Not More AI. Different Infrastructure.
IBM published a number last week that most people scrolled past.
Only 25% of enterprise AI initiatives deliver expected ROI. Only 16% have achieved enterprise-wide scale.
Think about what that means. These are large companies with serious budgets, access to the best models, dedicated engineering teams, executive sponsorship, and years of runway to figure it out. Three out of four aren’t getting the returns they expected.
The obvious explanation — that AI doesn’t actually work — is wrong. The tools work. The models are genuinely capable. The failure is happening somewhere else.
What IBM got right
At Think 2026, IBM’s CEO Arvind Krishna said something that cut through the product announcement noise:
“The enterprises pulling ahead are not deploying more AI — they’re redesigning how their business operates.”
That’s not a sales pitch dressed up as insight. It’s an accurate diagnosis of why the 75% are failing.
The companies that can’t close the ROI gap are not missing better models. They’re missing a coordination architecture — a system that connects AI agents, data, automation, and infrastructure into something integrated. Without it, you have a collection of capable tools that don’t know what each other is doing. Every tool executes independently. No tool understands what matters, what changed, or what the other tools just did.
IBM calls this the AI Operating Model. The frame is corporate, but the idea underneath is right: execution capacity and coordination infrastructure are different things. You can have unlimited execution capacity — every task automated, every query answered, every document generated instantly — and still have a coordination failure. The tools don’t talk to each other. The context doesn’t transfer. The loops stay open.
That’s not a technology problem. It’s a systems design problem.
The divide that’s already happening
IBM called this the AI Divide — a widening performance gap between enterprises that have built coordination infrastructure and those that are still treating AI as a collection of individual tools.
The split is already visible in the data. The companies investing in the full operating model — the ones connecting their agents, data flows, automation, and infrastructure — are pulling ahead of the ones running point solutions.
The ones running point solutions have more AI than they did two years ago. More capable AI. And they’re still failing at the same rate.
This pattern has a logic to it. When you add execution capacity without adding coordination infrastructure, you don’t solve the bottleneck. You move it. The bottleneck was intelligence. AI cleared that. The new bottleneck is coordination: who decides what to execute, in what order, with what context, and who confirms the loop actually closed.
More AI tools means more execution happening faster. Without the coordination layer, it means more open loops, more context that doesn’t transfer, more gaps between what the agent did and what the situation actually required.
IBM is building the infrastructure that makes coordination possible at enterprise scale. That’s what the operating model is. That’s what watsonx Orchestrate and IBM Concert are actually for.
The layer they’re not building
Here’s the part that matters more than IBM’s product announcements.
The AI Operating Model IBM is selling works within IBM’s infrastructure stack. Google’s version of this — Remy, Workspace Agents — works within Google Workspace. Microsoft’s version works within M365. Every major platform is building the coordination layer for their own ecosystem.
None of them are building it for the person who crosses all of them.
The average knowledge worker doesn’t live inside one enterprise ecosystem. They work in Gmail and Slack and Notion and Linear and Salesforce. They have a calendar that lives in one tool, commitments that were made in another, follow-ups that need to happen in a third. The loop opened in Slack. The context lives in email. The decision needs to get into the project management tool.
That’s not a Microsoft problem or a Google problem. It’s the coordination failure that happens at the seams — between platforms, between threads, between the place where a decision was made and the place where the person who needs to act on it will look.
The enterprise operating model doesn’t reach there. It was never designed to.
The personal version of the same problem
IBM’s diagnosis is correct. The AI divide is real. Coordination infrastructure is the difference between AI that compounds and AI that stalls.
But the divide isn’t only happening at the enterprise level. It’s already personal.
Knowledge workers who have a system — something that holds context across tools, tracks open commitments, surfaces what needs attention — are operating differently than those who are manually coordinating across twelve apps. The gap is the same as the enterprise gap. It just doesn’t have a $100M IBM implementation behind it.
The enterprises that get the operating model right will pull ahead. The evidence is in IBM’s own research: the 25% who see ROI are the ones with coordination infrastructure.
The question is what happens at the individual level as this plays out. IBM is solving for the company. Nobody has solved for the person who lives across companies — across tools, across ecosystems, across the seams that enterprise AI infrastructure was never designed to bridge.
That’s the version of the AI divide that’s still open.
IBM Think 2026 was held May 5, 2026. IBM’s CEO study data and Arvind Krishna’s keynote remarks were reported by IBM Newsroom.
