Two Years Later — What AI Actually Did to 80 Companies

The central finding

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%.

Operators acting on live data across a wall of dashboards in a modern control room

72%

of logistics employees use AI tools at work

Open Sky Group

71%

of U.S. acute-care hospitals run predictive AI in the EHR

ONC / AHA

77%

of manufacturers had implemented some form of AI by 2024

Rootstock

70%

of large U.S. farms run at least one AI decision tool

IMARC Group

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.

The hard part — and the edge

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.

An analyst scrutinizing financial data on a screen

What the hype said

Shell~$400M / year in AI savings

Chevron$900M over 3 yrs · 50% drilling cost ↓

John DeereUp to 90% herbicide cut

Walmart60% waste reduction · $1B / yr

What the operators actually disclosed

Shell20–25% maintenance cost ↓; $2M in trips prevented at Pernis in the first two weeks

Equinor$130M+ in AI savings, 2025 (company-disclosed)

John Deere43–77% (~50–59% avg at scale); 31M gal saved across 5M acres

Walmart$86M waste prevented by Eden in 6 months (Walmart-disclosed)

Five patterns, regardless of vertical

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.

Eight industries, the verified ranges

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.

Golden-hour view where a truck highway, an industrial facility, and farmland meet

Transportation & Logistics

Route & network optimization · lifecycle automation

3–4% incremental routing → 40%+ labor productivity

UPS ORION: ~100M miles & $300–400M/yr saved. C.H. Robinson: +40% productivity (shipments per person), headcount down ~10% as volume grew.

A brown parcel delivery van on an optimized route at dawn

Energy & Gas Pipelines

Predictive maintenance · subsurface & exploration

20–25% maintenance cost ↓ · up to 50% downtime ↓ · up to 50% drilling cost ↓

Equinor: $130M+ in AI savings in 2025. bp: 12 exploration discoveries in 2025 attributed to AI, including its largest find in a quarter-century.

A gas refinery and pipeline complex with a flare at dusk

Medical & Healthcare

Ambient documentation · imaging · operational AI

14 min/day clinician time saved · 10–40% OR idle-time ↓ · 21% burnout ↓

Cleveland Clinic Palantir command center: 40% OR idle-time ↓. Mass General Brigham ambient docs: 21.2% burnout reduction (JAMA Network Open, Aug 2025).

A modern operating room with an ambient clinical-AI screen

Agriculture

Input reduction through precision application

43–77% herbicide ↓ (avg ~50–59% at scale) · +2.0 bu/acre yield

John Deere See & Spray: 31 million gallons of herbicide saved across 5 million acres in 2025.

A green precision sprayer with a wide boom sweeping a field at golden hour

Industrial Manufacturing

Predictive maintenance · vision QC · process optimization

20–50% unplanned downtime ↓ · 50–70% QC labor ↓ · 1–5% process efficiency ↑

Caterpillar Helios: 50B data points/month from 1.6M connected assets. Danfoss: 80% of purchase orders automated, ~$15M/yr saved, 6-month payback.

A large yellow heavy machine fitted with sensors on a worksite

Insurance

Claims automation · fraud · underwriting acceleration

LAE ratio ~halved · 3–9× claims throughput · 50%+ underwriting triage time ↓

Lemonade: LAE ratio 13% → 7%. Allianz: £174M in fraud savings, 2025. AIG Underwriter Companion: 50%+ triage time cut.

A phone showing an insurance claim approved in seconds outside a home

Supply Chain Management

Demand forecasting · inventory optimization · warehouse automation

10–20% forecast accuracy ↑ · 15% logistics cost ↓ · up to 35% inventory ↓

Amazon: +10–20% forecast accuracy. Walmart: $86M waste prevented by Eden in 6 months (Walmart-disclosed).

A fulfillment warehouse with orange robotic drive units carrying shelves

Entertainment & Sports Emerging

Dynamic pricing · fraud/bot detection · personalization

Fraud blocking at scale · per-seat yield uplift (not publicly quantified)

Ticketmaster blocks ~200M bots/day. The thinnest published-outcome evidence of the eight — strong on fraud and pricing mechanics, light on disclosed ROI.

A crowd entering a stadium through modern smart entry gates at night

Why a two-year-old catalog needs re-verification

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.

Figures moving through a bright glass corporate atrium, suggesting constant change

Convoy Defunct

Ceased operations Oct 2023; tech acquired by Flexport (~$16M), later sold to DAT (~$250M, 2025). The brokerage no longer exists.

DB Schenker Acquired

Sold by Deutsche Bahn to DSV in a €14.3B deal completed April 2025.

GE Split ×3

Now GE Aerospace, GE Vernova, and GE HealthCare. The AI/additive work sits in GE Aerospace.

The Weather Company Divested

IBM sold it to Francisco Partners for $1.1B (Feb 2024); no longer an IBM product.

Blue Yonder Breached

Nov 2024 ransomware attack disrupted Starbucks, Morrisons, and Sainsbury's — the vertical's defining cyber-resilience lesson.

WWE → TKO Merged

Merged with UFC to form TKO Group (Sep 2023); FY2024 revenue $2.8B vs. the original WWE-only $1.3B.

SLB Reinvented

Rebranded from Schlumberger; now a full AI platform company (Lumi, Delfi) with $2.44B digital revenue.

Topgolf · MSG · 3M · Emerson · TKE Restructured

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.

What it means for infrastructure

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.

01 · Compute

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.

02 · Edge

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.

03 · Compliance

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.

04 · Resilience

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.

05 · Data

Integration is the universal prerequisite

In every industry the binding constraint was clean, integrated, sufficiently long-history data — the step that most stalls adoption.

SAVRN

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.

Two years ago we asked what AI would do to 80 companies.