At Davos this week, Demis Hassabis, the CEO of Google DeepMind, gave a standout talk about the path he believes AI will take in the coming years. While he thinks the tool will be utterly transformative, Hassabis still warned that the models have “these jagged edges that they’re good at and not good at.” Hassabis’s comments were a nod to AI pioneer Andrej Karpathy, who coined the term “jagged intelligence” to describe the uneven outputs from advanced models. In some domains, they are capable of superhuman performance while remaining incompetent in others.
Likewise, the adoption of AI models will be jagged. Real-world workflows are constrained by bottlenecks, where a single weak link can spoil the value of the entire process. Automating one task in isolation often does little, and can even backfire, if the remaining human tasks become more important, more time-consuming, or harder to automate. As a result, AI adoption is unlikely to proceed in a smooth fashion. Instead, firms may stall for long periods then jump in adoption once enough complementary pieces fall into place to clear the bottleneck.
One of the best surveys on these bottlenecks come from the Census’ Management and Organizational Practices Survey (MOPS), which Kristina McElheran analyzed for her talk at the Allied Social Sciences Association (ASSA) earlier this month. The two biggest constraints faced by businesses were the cost of AI adoption and finding business cases.
Still, this data comes from 2021: before ChatGPT went public. To better understand what’s currently happening, I collected a number of business surveys from the past year, which asked professionals about AI bottlenecks, use, and the productivity gains they’re experiencing from the tech.
These surveys should be taken with a grain of salt. They rely on self-reported data from executives who may have incentives to overstate adoption or returns. The samples skew toward large enterprises and many are produced by consulting firms that sell AI implementation services. On top of this, AI use is defined inconsistently across the board, making direct comparisons difficult.
That said, the gap between adoption and profitability is hard to dismiss.
window.addEventListener(“message”,function(a){if(void 0!==a.data[“datawrapper-height”]){var e=document.querySelectorAll(“iframe”);for(var t in a.data[“datawrapper-height”])for(var r,i=0;r=e[i];i++)if(r.contentWindow===a.source){var d=a.data[“datawrapper-height”][t]+”px”;r.style.height=d}}});
When looking across these surveys, a picture emerges. AI adoption is advancing rapidly in experimentation yet remains stubbornly uneven in deployment and value creation. Most firms now use AI in at least one business function, yet the majority are still stuck in pilots, with only a minority achieving enterprise-wide scale or measurable bottom-line impact. Returns take years to materialize and are especially elusive for more ambitious systems like agentic AI.
The profit gap is especially stark:
- MIT NANDA (2025): 95 percent have had no measurable P&L impact; only 5% are extracting significant value
- McKinsey (2025): more than 80 percent report no tangible impact on enterprise-level EBIT
- BCG (2024): Just 4 percent have developed capabilities that consistently generate significant value; another 22 percent are “beginning to realize substantial gains”
Many of the bottlenecks aren’t technical. MIT NANDA concludes that the divide between the 5 percent extracting value and the 95 percent that aren’t “does not seem to be driven by model quality or regulation, but seems to be determined by approach.” Fragmented and low-quality data, siloed IT architectures, and persistent difficulty measuring ROI are cited alongside broader organizational change as reasons for slow adoption. Still, inaccurate AI outputs have remained a key barrier on the technical side.
Human and organizational factors loom just as large. Workforce reskilling, change management, and redesigning workflows around AI are repeatedly identified as decisive for success. Meanwhile, leaders understand they are overestimating their readiness for AI. Some 63 percent of executives call AI implementation a high priority, but 91 percent don’t feel prepared to execute. There is also a startling attrition rate. S&P Global finds that 42 percent of companies have abandoned some AI initiatives before reaching production.
In short, AI adoption is jagged not because progress in models is uneven, but because production systems are. Until firms address the organizational, data, and governance constraints that bind their workflows, AI’s transformative potential will continue to arrive in fits and starts rather than as a smooth diffusion curve.
The post Jagged Intelligence, Jagged Adoption appeared first on American Enterprise Institute – AEI.
