Field reports/Briefing · Contrarian Thesis

Why 80% of predictive maintenance projects fail — patterns from 200+ Chinese factory floors.

The published research says 80% of predictive maintenance initiatives fail to deliver, falling apart within 12–18 months. The team has now walked 200+ Chinese factory floors with Inspectaman. The dominant pattern is not technological. It is organisational, and it is mostly invisible from the procurement table.

Published 16 May 2026Read time 12 min

The first time a reliability lead said it on a factory floor, the team thought it was a one-off joke. By the fiftieth audit it was no longer funny. The phrasing varied — Mandarin, Cantonese, broken English — but the meaning was identical:

"Yeah, we have predictive maintenance tools. They are on the shelf in the office."

Under a coating of dust. Three vendors deep, two consultants paid out, one champion who has since moved to a different company. The sensors are still attached to the asset. The dashboard URL still works. Nobody opens it. The published industry research (LLumin, 2025) puts the failure rate at 80%, falling apart within 12–18 months. The team has now walked enough floors to corroborate that number with on-site observation rather than survey response.

This piece is a structured walk through the three dominant failure modes. None of them is about the AI. Two of them are about the organisation. One is about the way the technology was sold. All three are addressable, and the addressing is mostly architectural rather than analytical.

Failure mode one — the champion exits, the program decays.

The pattern: a senior reliability engineer or maintenance manager champions the deployment, owns the relationship with the vendor, runs the weekly review, drives the work-order generation from the dashboard. Eighteen months in, the champion gets promoted, retires, or leaves for a competitor. The program has no second-tier owner. Within six months, dashboard usage drops to zero. Within a year, the sensors are decommissioned to free up SCADA capacity for a different project. The historian gets archived.

Across the audits, the team can map the rate of program decay against organisational depth. Plants with a three-person reliability function and a clear succession line keep the program running. Plants with a one-person reliability champion lose the program when that person exits — which, given a 38% SA engineer shortage and intense competition for experienced reliability talent, happens roughly every 28 months.

What working deployments do differently: program ownership is structural, not personal. The review meeting is on the calendar regardless of who is in the chair. Work-order generation is tied to a CMMS workflow that survives staff turnover. Dashboard usage is part of the daily standup, not a separate ritual. The team built Foundry Cortex's pilot scoping around this explicitly — the 60-day pilot exits with a documented operational rhythm, not just a working dashboard.

Failure mode two — the dashboard does not generate work.

The second pattern is structurally worse because it can persist for years without anyone noticing. The dashboard is live, the alerts are firing, the reliability lead is reviewing them — but the alerts do not generate work orders. The maintenance team continues to operate on the old run-to-failure schedule. The PdM system functions as a notification layer over an unchanged workflow.

The team has audited operations where the PdM alerts and the actual work-order log have zero correlation across an entire quarter. The alerts predicted seventeen failures; sixteen of them went on to fail; the maintenance team responded to all sixteen reactively after the fact. The forecasting was technically correct. The integration was missing.

The Foshan tier-2 plant the team has previously written about (field report here) is at the opposite extreme: 85% of predicted-failure alerts flow autonomously into the CMMS, generate a work order, and reach a technician at the next daily standup. The remaining 15% escalate to the reliability lead because the model confidence is low or the failure mode is novel. This is what closed-loop PdM looks like in practice. Anything less is a notification layer pretending to be a maintenance program.

Failure mode three — the technology was oversold for the context.

The third pattern is the one most vendors will not write about. A consultant or platform vendor pitches a deployment built for a Fortune 500 reference customer — multi-plant data fabric, asset-graph contextualisation, embedded LLM workflows — to a mid-tier operation that has neither the data hygiene nor the organisational depth to absorb it. The pitch lands because the buyer is impressed. The deployment goes live. Six months in, it is clear the platform is doing 20% of what was demoed because the underlying systems cannot feed it. The remaining 80% of the licence fee is being paid for capability nobody can use.

The team has now seen this pattern at scale across Chinese tier-2 operations and is starting to see early instances of it in South Africa. The structural risk is rising as enterprise vendors (Cognite, AVEVA, SymphonyAI) chase mid-market growth and apply the enterprise sales motion to buyers who need a different product entirely. The honest framing is: ask the vendor what their smallest deployment with this stack looks like. If the answer involves a multi-plant rollout or an integration partner billing R5M, the technology was built for someone else.

What the working 20% does differently.

The team can compile a short list from the audits where the deployment was still generating value at the 24-month mark. The differentiators are not glamorous.

  1. Program ownership is structural. At least two people own the review cadence. The work-order generation is a CMMS workflow, not a human-in-the-loop email.
  2. The technology was scoped to the operation.Twenty-five to fifty top-decile assets instrumented in detail. Not a multi-thousand-asset data fabric.
  3. The review cadence is part of an existing ritual. The daily standup or the weekly reliability review absorbs the PdM output. No separate dashboard meeting. No competing source of truth.
  4. There is a baseline of human pattern knowledge captured against the assets. The senior maintenance engineer's reasoning is in the system, not just in their head (see the Bill Problem field report).
  5. The pricing was honest enough to survive a CFO review at month 18. The deployment that gets defunded in year two was usually priced on the assumption it would justify itself in year one. The deployments that survive were priced on the assumption they would justify themselves over a three-year window — and are explicit about it.

What Foundry Cortex was engineered around.

The team built Foundry Cortex's pilot model around the working-20% pattern, deliberately and explicitly. The 60-day pilot scopes a 10–20 asset deployment — small enough to integrate, large enough to demonstrate. The pilot exits with: a working sensor estate, a closed-loop CMMS workflow, a captured Bill Inventory against the instrumented assets, and a documented review cadence absorbed into the existing operations rhythm. Not a dashboard that looks good in a board pack and decays in eighteen months.

The pricing model is engineered to survive the year-two CFO review. R100,000 fixed-price pilot. R575,000 per year ongoing for a 50-asset facility, quarterly billed, ZAR invoiced. Public, on the pricing page. No per-user fees, no integration uplift, no procurement loop. If the math does not work at year three, it does not work at year one. The team would rather lose the deal on day one than be defunded in month twenty-two.

The 80% failure rate is not a story about AI being immature. It is a story about industrial AI being sold the same way enterprise software has been sold for two decades — to champions, on the basis of capability rather than rhythm, without an explicit theory of how the program survives year two. Anything that does not address those three failure modes head-on is building toward the same outcome.

Want field intelligence every Tuesday morning?

The Industrial AI Briefing — six-minute read. What we saw on a factory floor this week and what it means for mid-market industry across Africa, India, Brazil, GCC, Australia, and SEA. Unsubscribe anytime.

Subscribe to the Briefing →