Ask a CIO to list every AI system running in production and you will usually get a confident answer. According to a study released on November 19, that answer is probably missing close to half the picture. Humatica, an operating model advisor, studied 252 enterprises for its report AI-Native by Design, sponsored by CoreWeave, and found that 42% of production AI use cases started in the business rather than in IT or a central AI team. Nearly one in four (24%) was started by an employee or team on their own initiative.
That would be a curiosity if those systems were side projects. They are not. The same study found that two in three enterprises (66%) now run a process, product or business model that would degrade or break if AI were switched off. Yet only about a third (35%) have a formal operating model for AI. The enterprise AI operating model, in other words, is lagging far behind the AI it is supposed to run.
What the Humatica study found
The research combined a panel of 252 listed, private equity-backed and family-owned enterprises across 14 industries in North America, Europe and Asia with more than 30 in-depth interviews with senior executives and investors. Panel companies were profiled from more than 1,900 public documents, and AI labs, AI and cloud vendors, startups and companies with under $50 million in revenue were excluded. The work ran from July to October 2026.
Three findings stand out for anyone responsible for AI spend or risk. First, dependency has arrived before structure: AI is load-bearing at 66% of enterprises, but formal operating models exist at only 35%. Second, leaders named their own organization as a constraint on AI 2.4 times as often as the technology itself. Third, AI is being built everywhere. The report’s own summary is blunt: building AI has gotten easier, but the work of owning it, which it describes as integrating, securing, monitoring and paying for it, has not gotten smaller.
Humatica partner Andy Knobloch put the ownership problem in one sentence: “Almost half of what runs in production started in the business, often with no sponsor and no budget.” Read that again from a finance or security seat. Systems that the business now depends on were launched without a sponsor, and in many cases without a budget line anyone in finance would recognize.
Why business-led AI breaks the old operating model
For most of the last decade, the enterprise playbook for new technology assumed a central gate. IT evaluated tools, procurement negotiated contracts, security reviewed access, and finance tracked the spend against a cost center. That model worked because building software was hard enough that most of it had to go through the gate.
Generative AI and agents removed that friction. A sales operations lead can wire a model into a CRM workflow in an afternoon. A finance analyst can stand up an agent that reconciles invoices using a personal API key. A marketing team can buy a seat-based AI tool on a corporate card. Each of those steps is reasonable on its own, and many of them create real value. The problem is what happens a year later, when the workflow has become critical and nobody can answer three basic questions: what does it cost, what does it return, and what is it allowed to do?
This is where business-led AI and shadow AI start to overlap. Not every use case that begins outside IT is unsanctioned, but every one of them begins outside the systems IT, security and finance normally use to see what is running. The 24% started by individual initiative are the sharpest version of the problem, because they often run on personal accounts, scattered API keys or free tiers that later convert into paid usage.
Data readiness is the other half of the gap
A second study released the same day points to why ownership matters for value, not only for control. Accenture’s research on AI-ready data, based on a survey of 2,000 C-suite executives and their direct reports at 2,000 large organizations across 15 countries, found that only 7% of organizations, which it classifies as Reinvention Ready, are leading in developing AI-ready data. More than 80% of executives said they delay, limit or alter AI or generative AI initiatives because of data.
The value gap is stark. Three in four Reinvention Ready organizations report measurable value from their AI-ready data, compared with just one in four of all others. Accenture also found that, on average, 40% of companies are leading at foundation capabilities, 23% at contextualization and only 19% at value realization. The pattern mirrors the Humatica finding: organizations are better at building the base than at connecting it to business outcomes they can measure.
Put the two studies together and the operating model question becomes concrete. If close to half of production AI starts in the business, and most organizations struggle to connect AI to measurable value, then the central team cannot just approve or block use cases. It needs a way to see every AI system, wherever it started, and judge each one on the same evidence. That is the signal Olakai was built for: a vendor-neutral system of record for AI that records what each tool, model and agent does, what it costs and what it returns.
What a working enterprise AI operating model needs
Humatica frames its answer around five practices, including empowering decentralized building and scaling proven results, and keeping every production system owned, measured and improving. Translated into day-to-day operations, the organizations that close the gap tend to put four things in place.
- An inventory that starts from usage, not from intake forms. Asking teams to self-report AI use misses exactly the 24% that started on individual initiative. Discovery has to come from observed activity across tools, APIs and agents, which is what complete AI monitoring is for.
- A named owner and a budget line for every production system. When a business-built workflow crosses from experiment to dependency, it should get a sponsor, a cost center and a clear outcome metric, so the next renewal is a decision rather than a surprise.
- One measurement standard across central and business-led AI. Cost per outcome, adoption and business impact should be compared the same way whether a use case came from the AI team or from operations. Agents in particular need visibility into tool calls, retries and spend, which is the focus of Olakai Agentic.
- Guardrails that scale with dependency. Policies should tighten as a system becomes critical, with AI governance applied to what agents can access and do rather than only to which tools are approved.
The measurement point is where most programs stall. This week we covered SnapLogic’s finding that 54% of enterprises have abandoned AI projects over unclear business value. Business-led AI makes that risk worse in both directions: valuable projects get cut because nobody can prove what they deliver, and weak projects survive because nobody is looking at them at all. In one engagement we supported, splitting value by agent showed that four agents carried about 70% of the measurable return, a pattern that is invisible without a portfolio view.
Advice by role
For heads of AI and CIOs, the job is shifting from building most AI centrally to running a platform and a standard that business teams build on. Your credibility will depend on being able to show the full portfolio, including use cases you did not launch. For CFOs, the priority is to bring business-built AI into the same cost and return discipline as any other investment; the AI ROI measurement questions do not change because a system started in a department rather than in IT. For CISOs, the 24% started on individual initiative are the place to look first, since they are the most likely to run on personal keys or accounts outside your controls.
In each case, Olakai gives the role the same underlying record. AI leaders see adoption and impact across every team, finance sees spend and unit cost by system and owner, and security sees which tools and agents touch which data. Because Olakai is vendor-neutral, that record covers whichever models, assistants and agent frameworks business teams chose, not only the ones procured centrally.
Owning the AI you depend on
Humatica’s conclusion is that the enterprises pulling ahead are not the ones with the most pilots but the ones that decide who owns each AI system, where it runs and how it gets better. That is an operating model decision, and it cannot be made without visibility. Organizations can build that visibility now, with Olakai’s Enterprise AI Intelligence Platform as the measurement layer across the entire AI stack, or they can wait until a business-built system that nobody owns becomes the incident, the budget overrun or the project that gets cancelled for lack of proof. To see what a full inventory of your production AI looks like, schedule a demo.
