Press Release
The leaders are pulling away on cost, on growth, on margin. The gap is not a technology gap, it is an operating-model gap. Three failures explain the gap, and the leaders have designed around all three.
The financial impact of AI at the leaders is real and audited, Klarna’s ~$60M in compressed customer-operations cost, JPMorgan’s ~$2B in annual benefit, GitHub Copilot now used by approximately 90% of the Fortune 100, and the conclusion was blunt: the leaders are not using fundamentally different models; they are rebuilding the operating model around them. So why has capturing the value proved so much harder than the demos suggested? This article is the answer.
Start with the scale of the problem. PwC’s 2026 Digital Trends in Operations survey of 767 US operations leaders found that 89% say their technology investments have not fully delivered the expected results. That number matters because it describes the majority, not the laggards: most organisations have already spent, and most have not yet been paid back.
PwC’s 2026 AI Business Predictions puts the diagnosis bluntly: technology delivers only about 20% of the value of an AI initiative; the remaining 80% comes from redesigning the workflow around what the technology can now do. That ratio matters because it inverts where most budgets actually go: spending concentrates on the technology, the 20%, and starves the redesign work that produces the other 80%. The diagnosis comes in three parts.
Reason 1: procurement, not redesign
The most common pattern in the 80% is the one the first article named, a procurement: issue an RFP, evaluate vendors, select a platform, sign a contract, roll it out across an unchanged workflow. The technology gets installed. The KPIs do not move. The CFO is told the pilot "did not deliver expected returns" and the budget gets reassigned.
The leaders do something different, because they take seriously the first article’s refrain: AI is a business problem dressed up as a technology problem. Klarna’s ~$60M in savings did not come from buying an OpenAI subscription. It came from redesigning the customer service operation, new role definitions, new escalation criteria, new performance metrics, new training for the humans now doing harder work. PwC’s 2026 findings make the gap quantifiable: only 27% of operations leaders have fully embedded an AI strategy across business units, and only 37% are comfortable assigning AI agents to execute full end-to-end processes. Those are organisational readiness statistics, not technology readiness. A budget that includes the technology line and not the workflow redesign line is a budget designed to fail.
Reason 2: the arithmetic of agent chains
The second reason is mathematical, and it is the single most important thing to understand about agentic AI before approving any deployment that runs autonomously through multiple steps.
As the chain of tasks gets longer, small per-step errors compound into system-level failure. However smart the model, the math says that without a human checkpoint to reset the error rate, the autonomous chain will eventually break.
If a single step is 95% reliable, a ten-step chain is around 60% reliable. A twenty-step chain is about 36%. The error compounds. And unlike a human, an agent that goes wrong rarely knows it has gone wrong, it confidently produces a wrong invoice, a wrong recommendation, a wrong customer email, and moves on.
A demo of a five-step agentic workflow looks impressive. The same agent running twenty steps on real customer data, three months after launch, each step now slightly less reliable than it was in the demo, can quietly slip below 30% end-to-end reliability, and start causing more problems than it solves, often invisibly. The leaders design around this by building human verification into the chain at the right moments, keeping the chains short, and treating the probabilistic system as exactly that. These are not technology arguments. They are risk-adjusted return arguments, and they belong in a CFO’s conversation with the business owner before any agentic system reaches production.
Reason 3: a turbocharger bolted onto a stock car
The third reason is best explained by analogy. The turbo, on its own, is extraordinary, the car accelerates and sounds like a race car. But the engine internals were never designed for that combustion pressure. The suspension was tuned for the old power band. The brakes, sized for the original horsepower, fade before they can stop the new mass-times-velocity. The car is faster, yes, and quietly, a more dangerous machine than the one you started with.
A proper performance build upgrades the whole system, engine internals, brakes, suspension, tires, an engine management system that keeps everything in balance. Agentic AI is the turbocharger. The foundation model is breathtaking. But the rest of the car, data governance, access controls, verification layers, human-in-the-loop checkpoints, observability, escalation paths, was tuned for the old power band. If you do not upgrade those at the same time, the system will perform impressively for a while, and then fail under load.
Most of the AI disasters in the news right now are not stories about bad models. They are stories about turbochargers bolted onto stock cars.
The leaders pulling away on the financials are not better at AI than everyone else. They are better at upgrading the rest of the car at the same time as they install the turbo. That is what an "AI operating model" actually is.
For a leadership team, the useful exercise this article leaves behind is a diagnostic one: take your most recent AI initiative and ask which of the three reasons describes it, a procurement without a redesign, a chain without checkpoints, or a turbocharger on a stock car. Most organisations will recognise themselves in at least one. The final article in this series turns that diagnosis into a playbook: where to invest, how to govern, and how to measure, with a 90-day plan a CEO can authorise on Monday morning.
