The most useful question about any AI deployment isn’t “what’s the technology?” or even “what’s the business case?” It’s: what made this one actually happen?
Internal roadmaps get written. Pilots get approved. Steering committees meet. And then, most of the time, the production deployment doesn’t come — because nobody inside the organization is forced to make it work. The deployments that do reach production almost always have a forcing function: something external, structural, or irreversible that makes the cost of not deploying higher than the cost of doing it.
Three cases from the past year, each with a different type of forcing function.
UPS - Supply Chain - From 21% to 90% of customs cleared without a human touch
When the U.S. de minimis exemption ended in August 2025, UPS suddenly had to file formal customs entries for shipments that used to flow through unexamined. In March 2025, UPS cleared about 21% of 13,000 U.S.-bound packages without manual intervention on a daily basis. By September 2025, the carrier cleared 90% of 112,000 daily packages with no manual intervention — a roughly 9x volume increase handled at higher automation rates. CEO Carol Tomé credited integrating Agentic AI into customs brokerage. The volume scaled because the regulation forced it to. What’s worth noting isn’t the percentage jump — it’s that a tariff policy change became the forcing function for production agentic deployment, not an internal roadmap. For logistics and trade ops leaders, the lesson is uncomfortable: regulatory shocks compress AI timelines more reliably than executive sponsors do.
Aviva - Insurance - £60 million from running 80 models against motor claims
Aviva’s motor claims operation runs on more than 80 AI models stitched together. The headline figure is a 23-day reduction in liability determination time on complex cases — the slow, contested files where adjusters spend most of their hours. Routing accuracy improved 30%, complaints fell 65%, and the company puts annual value generated from AI-driven claims optimization at £60 million. The interesting number is the complaints one. When claims move faster and land in the right queue the first time, customers stop calling — which is itself a measurable cost line most insurers underprice. The forcing function here wasn’t regulatory; it was operational pressure made visible. Complaints are a signal that the existing process is failing publicly. Once you can measure that cost precisely, the case for deeper deployment becomes hard to argue against. If you run a motor book and you’re benchmarking a single fraud model against incumbents, you’re already two architectures behind. The competitive moat is the orchestration of dozens of small models, not any one of them.
Bank of America - Financial Services - 15,000 advisors get AI inside the tools they already use
In April 2026, financial advisors at Bank of America’s Merrill Lynch and Private Bank units got access to AI preparatory aids integrated directly into the Salesforce CRM and Zoom software they already use. About 15,000 employees across Merrill and Private Bank will have access to the new tools in the first phase of the rollout. Note where the AI lives — not a new app, not a portal, but inside the software advisors already open every morning. That’s the forcing function: the existing tool stack. When the AI arrives in the tool you’re already using, adoption doesn’t require a decision. Inez Louzonis, managing director of Merrill Lynch, told American Banker that workflows have “already changed significantly in the last six months” alone and that further change will come in “another six months, 12 months, 24 months, 36 months.” The tell is “preparatory aids” — this is meeting prep and client research, not advice generation, which keeps it on the safe side of regulatory scrutiny. If you run advisor technology at a wealth competitor, the question isn’t whether to build an AI assistant. It’s whether your CRM vendor or your Zoom equivalent ships it before you do.
Throwback: American Express, 2006 — The Closed-Loop Advantage
In 2006, fraud detection was a rules engine. Most issuers ran static heuristics — flag transactions over $X, block foreign merchant codes, freeze the card and apologize later — accepting both the false positives that infuriated cardholders and the false negatives that bled the P&L. The industry consensus held that fraud was a cost of doing business, distributed across the network and priced into interchange. American Express disagreed. Because it owned both the issuing and acquiring sides of every transaction, it sat on something its competitors structurally could not replicate: a complete, closed-loop record of merchant and cardholder behavior, end-to-end. The contrarian bet was that this dataset, fed into machine learning models running in milliseconds, could turn fraud from a cost center into a competitive moat.
The build was patient. AmEx began deploying ML-based authorization in the mid-2000s, and by approximately 2010 had scaled real-time scoring to 100% of card transactions globally — every swipe, every online checkout, decisioned in under a few hundred milliseconds against models trained on the closed-loop history. The system did not replace human judgment so much as compound it: each caught fraud refined the next model, each false positive recalibrated the threshold. Through the 2010s, AmEx’s fraud loss rates remained consistently below the card network average — in some years materially so — while peers chased the same outcome with rule-sets that aged faster than the fraud patterns evolved. The investment paid back directly because, in a closed-loop model, the company eats its own fraud losses. The ROI was not a slide deck; it was a line item.
AmEx didn’t need a regulation or a volume shock to deploy at scale. Its forcing function was structural — baked into the company’s founding model since the 1950s. Owning both sides of every transaction meant the cost of fraud came out of AmEx’s own P&L, not a shared network’s. That alignment of incentives made the investment inevitable once the technology existed to act on it. The pattern worth carrying forward: whether the forcing function is a regulatory deadline, a measurable cost signal, or structural skin in the game, the companies that reach production AI deployment share one thing — they couldn’t afford not to.
A tariff change compresses six years of AI roadmap into six months. An operational cost signal — 65% fewer complaints — makes deeper orchestration the obvious next move. A tool you already open every morning removes the adoption decision entirely. AmEx had the forcing function built in from the start. The companies moving fastest now are finding theirs.