MES software
MES software exists because the plan and the shop floor disagree, and by the time the disagreement reaches a report it is a week old. The category is mature and the products are good. What decides whether one works for you is not the product — it is whether the floor tells it the truth.
The gap it is bought to close
Planning knows what should be running. The floor knows what is running. Between them sits a shift, a scrap rate, a machine that stopped for forty minutes, and an operator who wrote it on paper because the terminal is across the building. Every number downstream inherits that gap.
A manufacturing execution system is the instrument for closing it: what is on each machine, in what state, against what order, right now. Where it fails is rarely the software. It fails where the data enters — a scan nobody does, a reason code chosen at random because the real reason is not on the list, a downtime logged at the end of the shift from memory.
Buy the system, build the honesty
- Buy the MES. This is a mature, competitive category with decades of domain knowledge inside it. Building one is a bad idea and I would talk you out of it.
- Build what makes the entry true. Reason codes that match what actually happens, capture that takes seconds rather than minutes, and a check that refuses a shift record that cannot be right — a run longer than the shift, an output higher than the machine's rate.
- Reconcile the two versions. Where the MES and the ERP disagree about the same order, the difference gets surfaced the same day instead of being averaged into a monthly report.
- Read the free text. Operators write what happened in a comment box because no code fits. That text is the most honest data in the building, and it is the part no report ever reads.
- Watch for silence. A line that stops sending data looks identical to a line that is not running. Telling those apart automatically is worth more than most dashboards.
What it must not do on its own
- Change a production schedule. It can show that the schedule is impossible. Moving it is a decision with customers and materials attached, and it belongs to a planner.
- Overwrite an operator's entry. If the machine data and the person disagree, both are kept and the disagreement is raised. Silently trusting the sensor is how you lose the one person who knows why the reading is wrong.
- Judge an operator. This measures a process, not a person, and the moment it is used the other way the data quality collapses within a fortnight. That is not a prediction, it is what always happens.
- Decide a quality hold. It flags what looks out of tolerance and stops there. Releasing or holding a batch is a signature with a name on it.
The floor is not ready if
- You have one line and the supervisor can see all of it from where they stand. The instrument is for when the floor is bigger than the view.
- Nobody has agreed what a unit of good output is. If quality and production count the same run differently, the system will faithfully record both and settle nothing.
- The floor has no network and no terminals. That is the first project, it is ordinary infrastructure work, and it is cheaper than anything on this page.
Where the pieces of this already run
- Data platform — the same discipline at the layer underneath: every figure defined once, checked as it arrives, and traceable back to the row it came from. On a shop floor that is the whole argument.
- Systems watchdog — the difference between a line that is idle and a line that has stopped reporting, caught automatically — the failure that looks like nothing at all.
- Team briefing bot — the morning number in the chat instead of in a report nobody opens, which is what a floor manager actually reads.
What plant managers ask first
Should we buy MES software or build one?
Buy it. The category is mature, competitive and full of domain knowledge that takes decades to accumulate. What is worth building is the layer that makes the data entering it honest, and the reconciliation between it and whatever else claims to know the same numbers.
Our operators do not use the terminals. Does this fix that?
Only if the reason is friction. If entry takes ninety seconds and the shift is measured in seconds, no amount of software changes the behaviour — the capture has to get faster. If the reason is that no reason code fits what actually happens, the fix is the code list, not the screen.
Can it work with the ERP we already have?
That is the question to ask early. If the ERP can be written into and read from, the work is mapping your order and material records. If it cannot, you are buying a screen that somebody copies numbers off, and the copying is the job you were trying to remove.
What does it actually change in the first month?
Usually one thing: you find out how far the recorded state is from the real one. That number is uncomfortable and it is the point. Everything after it — scheduling, costing, quality — was being calculated on the gap.
Is AI involved, or is this ordinary software?
Most of it is ordinary software and should be. The machine-learning part is narrow: reading the free-text comments operators write when no code fits, and noticing that a pattern of readings does not look like normal running. Those two do not need a large model to be useful, and the rest of it should not pretend to.
Measure the gap before buying anything
Take one line, one shift. Write down what the system says was produced and what the supervisor says was produced, and put the two numbers next to each other. If they match, you do not need this. If they do not, the size of the gap is the size of the problem, and that sum costs you an afternoon. The first conversation is an hour and it is free.
Book that hour →