AI Has Landed on the Factory Floor. It Just Didn't Take Root.

The headline "AI has landed on the factory floor" sounds like victory. It's closer to pilot season.

McKinsey's own State of AI 2025 report finds that 88% of global enterprises have adopted AI. One-third have scaled beyond pilot projects. Just 6% qualify as "AI high performers" — the firms that achieve 5% or greater EBIT impact. Eighty-eight percent adoption, six percent payoff. That gap between "we tried it" and "it moved the P&L" is where the real story lives.

Any astute operator would have noticed this divergence by now. But the private equity world is running on a different set of numbers.

The Private Equity Value Creation Pitch Is Built on a Single Case Study

Private equity firms are treating AI as a structural lever in their industrial deal underwriting, not an afterthought. The pitch, as documented across industry guides and consulting whitepapers in 2025–2026, promises 200–300 basis points of EBITDA margin expansion across portfolio companies. The cost reduction narrative targets 8–15% improvement in the addressable cost base within 18 months.

That sounds engineered. It's actually aspirational.

The strongest evidence for this thesis is an anonymized case study from a $400M European lower-mid-market buyout fund. One portfolio company improved operational EBITDA by 280 basis points in 12 months, "directly attributed to AI deployment." A second fund closed two deals sourced through AI intelligence and cut its time-to-investment-committee memo by 55%.

That is the entire track record: one operational case study and one diligence-speed anecdote.

The same guide also documents the failure mode. A $1.5B US growth equity fund attempted portfolio-company AI deployment and failed — not because the technology didn't work, but because the operating partner team wasn't bought in. The initiative only moved forward after restructuring incentives in month 9. That is the dominant constraint: AI rollout in PE is a people and incentives problem, not a technology problem.

The pitch also carries a multiple arbitrage assumption. Some PE firms are buying platforms at 8–10x EBITDA and modeling exits at 15–20x, betting they can reclassify industrial services businesses to "look like software" after AI deployment. That is a valuation bet dressed up as an operational thesis. LPs are not yet pressure-testing whether those exit multiples hold when every firm tells the same story. But they will. Multiple compression is the quiet risk no one is underwriting.

What Actually Works on the Factory Floor

The independent evidence base is narrower than the pitch but real. An audited study of 200 B2B AI deployments from 2022–2025, published in early 2026, shows a median ROI of +159.8% over 24 months. The mean is +347%, but that's driven by outliers — the difference between what works and what doesn't is enormous.

Function-specific data tells a sharper story:

• Predictive maintenance — 10:1 to 30:1 returns. Payback within 12–18 months. 95% of deployments report positive ROI. This works because it replaces reactive or time-based maintenance (still used by 82% of manufacturers) with condition-based action. Low-hanging fruit, not magic.
• Quality control — 374% average 3-year ROI per Forrester, with 6–12 month payback. Detection accuracy of 95–99% versus 80–85% for human inspectors. Works best in high-volume standardized production; returns collapse in job shops producing low volumes of varied parts.
• Demand forecasting — 10–20% accuracy improvement, up to 5% inventory reduction. 18–24 month payback. The weakest ROI because it's harder to attribute AI-driven forecast improvement to specific inventory savings when production volumes fluctuate.
• Energy optimization — 10–15% cost reduction. Frequently fails to deliver the claimed ROI because the energy baseline was never established pre-deployment, making it impossible to separate AI gains from seasonal or rate variations.

The pattern is clear: AI delivers where the physical process is measurable, the baseline is documented, and the deployment can be normalized against production volume. The closer you get to soft metrics — forecasting, pricing, working capital — the harder it becomes to separate signal from noise.

Deloitte's 2025 Smart Manufacturing survey confirms how early this all is: only 29% of manufacturers are using AI/ML at the facility or network level. 23% are still piloting. The vast majority of industrial companies haven't even crossed the pilot threshold.

The Public Companies Are the Fastest Reality Check

If AI has truly landed on the factory floor, the public industrial AI companies should be reflecting that in their economics. They aren't — or they are, but the market has separated the real players from the pitch decks.

C3.ai — the ticker literally named "AI" on the NYSE — posted fiscal Q2 2026 results showing revenue growth of 7% quarter-over-quarter, driven by federal bookings that grew 89% year-over-year. But the full-year picture is brutal: revenue is down 35.7% year-over-year. Gross margin has collapsed to 30.9%. Operating margin is -199.2%. Free cash flow burned $192.1M over the trailing twelve months. The stock is at $10.22, down 24.2% year-to-date and 32.6% on a rolling annual basis. That is not a company whose customers are scaling industrial AI deployments at profit. That is a company burning $192M a year while revenue contracts by a third.

Palantir, by contrast, is the actual scaling story — but it's a government and defense scaling story, not a factory floor one. Revenue grew 78.9% year-over-year. Operating margin is 42.8%. Free cash flow hit $3.36B, up 96.5% year-over-year. ROIC is 32.6%. The stock is at $172, up nearly 40% over five days as of this week. But the vast majority of that revenue growth is US government contracts, which grew 115% year-over-year and represent 81% of total revenue. Palantir's AIP (Artificial Intelligence Platform) is genuinely capable — it layers AI on top of messy operational data to productionize models safely. But the factory floor is not the primary customer. The Pentagon is.

That distinction matters for the PE thesis. If private equity firms are betting that AI deployment will expand industrial EBITDA margins, they should be buying software whose customers are factories, not defense contractors.

The Gartner Contradiction Nobody Cites

Gartner reported in 2024 that 73% of AI projects fail to deliver expected business value. McKinsey's 2025 update finds only 5.5% of organizations across all sectors document financial returns from AI. In manufacturing, returns are slightly higher because output ties to physical throughput rather than productivity proxies — but "slightly higher" than 5.5% is still a very small denominator.

Seventy-three percent of AI projects fail. Six percent of organizations are high performers. Twenty-nine percent of manufacturers have deployed AI past the concept stage. One private equity case study claims 280 basis points of EBITDA improvement.

The headline "AI has landed on the factory floor" is not wrong. It's incomplete in the way that makes it dangerous. AI has landed. It's just sitting on the tarmac for most operators, surrounded by dirty data, unmigrated legacy systems, operators who don't trust false-alarm-heavy models, and management teams who haven't redesigned the workflows that AI is supposed to improve.

The World Economic Forum's Lighthouse Factory program — the gold standard for advanced manufacturing — shows what success looks like when all the prerequisites are met: 40% labor productivity boost, 48% lead time reduction, 66% defect rate reduction at Beko's clinching process. But Lighthouse factories are, by definition, outliers. They exist to prove that the ceiling is possible. They don't describe the floor.

The Investor Implication

Private equity's AI value creation thesis is premature, not dead. The mechanics are sound in specific functions — predictive maintenance, quality control at scale — and the one documented case study shows the EBITDA expansion is achievable. But treating it as a portfolio-wide lever that delivers 200–300 bps of margin improvement across industrial holdings confuses what's been proven in a single site with what's been proven across a network. It doesn't.

The cross-currents are:

• Directionally positive: Predictive maintenance and quality control have audited, repeatable ROI. These are not theoretical. They're plug-in applications with documented payback windows.
• Directionally negative: Data readiness is the single biggest barrier. Manufacturers with pre-2015 equipment, inconsistent part numbers, unreliable downtime codes, and tribal knowledge trapped in veteran operators' heads will not get AI ROI. AI amplifies existing mess; it doesn't clean them.
• Directionally uncertain: The PE multiple arbitrage play — buying at 8–10x and exiting at 15–20x by reclassifying an industrial business as software-adjacent — is a valuation bet that depends on LP discipline holding. It won't, if every firm makes the same pitch.

The headline claims AI has landed. It has. The question isn't whether it touched down. It's whether it's producing lift — or just weight.