every one of these depends on the same thing: clean, timestamped, categorized data about what actually happened on the floor. That data doesn’t come from AI. It comes from the systems that were supposed to be capturing it all along.
Why AI is only as good as the data feeding it
This is the manufacturing version of “garbage in, garbage out,” and it’s worth being direct about it: a plant that logs downtime in a notebook, or reconstructs shift reports from memory at the end of the day, has nothing for an AI model to learn from. Predictive maintenance can’t predict a failure pattern it’s never seen recorded. Anomaly detection can’t flag a deviation from a baseline that was never established.
The plants getting real value from AI right now aren’t the ones with the most sophisticated models — they’re the ones that already had disciplined, automated data capture in place before they started layering AI on top. That’s the same foundation covered in manufacturing execution systems (MES) in manufacturing: execution control that turns raw data into disciplined production, which turns out to be exactly the prerequisite AI needs too.
Where AI is already showing up on the floor
Three applications are past the pilot stage in real manufacturing environments today:
- Predictive maintenance. Instead of fixed maintenance schedules or reactive repairs, AI models trained on historical machine and downtime data can flag equipment likely to fail in the next days or weeks, prioritized by risk.
- Downtime pattern recognition. Rather than a supervisor manually noticing that Line 3 keeps stopping every Tuesday afternoon, AI can surface that pattern automatically from categorized (capcode-tagged) downtime data, and flag it before it becomes a trend.
- Automated, conversational reporting. Plant managers increasingly want to ask a question in plain language — “what caused the most downtime this week” — and get a direct answer instead of building a report by hand.
Notice what all three have in common: none of them work without a system already capturing structured floor data in real time.
Andon + AI: from reactive alerts to predictive ones
Andon systems generate exactly the kind of event data AI models need — timestamped stoppages, categorized by cause, tied to a specific machine and response time. Today, that data mostly powers reactive alerting: something breaks, a light turns on, someone responds. The next step, already underway at the leading edge of the industry, is using that same historical alert data to predict the next stoppage before it happens, shifting Andon from a reactive signal system to a predictive one. The 10–12% downtime reduction VersaCall customers see from Andon alerting today is the reactive-layer number; the predictive layer is where that number improves further.
MES + AI: turning execution data into recommendations
An MES’s job has always been to structure floor data — which is precisely what makes it the natural foundation for AI. Once execution data is organized by work order, machine, shift, and cause, it becomes possible to layer recommendation models on top: suggesting a changeover sequence that minimizes downtime, or flagging a work order at risk of missing its deadline based on current line performance. This is the layer where “disciplined production” (structured, reliable execution data) turns into “intelligent production” (that data actively informing decisions).
OEE + AI: forecasting performance, not just reporting it
OEE systems today report availability, performance, and quality after the fact — automatically, in VersaCall’s case, but still retrospectively. AI’s contribution here is forward-looking: forecasting where OEE is headed based on current trends, and flagging which of the three components (availability, performance, or quality) is most at risk before it drags the overall number down. For manufacturers already seeing an 81% improvement in on-time delivery from automated OEE reporting, AI forecasting is the difference between reporting a problem and anticipating one.
What manufacturers should do now
For most plants, the right next step isn’t buying an AI platform — it’s making sure the data foundation is actually there to support one. In practice, that means:
- Audit your current data capture. Is downtime logged automatically and in real time, or reconstructed manually at the end of a shift?
- Standardize your categorization. AI models need consistent capcodes and reason structures to find patterns; inconsistent tagging is one of the most common reasons manufacturing AI pilots fail.
- Get Andon, MES, and OEE talking to each other first. A manufacturing communication system that already connects alerting, execution data, and reporting is the platform AI gets built on top of — not a separate project.
- Start with one high-value use case, like predictive maintenance on your most failure-prone equipment, rather than a facility-wide AI rollout.
FAQs
Do I need AI to get value from an Andon or MES system?
No. Andon and MES systems deliver measurable downtime reduction and efficiency gains on their own — that’s been true for years, independent of AI. AI adds a predictive layer on top of a foundation that already needs to exist regardless.
What's the biggest reason manufacturing AI projects fail?
Incomplete or inconsistent floor data. AI models trained on gaps, manual reconstructions, or inconsistent downtime categorization produce unreliable predictions, which is why the data-capture foundation matters more than the AI model itself.
What's the first AI use case most manufacturers should try?
Predictive maintenance on the equipment responsible for the most unplanned downtime historically. It has the clearest ROI case and the most historical data (from Andon and OEE logs) to train on.
How does real-time floor data relate to AI search tools like ChatGPT or Gemini?
The same structured, well-documented information that trains an internal predictive maintenance model is what makes a manufacturer’s expertise citable by public AI tools — a plant (or its technology partner) that publishes clear, data-backed explanations of how its systems work becomes a more useful, more frequently cited source than one that doesn’t.
Build the data foundation first
Before evaluating AI vendors, it’s worth confirming your Andon, MES, and OEE data actually talks to itself. Schedule a free demo to see what a connected, real-time data foundation looks like, or explore case studies from manufacturers who built theirs before layering on the next generation of tools.
