Industrials are the rare place AI hits the P&L directly
Most AI spending in PE portfolios funds dashboards nobody acts on. Manufacturing is different — an hour of unplanned downtime, a percent of scrap, a warranty claim are all measured in dollars already. You don’t need a new business model. You need fewer machine stoppages and fewer bad parts.
The size is real but specific. McKinsey’s Global Lighthouse cohort — independently vetted factory sites, not vendor decks — reports AI use cases delivering up to a 99% reduction in defects, ~40% higher labor productivity, and ~48% shorter lead times, with one site lifting OEE by 10 points while halving unplanned downtime (McKinsey; WEF).
The trap: those are the winners. MIT’s NANDA study found 95% of generative-AI pilots deliver no measurable ROI, and BCG reports ~60% of companies generate no material value (via MIT/BCG reporting). The gap between Lighthouse results and the median pilot is the whole ballgame.
Play #1 — Predictive maintenance: the cleanest EBITDA line in the building
This is the most mature, most provable play. Sensors plus ML predict failure before it happens, so you replace a part on a planned Sunday instead of losing a Tuesday shift. McKinsey pegs the lever at up to ~50% less unplanned downtime, 10–40% lower maintenance cost, and 20–40% longer asset life (via Körber).
Named proof: Augury’s vibration monitoring delivered DuPont a 7x ROI in under a year, helped Colgate-Palmolive avoid a line failure worth 2.8 million tubes of product, and saved chemical maker ICL ~$1M in downtime at one plant in under 10 months (Augury). The napkin math: a plant with 6,000 production hours, 3% unplanned downtime, and $20K/hour contribution margin loses $3.6M/year to stoppages — so a 20% cut is ~$720K of EBITDA per plant, per year (Withum).
"What did unplanned downtime cost us last year in lost contribution margin — by line, not in aggregate? If you can't answer in dollars per hour per asset, we can't size the ROI and shouldn't fund the pilot yet."
Play #2 — Computer-vision quality inspection: kill scrap and warranty at the source
Cameras plus a defect-detection model inspect 100% of parts at line speed — something humans physically can’t do. Human inspectors miss 20–30% of defects and degrade after two hours of staring (Overview.ai). The EBITDA hits two lines: scrap/rework caught before more value is added, and warranty caught before it ships.
The marquee independent example is BMW, which uses AI image recognition across plants and reports up to ~60% fewer defects, with a single AI application saving over $1M/year at one site (AMS). The skeptic’s note: vision ROI is real but lumpier than predictive maintenance. Fund it where you have a known, expensive, recurring defect — not a blanket “inspect everything” mandate.
Play #3 — AI scheduling & throughput: more output from the same iron
This uses ML to sequence jobs against machine, labor, and material constraints — squeezing more saleable units out of equipment you already own. No capex. Deloitte’s 2025 survey found companies using AI for production planning unlocked up to 15% more capacity from the same floor, workforce, and machines (via TVSNext). Throughput on a capacity-constrained line flows almost entirely to EBITDA. Insist on a before/after OEE baseline.
Play #4 — Supply chain & inventory: free up the cash on the floor
This one moves the balance sheet, which matters for levered deals. AI demand forecasting lets operators cut inventory 20–30% and forecast error 20–50% while holding service levels (McKinsey). For a sponsor, that’s a one-time cash release that pays down debt or funds the next bolt-on — the most underrated industrial AI play because operators think “AI” means the factory floor, not working capital.
"Name our single most expensive recurring defect and its annual scrap + warranty cost. Vision inspection pays back fastest against a known, expensive, repeating defect — not a vague 'improve quality' goal."
What kills these initiatives
- No dollar baseline. “Improve quality” isn’t a target. If downtime/scrap isn’t measured in $/hour/line before the pilot, ROI can’t be proven.
- Pilot purgatory. The demo works, so nobody kills it — but the data pipes, MES integration, and operator workflow to run it daily were never built.
- Treating it as IT, not operations. AI bolted on by a vendor without floor buy-in; operators don’t trust the alerts.
- Dirty, siloed data. Most failures trace to data, not the model — sensor gaps, inconsistent lighting, no labeled defect history.
- Spreading thin. Five half-funded pilots lose to one fully-scaled play. Pick the asset with the biggest, best-measured pain.
- Energy/ESG theater crowding out the P&L plays. Don’t let a 3% energy story eat the budget for a 50% downtime story.
Sources
| Source | What it told us | Confidence |
|---|---|---|
| McKinsey — Lighthouse Network | Vetted factory results: 99% defect cut, ~40% productivity, +10pt OEE, halved downtime | STRONG |
| WEF — Lighthouse sites | Corroborates Lighthouse as independent benchmark | STRONG |
| Körber (McKinsey data) | Up to 50% downtime cut, 10–40% maintenance cost, 20–40% asset life | STRONG |
| McKinsey — supply chain | 20–30% inventory cut, 20–50% forecast-error cut, service maintained | STRONG |
| Augury — DuPont case | DuPont 7x ROI; Colgate failure avoided; ICL ~$1M saved | MEDIUM |
| AMS — BMW | BMW vision inspection: ~60% defect cut, $1M+/yr from one application | STRONG/MED |
| Overview.ai | Humans miss 20–30% of defects, degrade after 2 hrs | MEDIUM |
| Deloitte (via TVSNext) | Up to 15% more capacity from same assets | MEDIUM |
| Withum | The $3.6M downtime / $720K EBITDA napkin model | MEDIUM |
| MIT/BCG failure data | 95% pilots no ROI; 88% never reach production | STRONG |