Anchors and the Cautionary Cohort.
Section Eight · The Anchors and the Cautionary Cohort
Both readings are true. The discipline is reading which is which.
The empirical record on AI carries two truths simultaneously. AI works in production. The proof set is more than thirty years deep. And most contemporary AI claims do not survive to production. The cautionary cohort is dense, and concentrated in 2023 and 2024. The integrator's calibration discipline is to read which surface sits in which side of the timeline.
Pre-1995 anchors: production AI surfaces that have run uninterrupted at Plateau-grade scale for between thirty and forty-five years. 2018-2024 cautionary cohort: production AI surfaces that reached pilot or initial deployment and either collapsed, retreated, or drew regulatory action.
The four pre-1995 production anchors
Cognex shipped the first commercial machine-vision system in 1981. The lineage runs uninterrupted to the present and now sits inside every automotive manufacturing line at scale: dimensional inspection, weld-quality verification, surface-defect detection, parts identification. Forty-three years of continuous production. The technology generation has changed four times. The Plateau positioning has not.
American Airlines' yield-management programs — DINAMO among them — entered production through the mid-to-late 1980s, and the discipline they established became the foundation of modern airline revenue management. Roughly forty years later the discipline has spread from airlines to hotels, car rental, cruise lines, freight forwarding, and live-event ticketing. Every revenue-management vendor in production today traces its methodology to the DINAMO lineage: nested fare-class bid-price control, demand forecasting against constrained inventory, and dynamic re-optimization on the booking curve. The Plateau in pricing optimization runs through this anchor.
ABBYY, founded in 1989, launched FineReader OCR in 1993. Three decades later, OCR remains the deepest Plateau in Document AI. ABBYY FineReader, FlexiCapture, and the broader vendor cohort — Rossum, HyperScience, Kofax, and the GenAI-augmented stacks of 2024-2026 — all build against the OCR primitive that ABBYY commercialized. The Document AI Plateau in Section Four runs through this anchor.
FICO Falcon launched in 1992. Supervised machine learning on card-payment transaction features. The lineage runs uninterrupted across the global card-payment networks. Every Plateau reading on Fraud Detection in the Industry Matrix in Section Three — Banking, Fintech, Renewables and Utilities (smart-meter NTL detection), Outsource Call Center and BPO (voice biometric anti-fraud) — traces its methodology to the Falcon lineage. Thirty-three years of continuous production. The workstream's longest-running production AI surface.
IDeaS, acquired by SAS in 2008, consolidated the hospitality revenue management Plateau. The 2008 acquisition validated the durable production grade of pre-GenAI revenue management AI. Sixteen years of additional production lineage. The Plateau positioning in Hospitality on the use case matrix runs through this anchor.
The 2018-2024 cautionary cohort
The right side of the timeline carries the cautionary cohort. Twenty-five named operators, distributed across seven years, with two clear concentration points: 2023, which carried five collapses across healthcare, autonomous driving, freight, retail computer vision, and shipping blockchain; and 2024, which carried ten cautionary incidents across financial services, food and beverage, automotive, fraud, hospitality, and pricing.
The 2023 cohort was concentrated in three patterns. The healthcare-AI collapse cohort across 2022 and 2023 — IBM Watson Health divested in 2022, then Olive AI, Babylon Health, and Sensyne Health collapsing through 2023 — reflected the failure of the platform thesis in healthcare specifically: the surface is too fragmented across payor, provider, and patient to support a horizontal AI platform. The autonomous-driving retreats — Argo AI in October 2022, Cruise paused in October 2023 — reflected the gap between Level 3 demonstration and Level 4 production at scale in unconstrained urban environments. The retail-CV checkout-free retreats — Standard AI, Bossa Nova / Walmart, Amazon's Just Walk Out walk-back in April 2024 — reflected the gap between pilot-store demonstration and gross-margin economics at full-store scale.
The 2024 cohort was concentrated in four patterns, each tying to a specific failure mode from Section Seven. The GenAI customer-service retreats — Klarna, Air Canada, McDonald's-IBM — mapped to the hallucination-retreat failure mode. The fraud-attack incidents — the Hong Kong twenty-five-million-dollar deepfake CFO case — mapped to the AI-attack-outpacing-AI-detection failure mode. The pricing-action incidents — RealPage DOJ antitrust, Wendy's surge-pricing — mapped to the AI-pricing antitrust failure mode. And the platform-write-downs continued: Apple Project Titan, BenevolentAI, Biofourmis — mapped variously to industrial-platform overinvestment, healthcare-platform failure, and the underlying public-market discipline.
The empirical argument
Both halves of the timeline are real. The left half is the working empirical foundation for production AI. The right half is the cautionary cohort that any current AI program operates against. A senior executive reading the AI landscape gets both halves at once. The integrator engagement separates them: which surface in the client's portfolio sits on the left, which sits on the right, and which surface has not yet been positioned. The calibration is surface-specific and operator-specific.