As you may have noticed from prior issues, I’ve been thinking a lot about the very human challenge of change management that is driving AI adoption (or non-adoption) inside organizations. This week, I want to bring some evidence into that conversation, because for the first time in mid-2026 we have a meaningful body of it — operational results from organizations that moved past deployment into something that looks like real scale. What that evidence shows reframes the change management conversation.
The companies that successfully integrated AI at scale solved an organizational problem, and the ones that got there in 2026 seem to share the same underlying pattern: they built infrastructure alongside adoption, not after it. Four of them published enough detail this year to see the pattern clearly: EY, Shopify, JPMorgan, and Klarna, across different industries, different starting conditions, and different scales.
What “At Scale” Looks Like in 2026
EY deployed across 300,000+ professionals and reports over 80% active usage. Shopify measured a roughly 20% productivity increase in engineering. JPMorgan is running 450+ AI use cases in production daily. Klarna’s AI systems handled the workload equivalent of 853 agents and avoided roughly $60M in annual costs by Q3 2025. These are operating numbers, not pilot projections, and they come with enough detail to understand how they got there.
What Each Organization Did
EY ate their own dog food. When they built out their agentic AI deployment, the internal diagnosis was eye opening: employees didn’t need another tool, they needed a comprehensive operating system for agentic work. EY built the OS first, then scaled adoption into it. They called themselves “client zero,” practicing internally before advising clients on the same path. The 80%+ training completion number was a condition of deployment. More classically enterprise-style, training happened before the tools went live, which is the opposite of how most organizations have sequenced AI rollout.
Shopify standardized the infrastructure layer underneath the tools. Rather than mandating a single AI platform, Shopify built an LLM proxy that routes all AI requests through one gateway giving teams freedom to experiment with different models while the company maintained centralized cost control and usage analytics. Shopify took queues from leadership on how to inject AI adoption into the existing culture of work: leaders shared openly how they use AI to solve specific problems, which drove organic adoption across engineering, sales, finance, and HR. The productivity gain followed from the infrastructure being in place to support and effectively monitor. The Bessemer playbook on how they did it published in April 2026 is worth reading in full.
The warning Shopify is tracking that almost nobody else in the 2026 case study literature has identified is really worth calling out here: comprehension debt. If engineers stop thinking deeply about the systems they’re building, they lose understanding of how those systems work. Shopify’s guardrail is that engineers must understand systems two to three layers below where they’re working. That is a governance decision, and it only gets made if someone is thinking about the long-term organizational consequences of AI adoption alongside the near-term productivity metrics.
JPMorgan’s most cited number is COIN, the contract intelligence system that saves roughly 360,000 lawyer hours per year. But the interesting organizational insight is in a less-cited metric: the 20% gross sales increase in wealth management after AI agents helped advisors respond to market volatility with personalized, portfolio-specific messages. JPMorgan built a machine for generating and scaling use cases while supporting a relatively high touch human-to-human client service experience, and that system-level posture is what produced 450+ production deployments.
Klarna is the most useful case study precisely because it has two chapters. The headline numbers from Q3 2025 were striking: $60M in annual cost avoidance, AI handling the equivalent of 853 agents. Then, by May 2025, Klarna walked back its AI-only customer service claim and reintroduced human agents for complex cases. CSAT degraded on edge cases, hallucinations were appearing, and confidence thresholds needed tightening. Klarna adjusted methodically in public, and the system improved. The companies operating AI at scale built feedback loops tight enough to catch degradation early and course-correct without a crisis, and this is precisely what operational maturity looks like.
The Pattern Across All Four
Stanford’s Enterprise AI Playbook, published in March 2026 by Brynjolfsson and colleagues across 51 case studies, put a name to what all four companies above demonstrate: same technology, same use cases, vastly different outcomes. The difference was in the organization — its readiness, its processes, its leadership’s willingness to treat adoption as a measure of organizational performance. The seven cases that reached organization-wide transformation all hit what the researchers called “strategic integration,” where the sponsor made AI adoption a core criterion of organizational success.
BCG’s 10-20-70 rule gives this a concrete cost structure that enterprises can model around: 10% of what it takes to get AI to perform at scale is technology, 20% is data and analytics infrastructure, and 70% is people and process. Organizations that follow this ratio outperform those that don’t by 3x on reported ROI. The ratio feels counterintuitive until you look at where every stalled AI mandate has stalled, and it is almost never in the technology layer.
What This Means For You
For individual contributors doing grassroots AI enablement: The Shopify comprehension debt warning is aimed at you. The ICs who are compounding fastest right now built a system around their AI use that amounted to much more than habit-building and included tools that connect to traceable decisions. The risk of moving fast and shallow is that you lose your own understanding of the systems you’re building. Go two to three layers deeper than where you’re working.
For leaders whose mandates have stalled: The infrastructure gap is the one worth diagnosing, and the three questions that matter are: do your people know what’s safe to share with AI systems, can they trace a decision that came from one, and does the system learn from what’s working? Those are infrastructure questions, and training doesn’t answer them.
For leaders who haven’t mandated yet: So far this year, the case studies being published offer valuable insights into how organizations large and small can successfully wade through the AI adoption swamp. The pattern across EY, Shopify, JPMorgan, and Klarna is that the organizations that scaled built the conditions for adoption to compound alongside the tools — governance, data readiness, feedback loops, and leadership behavior all going in together. You don’t need all of it designed in advance, but you do need to know that you’re designing it as you go.
Your Takeaway This Week
The change management conversation most organizations are having is focused on getting people to use AI. The 2026 evidence says that’s a partial diagnosis and organizations that scaled were asking a different question: what organizational infrastructure does AI need to perform here? That’s a governance question, a data question, a feedback loop question, and a leadership behavior question.
If you’re an IC: map how your AI use connects to your decisions, and whether anyone could trace that chain. That’s your OS question for the week.
If you’re a leader: pick one of the three infrastructure questions above — safe sharing, decision traceability, system learning — and find out whether your teams can answer it. That is usually where the stall is.
Leave a Reply