Predictive Maintenance: The Model Was Fine and the Economics Weren’t
An automated warehousing operator (1,100 conveyor and sorter assets, ~11 significant failures a month at ~€18K each) had board approval for a €900K failure-prediction and auto-intervention system. Envion’s model reached 0.79 AUC — respectable — but the base rate destroyed the business case: 0.03% positive rate per asset-day meant ~248 alerts a month to catch 7.7 failures, ~240 of them false, at a 97% false-alarm rate technicians stop investigating within weeks. The verdict: no to the alert system; yes to a €70K ranked inspection list that reorders the existing 40-asset weekly inspection queue by risk. Twelve months on: failures down to 6.5/month, annual disruption cost €2.4M → €1.4M, €830K of capital not committed.
The challenge
Unplanned conveyor and sorter failures caused an average of 11 significant disruptions a month, each costing roughly €18K in delayed throughput, expedited freight and overtime. The client had two years of sensor telemetry — vibration, motor current, temperature, cycle counts — and a maintenance management system with work order history. They wanted a model to predict failures and trigger preventive intervention automatically.
The board had approved €900K. Envion's engagement was the last check before the build.
What makes this case interesting: the model worked. Working from the telemetry with a 72-hour prediction horizon, it reached 0.79 AUC on held-out data — not spectacular, entirely respectable, and better than the client expected. Then came the arithmetic that AUC hides.
Decision path
The base rate destroyed it. Significant failures ran at roughly 11 per month across 1,100 assets — a positive rate of about 0.03% per asset-day. At the threshold giving 70% recall — catching seven of every ten failures, which sounds good — precision was 3.1%.
In operational terms: to catch 7.7 failures a month, the system would generate around 248 alerts, roughly 240 of them false. Each alert meant sending a technician to inspect an asset — about €140 in labour and, for a third of assets, a short planned stoppage. Monthly alert cost: ~€35K against a generous €139K of prevented-failure value.
Alert fatigue is not a soft factor: at a 97% false alarm rate, technicians stop investigating properly within weeks, and the realistic prevention rate is maybe 40% after the first quarter — taking value to ~€56K against €35K of cost, before paying for the build, sensor upgrades, or the two FTEs to maintain it. Cross-referencing the model's positives against maintenance records showed ~30% coincided with assets already flagged by routine inspection — incremental information smaller than raw performance suggested. And the failures it predicted best were the cheap ones: gradual, well-signalled bearing degradation, while the expensive disruptions — sorter control failures, drive electronics — were abrupt and found barely better than chance. That inverse relationship is common, and fatal to the business case.
Envion contribution
Envion recommended against the €900K automated system — and was specific that this was not a "come back when you have more data" answer. More data would improve the model somewhat; it would not change the base rate, and the base rate was the problem.
The alternative proposal: a ranked inspection list, not an alert system. Same model, different deployment — technicians already perform routine inspection rounds with a fixed capacity of roughly 40 assets a week; instead of alerting, the model orders that existing queue by risk. No new labour cost, no false alarm concept, because nobody is being interrupted. The model doesn't have to be precise; it only has to beat the current inspection order, which was alphabetical by asset ID. Cost: about €70K.
Plus condition-based rules for the three failure modes with clear physical thresholds — where a simple threshold on vibration RMS outperformed the model and is explainable to a maintenance engineer — and better instrumentation on the drive electronics: the expensive failures were unpredictable largely because nothing that preceded them was being measured. Envion specified what to instrument and committed to an honest re-assessment in eighteen months.
Delivery
The deployment deliberately required no new decision from anyone: the inspection queue printout simply arrived in a different order. Technicians were told the ranking came from a model and nothing else changed — no alert fatigue, no new workflow, no trust to rebuild.
The condition-based rules went into the maintenance system as standard checks. The instrumentation specification for drive electronics was handed to the client's engineering team with an eighteen-month review date. Total programme cost: ~€70K against the €900K approved.
Outcome and evidence
Twelve months on: significant unplanned failures fell from 11 to 6.5 per month, annual disruption cost fell from €2.4M to €1.4M, technician inspection capacity was unchanged at 40 assets a week, false alarms numbered exactly zero because no alerts exist, and €830K of approved capital was never committed. The client got roughly 40% of the failure reduction the original business case promised, for about 8% of the cost, using the same model built during the assessment.
The advice that generalizes: do the alert arithmetic before approving the budget — failure rate, asset count, realistic precision at target recall, false alarms per week, then ask the person who'll receive them how long they'll keep responding; that conversation has killed more predictive maintenance projects than any technical finding, and it takes twenty minutes. Check whether the model predicts your expensive failures or your cheap ones. And ask whether there's a deployment that doesn't require a new decision from anyone — reordering work people already do is dramatically cheaper than creating work, has no false alarm cost, and tolerates a mediocre model.
| Metric | Before | After |
|---|---|---|
| Approach deployed | €900K alert system (planned) | €70K ranked inspection |
| Significant unplanned failures | 11/mo | 6.5/mo |
| Technician inspection capacity used | 40 assets/wk | 40 assets/wk (unchanged) |
| False alarms generated | — | 0 (no alerts exist) |
| Annual disruption cost | €2.4M | €1.4M |
| Capital not committed | — | €830K |
Client feedback
What the client says about this engagement
“Stanislav built us a working model and then spent the rest of the engagement explaining why we shouldn't deploy it the way we'd planned. The number that ended the discussion was two hundred and forty false alarms a month — he asked our maintenance manager how long his team would take those seriously, and the honest answer was about six weeks.
Turning it into a sorting order for inspections we were already doing is such a small idea and it's saving us a million a year. He also told us plainly that the model was worst at exactly the failures that cost us most, which no vendor was going to volunteer.”
From the engagement lead
What I’d tell anyone considering this
“Do the alert arithmetic before you approve the budget. Take your failure rate, your asset count, and a realistic precision at your target recall, and calculate how many false alarms per week your team will receive. Then ask the person who'll receive them how long they'll keep responding. That conversation has killed more predictive maintenance projects in my experience than any technical finding, and it takes twenty minutes.
Check whether the model predicts your expensive failures or your cheap ones. Gradual, well-signalled degradation is easy to predict and usually cheap to fix. Abrupt electronic and control failures are expensive and hard. If your business case is built on the expensive ones, verify the model actually finds them before anything else.
And ask whether there's a deployment that doesn't require a new decision from anyone. Reordering work people already do is dramatically cheaper than creating work, it has no false alarm cost, and it tolerates a mediocre model. Most predictive maintenance value is available this way, and it never gets proposed because it isn't impressive enough to put in a business case.”
Evidence gate. This page publishes only what Envion's project records and client disclosure permissions support. Outcomes are added once verified against a baseline, a measurement period, and an approved source.
FAQ
Questions about this case
Facing a similar challenge?
Approved budget for predictive maintenance? Do the alert arithmetic first — Envion’s feasibility read takes weeks and can save the entire build cost.
Discuss a Similar ChallengeKeep exploring
Similar case studies
Executive Technology Leadership
Support for high-stakes product and AI decisions
Bring senior technology leadership into the business when the roadmap is unclear, delivery is at risk, an AI initiative needs stronger ownership, or the company needs an experienced technical voice before hiring a permanent CTO.
Discuss Interim CTO SupportCore responsibilities
- Align product and technology priorities with business goals and measurable outcomes.
- Review architecture, delivery risks, data foundations, security needs, and AI readiness.
- Lead internal teams and external partners through a practical execution plan.
- Clarify team structure, ownership, decision rights, and delivery cadence.
- Support investor, board, partner, and due-diligence conversations with credible technical judgment.
New experience
Prompt-to-Page — try it right here
Describe the landing page you want, in your own words. We turn it into a finished page and email you a private link in 5–10 minutes — no briefs, no calls, $0 to see the result.
- Describe what you want to create.
- We structure, write, and compose the page.
- You receive a private link when it is ready.
Start with a sentence — the interactive builder takes it from there.
Generate My PageSafe, respectful content only. No obligation.