PLM, MES, ERP: these industrial software systems drive multi-million euro decisions without any reliability metrics

Your industrial software (PLM, MES, ERP) drives multi-million euro decisions every day. They rely on AI predictions. None of them tell you how reliable they are. This blind spot is expensive. Here is why, and how to close it.
The problem is not that AI is sometimes wrong. It is that it does not tell you when it is wrong. A model gives you a number, never the confidence you can place in it. And based on that number, a design team, a line operator, or a planner will commit capital, materials, and lead times.
The blind spot: a prediction without a confidence interval is a gamble
A point forecast is a single value, delivered without a confidence interval, and therefore without any indication of the risk it carries. Two forecasts of "100 units" can hide, one a deviation of plus or minus 5, the other of plus or minus 80. Same figure. Radically different risk.
Without the interval, the decision-maker treats them the same way. The literature in the field is clear: a point forecast "says nothing about the associated uncertainty" (Hyndman & Athanasopoulos, Forecasting: Principles and Practice, 2021).
In machine learning, the trap is worse. A model can be confident and wrong. A good average accuracy (RMSE, MAPE) hides precisely the predictions on which it is heavily wrong. Faced with an unprecedented case or a gradual drift (model drift), it continues to predict with the same assurance. It does not signal that it is going out of its domain of validity, a situation known as out-of-distribution.
What is missing from each inference is a per-prediction confidence metric. Not a reassuring average. A signal, here and now, on this precise output. Let's look at the cost, layer by layer.
PLM: the design decision that costs ×10 at industrialization
The scenario often recurs. The PLM's AI module notes an "optimal design, controlled cost". The team validates it. Six months later, during industrialization, the defect appears. The change now costs ten times more. No one saw that the prediction, that day, was fragile.
There is nothing anecdotal about this. Engineers know the rule. The cost of correcting a design change is multiplied by 5 to 10 at each stage of the product cycle. After release to production, a corrected defect can cost up to 100 times its initial price (Boehm, 1976; rule of 10, CENIT). A bad decision frozen upstream on an unreliable prediction propagates its cost throughout the downstream: purchasing, methods, tooling, production.
Generative design compounds the risk. The outputs "look visually plausible but offer no guarantee of functional correctness." And "you cannot trust the confidence score of a generative model" (colabsoftware, 2024). Translation: the software gives you an option, never its reliability.
With a per-prediction reliability attached to each PLM output, a design is frozen only if confidence is high. Predictions with wide intervals go to human review, before costing ×10 downstream. This is a reliability layer placed over the product lifecycle, not just another control.
MES: alarm fatigue, or the day the operator turns off the sound
Second layer, second story. The MES aggregates production data and IIoT sensors, which track vibration, temperature, and wear. The goal is to anticipate breakdowns, beyond classic preventive maintenance. The system generates continuous AI alerts from this. The operator receives so many, almost all of them false alerts, that they end up ignoring them. The day the alert is real, it gets drowned in the noise. And a stopped line costs tens of thousands of euros per minute.
This mechanism is documented and standardized. Beyond 6 alarms per operator per hour, the system is already degraded. Beyond 30 per hour, it is severely deficient. A human saturates around one alarm per minute (standard EEMUA 191). Worse: 10 to 20 alarms often generate 50% to 80% of the total volume. Resolving them cuts the flow by 60% to 80% (exida). A few bad actors therefore drown out everything else.
The price of a missed true alert, however, is catastrophic. Unplanned downtime costs about $22,000 per minute in the automotive sector, up to $50,000, or $2.3M per hour (Siemens, True Cost of Downtime, 2024). These incidents impact equipment availability, extend downtime, and inflate maintenance costs. They also distort your indicators, from MTBF to MTTR, up to OEE. During the Texas City accident, 3,712 alarms during the event delayed the critical response. The overwhelmed operator did not see the tower filling (investigation US CSB, 2007). The alarm flood was identified as a root cause.
A reliability layer on MES alerts distinguishes a reliable alert from one that is not. Only high-confidence alerts are escalated. The rest is prioritized. We signal model drift before they become noisy. The operator gets a usable flow back, and the real alert stops being drowned out.
ERP: the demand forecast we believe, and which fills the warehouse
Third layer. The ERP outputs a demand forecast by SKU. The planner must decide how much to produce and stock. If they believe a false forecast, they either drown in stock or run out. And nothing tells them how reliable that specific forecast is.
For lack of this signal, two costly reflexes take hold. They over-dimension safety stock "just in case." Or they adjust the system manually, even though nearly 40% of these manual adjustments degrade accuracy (McKinsey, 2024). Distrust is massive: more than 50% of supply chain teams cite explainability as a barrier to adopting models (the return to "Shadow Excel", McKinsey, 2024).
Both errors carry a price. Too high: holding inventory costs 20% to 30% of its value per year in carrying costs (APICS), not counting obsolescence and markdowns. In 2022, US retail inventories surged by $78 billion; Target's operating income fell by 43% in one quarter, forecasts revised twice (Modern Retail, 2022; CNBC, 2022). Too low: stockouts and lost sales. And the bullwhip effect amplifies every error upstream, with an estimated extra cost of 12% to 25% per company in the chain.
The stakes are not minor. Better forecast reliability can reduce error by 20% to 50%, stockouts and lost sales by up to 65%, and inventory by 20% to 50% (McKinsey, 2024). A confidence interval per demand forecast tells the planner where to act firmly and where to cover their risk. It acts on confidence, not fear.
The Total Cost of Reliability (TCR): what an unreliable prediction really costs
Taken individually, these three layers seem to pose three distinct problems. In reality, they share the same cause: the lack of a per-prediction confidence metric. To make it visible, we use a purchasing reading grid.
The Total Cost of Reliability, or TCR, refers to the real cost, accumulated over about 24 months, of an AI on which no one can decide with confidence. It is not a single invoice. It is a silent addition, carried without being measured by any owner of "AI-augmented" PLM, MES, or ERP.
Cost Category (over ~24 months)
Layer
Sourced Benchmark
Re-design / late change
PLM
cost ×5 to ×10 per stage; up to ×100 after release to production (Boehm, 1976)
Overproduction / tied-up stock
ERP
20% to 30% of inventory value per year in carrying costs (APICS)
Obsolescence and markdowns
ERP
2022 case: Target operating income down 43% over one quarter (CNBC, 2022)
Stockouts and lost sales
ERP
reducible by up to 65% with better forecasting (McKinsey, 2024)
Bullwhip effect
ERP
estimated extra cost of 12% to 25% per company in the chain
Operator time lost in nuisance
MES
10 to 20 alarms = 50% to 80% of volume (exida)
Missed real breakdown / unplanned downtime
MES
~22,000 $/min in automotive; 2.3 M$/h (Siemens, 2024)
Distrust and return to manual spreadsheets
ERP
~40% of manual adjustments degrade accuracy (McKinsey, 2024)
The logic of TCR is simple: do not try to treat each symptom separately. Treat the common cause. We do not publish a total in euros, because it depends on your volume, your margins, and your sector. The calculation, however, is done on your line, with your figures.
Per-prediction reliability before decision, not aggregated monitoring after the fact
This is where the real distinction lies. Monitoring and drift detection tools measure model health in aggregate and after the fact. They tell you "your model drifted last month". Useful for the review. But the production, design, or procurement decision has already been made, based on the bad prediction.
Aggregated monitoring looks at an average over a past window. Post-mortem analysis explains an incident after it has cost. Neither qualifies the forecast that the planner is about to follow, now.
Per-prediction reliability attaches a confidence indicator to each individual inference, in real time, before the decision. On a time series, this is decisive: it is not the average accuracy of the model over 30 days that counts, it is the reliability of today's forecast. In other words: others monitor the model. We qualify the prediction.
A guide to avoid confusing the two:
Monitoring / drift: aggregated, a posteriori, guides model maintenance. Answer to "is the model doing well, on average?"
Per-prediction reliability: by prediction, in real time, guides business decisions. Answer to "can I follow this precise output, now?"
Let's be clear on the scope: a reliability layer contributes to the quality of the decision. It does not guarantee a commercial result. It reduces the share of decisions made blindly.
TrustalAI Predictive: a plug-and-play reliability layer on your software
TrustalAI Predictive is the per-prediction reliability block for your predictive models, including your deep learning models. This reliability layer adds a 95% confidence interval to each forecast, in real-time. It detects model drift, signals anomalies and out-of-distribution situations. The decision-maker finally receives a figure, its confidence score, and its level of risk, before acting.
The product metrics are validated and published: −81% errors and −84% false positives, for 20 ms latency. The integration is plug-and-play and black-box compatible: no access to your model, no retraining, no modification of your process. The calculation runs in real time, in edge under 20 ms or in cloud under 80 ms (so <100 ms). The product is at TRL9, validated in a real operational environment.
The same mechanics proved themselves on the vision side, on a multi-sensor perception case (PoC VEDECOM): 83% reduction in critical false positives, without retraining the client model. Proof that the reliability measurement holds on real data, not just in the lab. The principle transposes to the management of predictive model drift.
What you gain, layer by layer
PLM: safer design decisions, fewer costly re-design cycles.
MES: end of online alarm fatigue, more reliable predictive maintenance, faster response to real failures.
ERP and supply chain: less overproduction, fewer stockouts, better resource allocation.
The deployment model follows two paths. Direct, where a CTO or Head of Digital Manufacturing mandates the layer in their PLM, MES, or ERP environment. In B2B2B, the software vendor embeds TrustalAI Predictive as an OEM module or SDK. For this vendor, reliability becomes a product argument: its predictions come out with a confidence interval, its software knows when it doesn't know. For the system integrator delivering the line, it is leverage for their obligation of result: a defensible decision, even when the model remains a black box.
And what about compliance?
The perspective here is financial before it is regulatory. Demand forecasting and PLM, MES, or ERP planning do not generally fall under the high-risk systems of Annex III of the EU AI Act (Regulation EU 2024/1689). The exception targets AI as a safety component of critical infrastructure, energy or water for example, relevant for a smart grid, not for a classic demand forecast.
Beyond the EU AI Act and the Machinery Directive, the issue therefore remains business risk. That said, documenting the reliability of each decision, with timestamped logs and traceability, also prepares the ground for compliance and human supervision. The proof of reliability serves both causes at once.
FAQ: industrial software and per-prediction reliability
What is a per-prediction reliability metric?
A per-prediction reliability metric is an indicator attached to each output of the model, which indicates whether this specific prediction deserves to be followed. Concretely, TrustalAI Predictive adds a 95% confidence interval to each forecast, in real time. The decision-maker no longer receives just a figure, but a figure and its confidence level, before acting. It is this signal that transforms a raw prediction into a manageable decision.
Why is a demand forecast without a confidence interval a problem?
Because a single value says nothing about the risk it carries. Two identical forecasts can hide very different uncertainties, and the planner treats them the same. Result: they over-dimension stock "just in case," or adjust the system manually, even though nearly 40% of these adjustments degrade accuracy (McKinsey, 2024). The confidence interval finally tells them where to act firmly and where to cover.
Does TrustalAI Predictive replace my PLM, my MES, or my AI model?
No. TrustalAI Predictive does not replace your software or your models. It is a plug-and-play reliability layer, black-box compatible, which is added on top of the existing system without retraining or accessing your intellectual property. It qualifies each prediction in real-time, in edge under 20 ms or in cloud under 80 ms. A vendor can also embed it as an OEM module in their own software, via a B2B2B channel.
Is a demand forecasting software a high-risk system under the EU AI Act?
In general, no. Planning and demand forecasting in a PLM, MES, or ERP are not listed in Annex III of the EU AI Act. The exception concerns AI used as a safety component of critical infrastructure, such as an energy or water grid. For most Industrial Fabric uses, the decisive angle is financial and operational risk, not compliance. Any specific date must be validated on a case-by-case basis.
Take Action
You only manage well what you measure. Today, your industrial software drives multi-million euro decisions on predictions whose reliability no one measures.
Change this starting point. Quantify the cost of unreliability in your PLM, MES, and ERP forecasts. Request a 2-week PoC. On your line, with your data, without retraining your models. We apply the reliability layer, you decide on solid ground.