MES gets sold as a category, but it is really a job description: whatever software sits between the plan and the machine has to answer "what should be running right now, is it running correctly, and what actually happened." Different vendors package that job differently. What they package rarely varies.
What an MES actually does
Strip away the vendor branding and an MES is doing five things, and you can trace each one to a plant that runs it.
| Function | What it means on the floor | Corpus example |
|---|---|---|
| Work order execution | Dispatches the right job to the right line and tracks it to completion | Daimler India DICV runs multi-brand commercial vehicle production on FactoryTalk MES, sequencing different vehicle brands down shared lines |
| Routing enforcement | Blocks a unit from moving to the next step if the prior step wasn't done, or wasn't done correctly | Mullins Food Products eliminated batch deviations almost entirely once the MES enforced the recipe instead of trusting operators to follow the paper version |
| Genealogy / traceability | Links every finished unit back to its lot, machine, operator and timestamp | An anonymous tool manufacturer cut defect investigation time from 5 days to 30 minutes, avoiding $266,000 in rework by finding the source fast enough to contain it |
| Electronic work instructions | Delivers the current version of the build steps to the operator, not the version in the binder on the wall | A prosthetic-arm startup cut assembly time from 3-5 days to 1-2 days by replacing paper instructions with guided digital ones |
| Data collection | Records a transaction for every unit and every operation, not a summary at shift end | BIC's Tunisian plant runs 100,000+ finished-good transactions per shift on Plex MES |
None of these is optional in a working system. A platform that does routing and data collection but has no real genealogy isn't a lighter MES, it's a different tool wearing the name.
Where MES sits
In the ISA-95 model vendors cite constantly, MES occupies Level 3: above the control system that actually runs the equipment (Levels 0-2) and below the ERP that plans and pays for the business (Level 4). That single sentence is as much of the standard as you need for a working mental model. What actually separates MES from ERP in practice, and where the two disagree expensively, is its own subject: see MES vs ERP.
Who actually needs one
Genealogy is the tell. If a regulator, a customer, or a recall could someday ask "which machine, which operator, which lot" and the honest answer is "we'd have to reconstruct it," you need an MES. That's why the deployments cluster so heavily in pharmaceuticals — Dong-A ST built a cGMP-compliant smart factory around its MES in Korea — and in automotive and food, where a bad lot has to be traceable to an exact production window.
Inventory accuracy is the other honest signal. Newman Technology hit 98% inventory accuracy and cut inventory cost 25% after moving from manual counts to MES-tracked transactions; Arrow Fabricated Tubing landed in the same 98-99% range. Both were running spreadsheets and count sheets before. If your plant runs one product, one routing, one shift, and nobody has ever asked where a specific unit came from, an MES is solving a problem you don't have. Buy one anyway and you've bought a very expensive way to enforce a process on a floor that didn't need enforcing.
What actually breaks these projects
An MES doesn't design your process. It encodes the one you already have. That is the entire mechanism, and it's why implementations fail in a specific, predictable way: if the process on the floor is undocumented, or three shifts each run it slightly differently, the MES forces someone to pick one version and write it down before a single line of configuration happens. Plants that skip that step end up "configuring" the software during go-live, discovering the disagreements live, and burning the project timeline on a fight that should have happened in a conference room. The software isn't what's expensive. Reconciling a process nobody had actually agreed on is.
This is also why the traceability payoff is so large relative to the setup cost. Reconstructing what happened to a defective batch by hand, from paper travelers and operator memory, is a multi-day forensic exercise. An MES makes it a query. That's the mechanism behind the 5-day-to-30-minute result above: the data was already being captured as a byproduct of running the line, so answering the question was free at the point someone needed the answer.
How AI changes the picture
MES was built to record what happened. The newer layer is software that looks across everything it recorded and finds what a person reading one shift report would miss: a die wearing faster on second shift, a supplier lot correlating with a defect spike three operations downstream. Across the 416 process-optimization deployments in our corpus that this kind of MES-fed analysis feeds into, productivity and OEE gains land at a median 25% (3%-80%, n=40) and reported annual savings at a median $600K ($34K-$100M, n=10). Those numbers come from the data an MES was already collecting, pointed at a model instead of a filing cabinet.
Where to go next
- For the sharper version of the ERP question, MES vs ERP covers where the two systems actually disagree and what the integration between them costs.
- The process optimization use cases and benchmarks pages break down outcomes across the full 416-deployment corpus.
- Vendors that show up repeatedly in MES deployments are ranked on the process optimization software hub.
- Special Products Manufacturing cut inventory 30% after moving onto an MES, one of the cleaner examples of what the data-collection layer alone is worth.
AI Roadmap — Eight workflows in your plant where AI would pay back this quarter. We know which. 12 pages, in 72 hours.