Why enterprise AI fails at the org chart, not the model layer

Let me describe an enterprise you might recognize. It's based on a client I've advised for the past 10 years. I'll call this Fortune 500 insurer Meridian (not the real name). They greenlit an AI initiative to read adjuster notes, flag fraud patterns and cut settlement times in half. The CIO hired a respected consultancy and, within four months, had a dazzling proof of concept running on clean, curated data.

Sound familiar? 

Then came production. The model needed claims data owned by the claims operations group, whose vice president saw no reason to prioritize an integration that wasn't on her roadmap. The fraud patterns required actuarial context guarded by an entirely different division. The adjusters, hearing rumors the system would optimize their roles, quietly stopped logging the structured notes the model depended on.

Eighteen months and seven figures later, the initiative was shelved and the postmortem blamed data quality issues. Nobody wrote down the real finding.

The org chart killed it.

I've watched this movie enough times to recite the ending. Meridian isn't an outlier. A preliminary 2025 report from MIT's NANDA initiative found that 95% of the organizations it examined were getting no return from their GenAI investments. Tellingly, the report attributes the failures not to model quality or regulation but to flawed enterprise integration.

I'll go one step further: In my work as an independent AI architect over the past five years, I have rarely seen enterprises demonstrate a clear return on AI.

Organizational readiness is too broad a diagnosis. The more specific problem here is an authority gap: The people accountable for an AI outcome often do not control the data, workflows, budgets or incentives on which it depends.

The model layer is rarely the problem

Here's the uncomfortable truth: The technology largely works. Foundation models are increasingly commoditized, reasoning continues to improve and APIs are widely accessible.

In 35 years, I've watched enterprises absorb client-server, the web, cloud and big data, and I can tell you the current failures don't look like technology failures. They look like organizational failures wearing technical costumes. The same pattern that killed your data warehousing program in 2005 and your cloud migration in 2015 is now killing your AI initiative.

Models can be rented. Data is entangled in everything that makes an enterprise what it is: its departments, politics, budgets and fear. When your AI initiative needs data sitting in five business units, each with its own definition of "customer" and its own incentives to hoard, no model can save you.

When a central team builds the pilot and reports to the CIO, but production requires cooperation from business units that report to someone else, you don't have a project. You have a negotiation. And when the model produces a wrong answer, and it will, nobody's performance review depends on fixing it.

As I used to say in my CTO days, when taking over the technology strategy for an entire company: "If I can't fire people or control their budgets, then I cannot change it." That principle shaped the authority I sought in executive roles throughout the 1990s and 2000s.

IT owns the tool. The business owns the process. Nobody owns the outcome.

The most common failure mode I see is the handoff problem. A central AI team builds something impressive in isolation, then tosses it over the wall to the business. The business unit didn't ask for it, doesn't understand it and has no budget line to maintain it. The data scientists move on to the next pilot. The application rots. In 2024, Gartner projected exactly this collapse point, predicting that at least 30% of GenAI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. Read that list again. Each one also raises an ownership question.

The second failure mode is misaligned incentives. AI-driven workflow redesign asks middle managers to absorb disruption, expose their data and accept head count pressure, all in exchange for benefits that land on someone else's dashboard. When your pilot stalls due to lack of engagement, understand that those managers are doing exactly what your company pays and rewards them to do. That's management's design failure, not their obstinacy.

The third is the accountability vacuum, where decision rights come in. Who arbitrates when two divisions disagree on a data definition? Who signs off when the model is wrong? Who authorizes the system to act on its own output? IT owns the tool. The business owns the process. Nobody owns the outcome.

Why buying more technology can't fix this

The reflexive response is to buy more tooling. A governance suite, a data mesh platform, an orchestration layer. This is normally where I get engaged. They want me to pick some magic tech that will get them out of this pickle.

I've watched enterprises spend hundreds of millions this way, and it doesn't work because the problem was never missing capability. It was missing organizational design. Tooling acquired faster than processes change just digitizes the dysfunction.

Let me render a verdict the vendors won't: if a vendor tells you their platform will fix your data governance, they're selling you a filing cabinet for a house with no rooms (the best analogy I could come up with).

Here's my overhyped list.

Agent platforms: Much of what's sold as agentic AI today is still closer to a rebranded chatbot.

Governance platforms: Governance is a decision you make, not a product you buy.

Enterprise-wide AI transformation programs: Change theater with a slide deck.

What deserves your attention is boring: decision rights, data ownership and workflow redesign.

What the winners do differently

Of course, I have a few clients on the right path -- very few. The enterprises consistently moving AI into production share three traits.

They assign a single accountable owner per use case, with budget and a P&L stake, embedded in the business unit. Remember my demand that I control the money and careers?

They fuse technologists and domain experts into cross-functional pods, eliminating handoffs rather than managing them.

They redesign the workflow before deploying the model, deciding in advance whose metrics change.

Now, the uncomfortable truth.

The central AI center of excellence (CoE), the structure most of you defaulted to, is one reason these failures happen. In my experience, the CoE as most companies run it is where AI goes to die.

It collects the best talent, builds disconnected demos, accumulates no business authority and gets reorganized the moment the board asks pointed questions. I'm not saying central capability has no place. I'm saying a CoE without embedded business ownership is a very expensive innovation theater troupe. I saw the CoE model fall short in cloud computing; I'm not sure why we're trying it again with AI, but we are.

Tooling acquired faster than processes change just digitizes the dysfunction.

Three things CIOs should do right now

First, renegotiate decision rights explicitly. For every AI initiative in flight, draft a one-page charter: who owns the data contract, who owns the model's performance, who decides when it acts and who answers for the consequences. If you can't name the accountable owner in the business, the initiative is already failing, and you just haven't seen it yet. You can start this Monday, peer-to-peer with your business unit leaders, without waiting for a CEO mandate.

This is where things go south quickly. People are sensitive about who controls what, and many leaders avoid the conflict that follows. If the process still depends on "influencing" or "convincing" people after decision rights are supposed to be settled, you still have a negotiation rather than an operating model.

Second, fix the incentive mismatch where you have leverage. Make adoption outcomes a stated success measure for the managers whose cooperation your roadmap depends on, with real change management budget behind it, not a training deck. If you can't change their KPIs, escalate or descope. A pilot nobody measures is theater for your board deck.

Third, stop launching pilots and move two or three existing ones into production inside redesigned workflows. The chasm between demo and deployment is where your AI ROI goes to die and closing it for a handful of high-value use cases beats ten more demos.

One last thing for your radar. As agents gain authority to act -- placing orders, approving transactions and modifying systems -- somebody has to decide what they're allowed to do and own the consequences. Agent permissions are becoming an operating model problem, not just a security problem. In my judgment, the leaders who figure out agent decision rights in the next 18 months will own a real advantage. The rest will be running this same postmortem in 2028.

The hardest problem in enterprise AI isn't inference. It's the org chart. I've watched this pattern play out across four technology waves now, and the same gap between accountability and authority keeps returning. Fix the chart, and the stack has a much better chance of taking care of itself -- trust me.

David Linthicum is a globally recognized thought leader, innovator and influencer in AI, cloud computing and cybersecurity. He has more than 30 years of experience in enterprise technology and, until early 2024, served as managing director and chief cloud strategy officer at Deloitte Consulting LLP.