The Pilot Says Maybe: Deciding on a Predictive Maintenance Rollout
You lead reliability at a three-line bottling plant. A vendor proposes vibration and thermal monitoring on 12 critical motors: $180,000 capital, $24,000 per year in subscription, and a projected $310,000 annual saving built from a case study at another bottler. Your own ten-week pilot on three motors produced nine alerts: five were confirmed bearing degradation caught early, three were false positives that triggered unnecessary inspections, and one seal failure was missed entirely. Median warning time on true positives was 11 days, but one flagged motor failed two days after the alert. Baseline unplanned downtime last year was 42 hours at about $9,800 per hour, though only a subset of that was motor-related. The capital committee meets in five weeks; the vendor's 20 percent discount expires in three. Your maintenance planner is enthusiastic; your controls engineer worries about sensor data ownership and who maintains the models after year one.
1. What does the pilot evidence actually support, and what does it not?
Read the worked answer
What it supports: on three motors over ten weeks, eight of nine alerts were actionable enough to inspect; five corresponded to real bearing degradation; the system can generate usable early warnings on at least one failure mode. That is a plausible signal, not a proven business case.
What it does not support: precision is 5/9 (about 56 percent), meaning roughly one inspection in two was avoidable, and the false-positive load falls on a maintenance team that is already stretched. Recall is unknown - you caught one known miss, but you do not know how many degradations occurred without an alert. Median lead time of 11 days is encouraging, but the two-day case matters more than the median: for fast-developing failure modes, a late alert is not a planned intervention, it is a notification. Nine alerts is far too small a sample to estimate savings, and the vendor's $310,000 figure comes from a different plant, different duty cycle, and different baseline.
The downtime denominator needs work before any ROI exists: 42 hours at $9,800 is about $412,000, but only a subset is motor-related, and only a subset of that is detectable and avoidable with 11 days' notice. Until you break downtime down by failure mode and cause, the savings number is a guess with a decimal point.
Conclusion: the pilot justifies a controlled expansion, not a plant-wide commitment, and certainly not a decision made in three weeks to capture a discount.
2. Recommend a decision with explicit conditions, including what to do about the expiring discount.
Read the worked answer
Recommendation: stage the rollout to the full 12 motors, but as a gated six-month deployment with contractual and operational conditions, not a one-way purchase.
Conditions: (1) Data ownership: plant owns raw and derived data; vendor gets a licence for model improvement only, with an opt-out. (2) Model maintenance named: who retrains, how often, what happens at year two, and what the subscription covers when the model drifts. (3) CMMS integration so alerts become work orders with a named owner and a response standard. (4) Success gates at month six: alert precision at or above 70 percent on targeted modes, median lead time at or above seven days for slow modes, and a documented avoided-downtime figure from actual work orders. (5) Exit clause if gates are missed, priced in advance.
On the discount: $180,000 at 20 percent is $36,000. That is real money, but it is smaller than the cost of a three-year commitment to a contract with unresolved data and maintenance terms, and far smaller than the cost of a monitoring system the planners stop trusting after a false-alarm summer. Ask for the discount to be held against a signed letter of intent with the conditions attached; if the vendor refuses, let the discount go and say why in the committee paper.
Expand scope cautiously: add the highest-criticality motors first, keep the three pilot motors instrumented as a control, and do not let enthusiasm turn the six-month review into a formality.
3. The CFO wants a single payback number. How do you present the decision?
Read the worked answer
Refuse the single number, but not by lecturing. Give a one-line headline - "between one and four years, most likely around two and a half" - then show the table behind it. A CFO who is given a range with named drivers can make a decision; a CFO given a point estimate later discovers it was fiction and discounts everything else you say.
Scenario table: best case assumes 12 motor-related downtime hours avoided at full cost, high alert precision, and no additional inspection labour. Base case assumes six to eight hours avoided, 70 percent precision, and two extra inspection hours per false alert. Worst case assumes the missed-failure pattern repeats, precision stays near 56 percent, and the team loses confidence and stops responding to alerts - which is a realistic failure mode and should be costed.
Name the assumptions explicitly: downtime hours attributable to motors, cost per hour, fraction detectable with 11 days' notice, and inspection labour. Invite the CFO's finance analyst to challenge the downtime figure - the number improves when it is theirs, not yours.
Then give the decision structure, not just the arithmetic: the ask this cycle is the staged deployment with gates and an exit clause, not the full three-year commitment. That is a smaller, reversible bet with the same learning value. Frame the gates as the CFO's protection: if the six-month numbers do not hold, the exit is pre-priced and the downside is capped. That is a stronger argument than a payback figure, and it is the honest one.