IBM held its annual Think conference in Boston on May 5 and led with a statistic that should be uncomfortable for anyone who has approved an AI budget in the past two years: only 25% of AI initiatives deliver expected ROI, and only 16% have scaled enterprise-wide.
The number comes from IBM's own CEO study, but it rhymes with every survey running right now. A separate 2026 enterprise survey found that 97% of executives report benefits from AI but only 29% saw significant ROI from generative AI. MIT research from mid-2025 put the generative AI pilot failure rate at 95%. The specific numbers vary by methodology, but the shape of the data is consistent: AI is almost universally "useful" and almost never "profitable at scale."
This is the AI divide that IBM named its conference around. The gap between AI-first organizations — which the data credits with 70% greater productivity improvement and 74% faster cycle times than peers — and organizations still running fragmented pilots is real and is widening. IBM's diagnostic is that the problem is operational integration: governance bolted on too late, data too fragmented to feed agents at enterprise speed, no cross-functional ownership of AI outcomes.
"The enterprises pulling ahead are not deploying more AI — they're redesigning how their business operates," CEO Arvind Krishna said in Boston.
The product response is what IBM is calling the AI operating model: four pillars (agents, data, automation, hybrid cloud) held together by next-generation watsonx Orchestrate for multi-agent coordination, IBM Concert for intelligent operations, and IBM Sovereign Core for organizations that need to avoid dependence on US-hosted infrastructure. Bain, IBM's consulting partner for the Think narrative, put it as a transition from "AI pilots" to an "operating model" — the same framing IBM used internally for its own AI deployments, which it claims have saved $3.5 billion in operating costs since 2023.
The diagnostic is right. The prescription is complicated.
IBM's diagnosis is credible because the failure mode it names is real. Organizations do not have an AI problem. They have a governance problem, a data problem, and an accountability problem — and they have been trying to solve all three by buying model subscriptions.
A generative AI tool that sits on top of fragmented, siloed data cannot close a meaningful business gap. A proof-of-concept agent that has no designated owner, no success metric, and no integration with existing approval workflows will be abandoned after the demo. These failure modes are not exotic edge cases. They are the median AI initiative at a company that started AI investments in 2024 and is now explaining to its board why the ROI has not materialized.
The prescription is complicated because IBM is selling consulting services and middleware to fill a gap that most organizations would prefer to fill with internal talent and process. The watsonx stack is coherent on paper. It is also a significant vendor commitment for an organization that is still running AI initiatives as isolated experiments. Buying an operating model before you understand what you are operating is the enterprise equivalent of installing a factory scheduling system before you know your production plan.
The harder question is what the 5% of organizations achieving substantial ROI are doing differently. Based on the pattern across the IBM data, the Bain analysis, and the MIT research, the answer is not that they have better tools. It is that they assigned clear ownership before deployment, defined success metrics before deployment, integrated AI outputs into existing workflow approval chains rather than creating parallel AI-specific workflows, and treated data quality as a prerequisite rather than a future state. None of those actions require a vendor. All of them require someone in the organization to be accountable for the outcome.
That person, in most organizations, has not been named.
From an operator's vantage point: we build systems for clients who have lived the pilot failure cycle. The tell is always the same. The AI initiative has a champion (usually a CTO or a CDO who attended a conference), a vendor, and a demo environment. It does not have a business owner, a data source it can actually read, or a definition of success that a CFO would recognize. The model is fine. The organizational infrastructure around the model is missing.
The short of it.
IBM Think 2026 centered on the "AI divide" — data showing that 97% of executives see AI value, but only 29% see significant ROI and only 16% have scaled it enterprise-wide. IBM's diagnosis (fragmented data, late governance, no cross-functional ownership) is accurate. Its prescription (buy the watsonx operating stack) is selling a structural solution to what is fundamentally an accountability problem. The organizations beating the average are not using better tools — they named a business owner, defined a success metric, and plugged AI outputs into existing approval workflows before deployment.