Two years ago we asked what AI would do to 80 companies.
The answers came back in the future tense — can optimize, promises to reduce, is piloting. We pulled the data two years later. This is what actually happened — verified company by company, with the vendor hype stripped out.
Two hosts read the whole study, and argue with it.
Eight industries, eighty case studies, and the line between what an operator actually disclosed and what a vendor claimed — in nine minutes.
AI crossed the line from pilot to operating standard.
In every one of the eight verticals, the leading firms have moved AI out of the lab and into the run-the-business layer. Adoption in the most data-mature sectors now sits above 70%.
The gains are real — but they are bounded, use-case-specific, and unevenly evidenced, not the blanket transformation vendor marketing implies. So we did the harder thing.
We separated what the operators disclosed from what the vendors claimed.
The eye-catching numbers almost always trace to a vendor projection or a secondary blog — not the company's own books. The operator-verified figures are still strong. They're just materially lower. Knowing which is which is exactly the advantage a sophisticated buyer demands.
What the hype said
What the operators actually disclosed
The durable winners don't have a better algorithm. They have a data moat no competitor can copy.
Cross-industry pattern № 2
The same structural lessons recur across all 80 companies.
Predictive maintenance is the universal first win
In every asset-heavy industry it's the highest-ROI, lowest-risk entry point — converting unplanned downtime into scheduled work, with verified 20–50% downtime reductions. The most reliably evidenced use case in the entire dataset.
The data moat beats the algorithm
UPS, C.H. Robinson, Caterpillar, Lemonade win on proprietary, high-volume, well-structured data history — not a uniquely clever model. Competitors copy the algorithm; they can't copy the data.
2024–26 is the pivot to agentic action
The leading edge is no longer AI that recommends — it's AI that acts: C.H. Robinson's 30+ agents (3M+ tasks), AIG's Underwriter Companion, ambient clinical scribes. The next wave of leverage is concentrated here.
Vendor numbers systematically overstate
The headline figures almost always come from vendor projections or secondary blogs, not operator disclosures. Build any business case on the conservative, operator-disclosed figure — it's still strong enough.
AI improves the average case, not the tail
Lemonade still absorbed ~$45M in 2025 California wildfire losses despite best-in-class AI. AI compresses routine cost and variance; it does not erase the catastrophic, low-probability events that define an industry's economics.
What this means for you
Lead with operator-disclosed figures. The flagged, unverifiable items are precisely the numbers a sharp buyer will challenge — surfacing them first is the credibility move.
Where the value actually landed.
Each line below is the strongest operator-disclosed evidence in that vertical — the defensible ROI anchor, not the marketing number.
Transportation & Logistics
UPS ORION: ~100M miles & $300–400M/yr saved. C.H. Robinson: +40% productivity (shipments per person), headcount down ~10% as volume grew.
Energy & Gas Pipelines
Equinor: $130M+ in AI savings in 2025. bp: 12 exploration discoveries in 2025 attributed to AI, including its largest find in a quarter-century.
Medical & Healthcare
Cleveland Clinic Palantir command center: 40% OR idle-time ↓. Mass General Brigham ambient docs: 21.2% burnout reduction (JAMA Network Open, Aug 2025).
Agriculture
John Deere See & Spray: 31 million gallons of herbicide saved across 5 million acres in 2025.
Industrial Manufacturing
Caterpillar Helios: 50B data points/month from 1.6M connected assets. Danfoss: 80% of purchase orders automated, ~$15M/yr saved, 6-month payback.
Insurance
Lemonade: LAE ratio 13% → 7%. Allianz: £174M in fraud savings, 2025. AIG Underwriter Companion: 50%+ triage time cut.
Supply Chain Management
Amazon: +10–20% forecast accuracy. Walmart: $86M waste prevented by Eden in 6 months (Walmart-disclosed).
Entertainment & Sports Emerging
Ticketmaster blocks ~200M bots/day. The thinnest published-outcome evidence of the eight — strong on fraud and pricing mechanics, light on disclosed ROI.
The companies didn't stand still.
A meaningful share of the original 80 had changed shape entirely — proof that a stale catalog is a liability, not an asset.
Ceased operations Oct 2023; tech acquired by Flexport (~$16M), later sold to DAT (~$250M, 2025). The brokerage no longer exists.
Sold by Deutsche Bahn to DSV in a €14.3B deal completed April 2025.
Now GE Aerospace, GE Vernova, and GE HealthCare. The AI/additive work sits in GE Aerospace.
IBM sold it to Francisco Partners for $1.1B (Feb 2024); no longer an IBM product.
Nov 2024 ransomware attack disrupted Starbucks, Morrisons, and Sainsbury's — the vertical's defining cyber-resilience lesson.
Merged with UFC to form TKO Group (Sep 2023); FY2024 revenue $2.8B vs. the original WWE-only $1.3B.
Rebranded from Schlumberger; now a full AI platform company (Lumi, Delfi) with $2.44B digital revenue.
Topgolf spinoff abandoned (60% sold to PE, ~$1.1B); MSG split from Sphere; 3M spun off Solventum; Emerson spun off Copeland; TK Elevator now private, pending KONE combination.
The recurring high-volume cost isn't training. It's inference — pricing every six minutes, scoring every claim, reading every scan.
The infrastructure read
The same five requirements show up in every vertical.
Synthesized across all eight industries, the binding constraints on AI value point to one architecture: modular, compliance-driven infrastructure delivered close to where the data is generated.
The training/inference split is industry-specific
Heavy GPU training concentrates in model-building. The recurring cost is inference at volume — separating the two maps directly to how these workloads actually bill.
Low latency matters in physical industries
Logistics, manufacturing, energy OT, and field agriculture all require processing near the data source. Centralized-only cloud is insufficient.
Regulation is the gating factor
HIPAA, the NAIC AI bulletin and EU AI Act, PHMSA safety rules — regulated verticals cannot adopt without governed, auditable, often on-prem/regional infrastructure.
Redundancy is now a procurement requirement
The Blue Yonder attack reframed concentrated SaaS dependency as systemic risk. Modular, redundant deployment is a selling point, not just an architecture choice.
Integration is the universal prerequisite
In every industry the binding constraint was clean, integrated, sufficiently long-history data — the step that most stalls adoption.
This is the thesis the data validates
Modular GPU training, inference close to the data, a compliance-grade monitoring layer, and a data-ingest foundation — built once, deployed where each industry's constraints demand.
Every figure here is a citable, current proof point.
Pair the verified outcome with the matching infrastructure requirement when you walk into a vertical. Lead with the operator-disclosed number. Let the rigor do the selling.