Humatica finds AI dependence outpaces formal oversight in 252-enterprise study
The CoreWeave-backed research separates using AI from being equipped to run it. Its sample includes only enterprises with active AI programs.
Loading page…
The CoreWeave-backed research separates using AI from being equipped to run it. Its sample includes only enterprises with active AI programs.
Listen to this story
Humatica and CoreWeave’s November 19 report points to an operating gap: many established companies depend on AI in core operations, but formal ownership and governance have not kept pace. In the 252-company, AI-active sample, only about a third had a formal AI operating model, and 20% had appointed a formal AI leader. The findings are not an economy-wide adoption estimate: the study excluded startups and AI vendors, among others. For operators, the report’s practical focus is assigning owners and putting shared safeguards around systems before others rely on them.
The sample included enterprises with active AI programs and more than $50 million in revenue; it excluded startups, AI labs, and AI and cloud vendors.
Organization and governance were cited as constraints by 68% of leaders, compared with 28% who cited technology and infrastructure.
Twenty-four percent of production AI use cases began without a sponsor or budget, creating a risk of systems with no clear owner.
AI has become essential to operations faster than enterprises have assigned responsibility for running it. In research published November 19, Humatica and CoreWeave found that 66% of 252 enterprises studied ran a process, product, or business model that would break without AI. Only about a third had a formal AI operating model.
The study, AI-Native by Design, combines profiles of enterprises across 14 industries with more than 30 interviews with executives and investors. Humatica examined more than 1,900 annual reports, filings, and investor documents between July and October 2026. Every enterprise had an active AI program, and the study excluded startups, AI labs, AI and cloud vendors, and businesses with less than $50 million in revenue. These findings describe an AI-active panel, not enterprise adoption across the economy.
Within that group, 95% had moved beyond individual experimentation. CoreWeave’s accompanying blog gives a more specific view of dependence: 44% had AI in the products they sold or in their business economics. Twelve percent could not deliver their offering at its current price or cost base without AI. Those measures go beyond counting users or trials.
The report defines an operating model as the arrangements for deciding what gets built, who owns it, and how it is funded and governed. It identifies four approaches, ranging from a central AI team to builders embedded across business functions. Each trades control against speed. Separately, just 20% of the enterprises had appointed a formal AI leader.
CoreWeave’s account of the research says 68% of leaders cited organization and governance as constraints on AI progress.
Technology and infrastructure were cited by 28%, putting the organizational constraint well ahead in the same comparison.
The ownership problem starts before a system becomes essential. The report found that 24% of production AI use cases began without a sponsor or budget. CoreWeave’s blog adds that 42% originated in the business, while one in four started through an individual’s or team’s own initiative.
Humatica describes systems that teams rely on but nobody clearly owns as “orphan services,” noticed only when they break. Its proposed response is not to ban decentralized building. Instead, it recommends giving employees a shared way to develop AI, with a stricter gate before others depend on the result:
The research also contrasts two reasons for controlling important AI workloads. Among leaders asked, 73% cited differentiation and proprietary intellectual property. Cost and predictability were cited by 36%, making them the least-cited reason. That finding measures leaders’ motivations; it is not a comparison showing that one infrastructure choice delivers better returns.
The authors recommend buying AI services for everyday work while keeping differentiating or sensitive workloads on infrastructure the enterprise controls. They also advocate routing work across models according to task, cost, and sensitivity. That recommendation comes from research produced with CoreWeave, which markets its own AI cloud infrastructure alongside the findings.
For executives, the report reserves another task beyond selecting technology: capturing expertise that was never written down. Its recommendation is to curate that knowledge, give it an owner, and build on it where the business competes.
Loading discussion...
Join the conversation
Explain where experimentation should end and formal responsibility should begin.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.