Predictive Maintenance ROI for Indian Manufacturers: The Business Case Your CFO Will Approve

The short answer

Predictive maintenance gets approved when you stop pitching it as a technology project and start presenting it as the recovery of a quantifiable leak: the cost of unplanned downtime you are already paying every month. Deloitte's own analysis puts the price of unscheduled downtime for industrial manufacturers at roughly 50 billion dollars a year, and reports that predictive approaches can increase equipment uptime by 10 to 20 percent while cutting overall maintenance costs by 5 to 10 percent (Deloitte Insights). The way to get a CFO to sign is not a plant-wide rollout. It is a narrow phase one on the 5 to 10 assets whose failure hurts most, sized so the downtime you recover pays for the work. This is a data problem first and a sensors problem second, and it is exactly the kind of leak we help capital-intensive plants (steel, tyres, auto components) put a rupee figure on.

The rest of this piece walks through the argument in the order a CFO hears it: why downtime is a leak rather than a line item, how to pick the handful of assets that make phase one self-funding, the four numbers that turn a vague ambition into an approvable case, the data you need before a single sensor goes on the wall, and the two extra recoverable lines that Indian steel and energy-heavy plants keep leaving on the table.

Why unplanned downtime is a leak, not a line item

A CFO does not fund "predictive maintenance." A CFO funds the recovery of money currently walking out the door. So the job is to make the leak visible in the units the CFO already trusts: rupees per hour of unplanned stoppage, and lakhs per unplanned failure.

In most plants that number does not exist as a report. Breakdown history sits in a maintenance register or CMMS, production loss in the OEE or shift logs, energy and consumable cost in the ERP. The true cost of a single unplanned trip (lost tonnes, restart energy, overtime, expedited spares, contractual penalties) lives across systems that never get joined. That gap is why downtime feels like an operational nuisance rather than a board-level number. Reconnect the systems and "the furnace tripped again last week" becomes "unplanned downtime on this asset cost us X lakh last quarter."

The scale of what stays hidden is easy to underestimate. Siemens and Senseye's True Cost of Downtime 2024 study found that two-thirds of surveyed plants hit unplanned downtime at least once a month, and among firms that tracked their outages the average event ran about four hours. The report also pegs an unproductive hour in automotive at roughly 2.3 million dollars, a figure that has climbed sharply since 2019. Indian plants rarely run at those absolute numbers, but the pattern holds: the per-hour cost is far higher than the maintenance team quotes, because the maintenance team is only counting the labour and the spare, not the lost contribution margin, the restart energy, or the customer penalty. When we reconcile the data, the fully loaded number is routinely several times the figure the plant had in its head.

Start with the 5 to 10 critical assets, not the plant

The mistake that kills these business cases is scope. A plant-wide programme has a large upfront number, a long horizon, and no clean way to attribute savings, so it dies in the capex committee. The winning move is the opposite: pick a handful of assets and let phase one self-fund. Three filters decide which qualify:

1. High failure cost. The asset whose stoppage halts a whole line or spoils a heat: an EAF transformer, a continuous caster, a Banbury mixer, a critical press. One prevented major failure can cover an entire monitoring programme.
2. Measurable degradation. The failure must announce itself through vibration, temperature, current draw, or oil condition, a signal that drifts before the asset fails. Instant, random failure gives prediction nothing to learn from.
3. Data history. Enough recorded failure and sensor history for a model to tell normal from pre-failure. An asset with years of logged trips beats a new machine with none.

Assets that clear all three earn payback fastest. Everything else waits for phase two, funded by phase one's results. In practice we ask the plant to rank its top 20 or 30 assets by how badly a stoppage hurts, then run each candidate through the three filters. The list usually collapses to a shortlist of 5 to 10 very quickly, because the assets with the highest failure cost are often the same ones the plant has been nursing for years and therefore has the richest breakdown history on. That overlap is the reason a tightly scoped phase one tends to hit its number: you are aiming the first model at the assets that both hurt the most and have the most to learn from.

There is a discipline point here too. Sequencing the work asset by asset, rather than instrumenting everything at once, keeps the programme honest. Each asset you bring online produces an attributable result, and that result is what funds the next one. A capex committee will approve a small, self-funding first step far sooner than an open-ended platform.

Building the four-line number your CFO will approve

A business case a CFO approves has four lines, and every one comes from your own data, not a vendor slide:

• Baseline downtime cost. Unplanned stoppage hours per critical asset times the fully loaded cost per hour: lost contribution margin, restart energy, overtime, penalties. This is the leak, and it is the line most plants have never actually computed.
• Recoverable share. A hedged, sourced range, not a promise. Deloitte's 10 to 20 percent uptime gain and 5 to 10 percent maintenance-cost cut are conservative anchors. McKinsey's operations research on analytics-based maintenance reports larger reductions in machine downtime and maintenance cost for plants moving off reactive regimes, and independent reviews cite 30 to 50 percent unplanned-downtime reductions for reactive-heavy operations. Present a conservative and a stretch case, and be clear which published range each rests on.
• Programme cost. Sensors where they are missing, the data pipeline that unifies CMMS, OEE, and ERP, and the model itself. In a partly instrumented plant the pipeline is often the larger line, not the hardware, and pretending otherwise is how programmes overrun.
• Payback window. Recoverable leak divided by programme cost. Published critical-asset programmes commonly show payback in a 6 to 18 month range, fastest in high-cost sectors where one averted failure pays for the year.

Present those four lines and the conversation stops being "should we buy an AI tool" and becomes "here is a leak of this size and here is what closing it costs." That is a decision a CFO can make. The honesty of the recoverable-share line matters more than its size: a case built on a conservative range that then overdelivers earns you phase two, whereas a case built on a vendor's best-day number that underdelivers ends the programme.

The data prerequisites most plants skip

Before any model can predict anything, three streams of data have to exist and, more importantly, have to be joinable. The first is a clean failure and maintenance history: what broke, when, why, and how long it took to recover, ideally coded consistently rather than as free-text notes. The second is condition or sensor data on the asset, whether that is existing PLC and SCADA tags or new vibration and temperature sensors. The third is the cost and production context, from OEE and the ERP, that turns a predicted failure into a rupee figure.

The recurring problem is not that the data is absent. It is that the three streams sit in three systems with no shared asset identifier, no common timestamp convention, and no agreement on what a "failure" even is. One system logs a breakdown, another logs the production loss for the same event, and nothing links them. This is why we say predictive maintenance is a data problem before it is a sensors problem: you can bolt on the most accurate vibration sensor in the world, but if its readings cannot be joined to the failure that followed and the cost that resulted, the model has nothing to learn the pattern from. Getting the join right, so that every past failure carries its sensor signature and its rupee cost, is the unglamorous work that decides whether the model is trustworthy. It is also the work that produces the baseline number in the CFO case, which is why we do it first regardless of whether the plant ever buys another sensor.

The India angle: energy per tonne and CCTS

For Indian steel and other energy-heavy plants there is a second recoverable line most cases miss: energy. An EAF or furnace running with degrading equipment burns more power per tonne than a healthy one, and power is one of the largest variable costs in steelmaking. Catching degradation early protects cost per tonne, not just uptime, and on a high-throughput asset that quiet efficiency drift can rival the downtime line in size.

That line is about to matter more. India is transitioning from the Perform, Achieve and Trade (PAT) energy-efficiency scheme to the Carbon Credit Trading Scheme (CCTS), which shifts compliance from energy consumption to carbon emissions measured against sector benchmarks. Equipment that runs efficiently emits less, which now carries direct compliance and credit value on top of the fuel saving. A plant that keeps its critical assets in healthy condition is not only avoiding downtime, it is protecting its position against the emissions benchmark it will be measured on. A business case that ignores the energy-per-tonne and CCTS lines understates its own return, sometimes substantially.

Key takeaways

• Predictive maintenance gets funded as leak recovery, not as a technology purchase. Present unplanned downtime as rupees per hour and lakhs per failure, in the CFO's own units.
• The number rarely exists as a report because breakdown, production, and cost data sit in separate systems (CMMS, OEE, ERP). Joining them into one view is the first build, and it produces the baseline number the whole case rests on.
• Scope phase one to 5 to 10 assets that clear three filters (high failure cost, measurable degradation, and enough data history) so the recovered downtime self-funds the work.
• Hedge every figure. Deloitte cites 10 to 20 percent uptime gains and 5 to 10 percent maintenance-cost cuts, McKinsey and independent reviews report larger reductions for reactive-heavy plants, and published critical-asset paybacks commonly fall in a 6 to 18 month range. Your own baseline decides where you land.
• For Indian steel and energy-heavy plants, add the energy-per-tonne and CCTS compliance lines. Degrading equipment quietly inflates both.

How to find your own number

The ranges above are industry figures, not promises. Your real payback depends entirely on your own downtime cost and how instrumented your critical assets already are. The only way to know is to reconcile your maintenance history, production loss, and cost data into a single view and put a rupee figure on the leak per asset.

That reconciliation is the first thing we build. Our AI Stack Audit x-rays your CMMS, OEE, and ERP data, quantifies what unplanned downtime is actually costing you per critical asset, and tells you which 5 to 10 assets make the strongest phase-one case, before you spend on a single sensor. For how we work with plants specifically, see our manufacturing practice page.

Frequently asked questions

How do I justify predictive maintenance to my CFO?

Present it as recovery of a quantifiable leak, not a technology purchase. Calculate the fully loaded cost of unplanned downtime per critical asset, apply a hedged recoverable range from published sources, and divide the recoverable amount by the programme cost to show a payback window. A CFO approves a leak-and-payback case far more readily than an "AI tool" case.

Which assets should I start with?

Start with 5 to 10 assets that clear three filters: high failure cost (stopping one halts a line or spoils output), measurable degradation (a physical signal like vibration or temperature that drifts before failure), and enough recorded history for a model to learn from. Scoping tightly lets phase one self-fund the rest.

What payback period should I expect from predictive maintenance?

Published case studies commonly report payback in a 6 to 18 month range for critical-asset programmes, with the fastest returns in high-cost sectors where one prevented failure covers the year. Deloitte cites uptime gains of 10 to 20 percent and maintenance-cost reductions of 5 to 10 percent. Treat these as planning ranges. Your own downtime baseline sets your actual number.

What's the real ROI difference between preventive and predictive maintenance?

Preventive maintenance services assets on a fixed schedule regardless of condition, so you replace parts that still had life and still miss failures that arrive off-schedule. Predictive maintenance acts on the asset's actual condition, targeting intervention only when degradation appears. The gain is fewer unplanned failures and less wasted preventive work, but it only pays off on assets whose degradation is measurable and whose failure is expensive.