Maximo APM + Visual Inspection: A 5-Asset Pilot That Paid for Itself in 90 Days

The Maximo Application Suite includes a powerful set of reliability and asset performance management capabilities: Monitor (real-time telemetry), Health (asset health scoring), Predict (failure prediction models), Visual Inspection (computer vision on images and video), and Condition Insight (the human-friendly dashboard). The full stack is a multi-million-dollar, multi-year initiative. Most plant managers do not have the stomach for it.

The good news: a 5-asset pilot with a clear scope, a tight timeline, and a defensible ROI will pay for itself in 90 days and give the executive sponsor the confidence to fund the full rollout. This article is the field-tested pattern for that pilot.

The 5-asset rule

The number 5 is not magic. It is the result of a trade-off:

• Too small (1-2 assets): The pilot is not representative. The executive sponsor cannot extrapolate the ROI to the full fleet.
• Too large (20+ assets): The pilot becomes the rollout. The timeline balloons, the budget gets cut, the executive sponsor loses patience.
• 5 assets: The pilot is representative (different asset classes, different failure modes, different data sources). The timeline is bounded (90 days). The budget is defensible (under $200K all-in).

Choose the 5 assets carefully. The ideal mix:

• 2 critical assets — high consequence of failure, high downtime cost
• 2 mid-criticality assets — moderate consequence, moderate cost
• 1 problem asset — known history of failures, the sponsor wants it fixed

The problem asset is the most important. If the pilot solves a known problem, the sponsor will be a champion. If the pilot is on a "we think this might be useful" asset, the sponsor will be a skeptic.

Phase 1: Define the failure modes (Week 1-2)

The most common pilot failure is scope creep on failure modes. The team says "let's predict every failure," the project is in scope for everything, and the pilot never delivers. The fix: define 1-3 failure modes per asset, total 5-15 across the pilot.

A failure mode is a specific, measurable, preventable event. Examples:

• Pump cavitation — measurable by vibration at the pump bearing, preventable by adjusting suction pressure
• Bearing wear on a motor — measurable by vibration RMS trending up, preventable by lubrication
• Valve seat leak — measurable by acoustic emission or visual inspection, preventable by re-seating
• Corrosion under insulation — measurable by Visual Inspection (UT thickness, IR thermography), preventable by coating repair

For each failure mode, define:

• The detection method (vibration, acoustic, IR, visual, AI vision)
• The data source (online sensor, manual reading, image capture)
• The lead time (how many days/hours between detection and failure)
• The cost of failure (downtime, repair, safety incident)
• The cost of prevention (the action you would take if alerted)

The cost-of-failure column is what makes the pilot defensible. If a pump cavitation failure costs $250K in downtime and you can detect it 30 days early, the prevention action is essentially free.

Phase 2: Configure Monitor (Week 3-4)

Monitor is the real-time data ingest service. For the pilot, you do not need a full Industrial IoT integration. You need:

• A small number of sensors — vibration, temperature, current draw
• A gateway — a Siemens RUGGEDCOM, a Honeywell edge box, or a Raspberry Pi running the MAS Edge agent
• A Monitor pipeline — a defined YAML that ingests, normalizes, and persists

The Monitor pipeline is configured in the Monitor Application Designer. A typical pump-cavitation pipeline has 5-10 stages:

name: pump-cavitation-monitor
source:
  type: kafka
  topic: edge-iot-pump-001
stages:
  - type: filter
    where: sensor_id = 'pump-001-bearing-vib'
  - type: rolling-window
    window: 60s
    aggregation: rms
  - type: threshold
    field: rms
    warning: 4.5
    critical: 6.2
  - type: anomaly-detector
    model: pump-bearing-baseline
    threshold: 0.85
  - type: kafka-publish
    topic: mas.health.scores

The pipeline is hot-reloadable. You can iterate on the thresholds in 10 minutes, watch the alerts in real time, and tune the model in the same week.

Phase 3: Train the Predict model (Week 5-6)

Predict is the failure prediction service. The training data is the work order history plus the sensor history for the past 18-36 months. If the asset does not have 18 months of history, the pilot is too early for Predict. Use Health (rule-based scoring) until you have the data.

For the assets with history, the training flow is:

1. Extract the work order history for the asset and the related failure codes.
2. Extract the sensor history for the same period.
3. Join the two on the failure date — for each failure, the 30 days of sensor data leading up to it.
4. Train the model — Predict has a built-in auto-ML pipeline that picks the best of {random forest, gradient boosting, LSTM, transformer} for the data.
5. Validate the model — hold out 20% of the failures, predict them, measure recall and precision.
6. Deploy the model — push to the Predict service, configure the alert threshold.

The validation step is non-negotiable. A model that predicts 100% of the failures but generates 10 false positives per day is worse than no model. Target >70% recall and >50% precision in the first pilot. Tune later.

Phase 4: Configure Visual Inspection (Week 5-7)

Visual Inspection is the computer vision service. The pilot use case is usually corrosion, leak, or anomaly detection on a small set of camera feeds. The flow:

1. Capture 200-500 images of the asset in its normal state.
2. Capture 50-100 images of the asset in its failed state (or simulate via annotation).
3. Annotate the images — the Visual Inspection labeling tool accepts bounding boxes, segmentation masks, or classification labels.
4. Train the model — Visual Inspection uses IBM's foundation model for vision, fine-tuned on your labels.
5. Deploy the model — push to the Visual Inspection service, configure a camera feed (live or periodic image capture).
6. Alert on anomaly — when the model flags an image, send an alert to Maximo via a work order or a notification.

The 200-500 image count is the minimum for a working model. If you have fewer, the model will not generalize. The pilot team should be capturing images from week 1 in parallel with the other phases.

Phase 5: Wire Health scoring (Week 7-8)

Health is the asset health scoring service. It consumes the Monitor thresholds, the Predict scores, and the Visual Inspection anomalies, and produces a single 0-100 health score per asset. The health score is what the operator sees in the dashboard.

A typical health score formula for a pump:

asset: pump-001
health:
  inputs:
    - source: monitor.vibration_rms
      weight: 0.4
      function: inverse_distance_to_threshold
    - source: predict.failure_probability_30d
      weight: 0.3
      function: 1 - probability
    - source: visual.inspection_score
      weight: 0.2
      function: score / 100
    - source: workorder.open_criticality
      weight: 0.1
      function: 1 - (1 / (1 + open_criticality))
  bands:
    green: 80-100
    amber: 50-79
    red: 0-49

The inverse_distance_to_threshold function is the most important. As the sensor reading approaches the threshold, the score drops. As it crosses the threshold, the score collapses. The dashboard turns amber, then red, then the operator gets a notification.

Phase 6: Measure the ROI (Week 9-12)

The last phase is the one the executive sponsor cares about: what did this cost, what did it save, and what is the next step?

The cost side:

• MAS APM subscription (5 assets, 90 days) — $20K
• Sensors and edge hardware — $30K
• Implementation services (consultant, 90 days, part-time) — $60K
• Internal labor (reliability engineer, part-time) — $40K
• Total — $150K

The savings side (from the 5-asset pilot):

• 1 pump cavitation avoided — $250K
• 2 motor bearing replacements caught early — $80K
• 1 valve seat leak avoided — $60K
• 2 unscheduled downtime events avoided — $120K
• Total — $510K

The ROI is 3.4x in 90 days on a $150K investment. The payback period is 45 days. The annual run rate, extrapolated to the full fleet, is $4M in savings on a $1.2M subscription cost. That is the executive sponsor's number.

The rollout plan

After the pilot, the rollout is straightforward:

• Year 1: Add 50 critical assets
• Year 2: Add 200 more
• Year 3: Add the remaining 500-1000

The 5-asset pilot has already validated the architecture, the workflow, the data integration, and the ROI model. The rollout is execution, not innovation.

The pitfalls

The most common pilot failures are:

1. No problem asset — the pilot is on a "we think this might be useful" asset, the executive sponsor is not a champion. Fix: include at least one problem asset.
2. No failure mode discipline — the pilot is in scope for everything, the timeline balloons, the executive sponsor loses patience. Fix: 1-3 failure modes per asset.
3. No image capture from week 1 — the Visual Inspection model is starved for data. Fix: image capture starts in week 1, in parallel.
4. No early success — the pilot delivers in week 12 with no earlier wins, the executive sponsor has moved on. Fix: ship a small win in week 4 (a threshold alert, a single health score, a single anomaly detected).
5. No ROI narrative — the pilot is technically successful but the business case is not built. Fix: the cost-of-failure and the cost-of-prevention columns are filled in week 1.

The bottom line

The 5-asset pilot is a 90-day, $150K, 3.4x-ROI project that is repeatable, defensible, and sets up the full rollout. The Maximo APM stack is mature enough in 2026 that the pilot is execution, not experimentation. The customers who have done it in the last 12 months have a clear path to the full fleet. The customers who are still "evaluating" are running the same 5-asset pilot in 18 months, having lost a year.