IBM CEO Krishna Was Half-Right About AI

86% of CEOs say their workforce has the skills to collaborate with AI. Only 25% of that workforce actually does. That single gap, buried in IBM's new CEO Study, tells you more about why "AI operating model" will stall than any of the headlines from last week.

IBM CEO Arvind Krishna said the quiet part out loud last week. AI is not helping your business. It is your business. The WSJ ran it. LinkedIn caught fire. Other leaders I know are having the same "we should talk about this" conversation. 

He's directionally right. And at the same time, also missing half of the sentence that decides whether any of it works.

A phrase that's about to calcify

I want to flag something before it sets. "AI operating model" is heading toward the same fate as "digital transformation." A power phrase that enterprises often spend years attempting to define for themselves: hiring consultants, building steering committees around, putting on a roadmap, and often never operationalizing.

We've run this play before. Twice actually. "Digital transformation" took a lot of investment and budget before anyone asked what got transformed. "Data-driven enterprise" had a similar arc.

Both phrases described real shifts. Both became scaffolding for inertia.

The reason isn't a vision gap. It's that each phrase assumed the institution underneath could absorb what the new layer produced. Each time, it couldn't. Not because the people were slow. Because the infrastructure was wrong for the speed being asked of it.

AI is about to do the same thing, faster, with sharper consequences. Which brings me to the new IBM CEO Study, because the data inside it is more interesting than the headline.

What the numbers say

IBM surveyed 2,000 CEOs across 33 geographies. The findings that matter, and everyone will quote:

• 76 percent of organizations now have a Chief AI Officer, up from 26 percent a year ago.
• 79 percent of executives report they're decentralizing decision-making to absorb AI's expanding role.
• 64 percent of CEOs say they're comfortable making major strategic decisions on AI-generated input.
• By 2030, CEOs expect 48 percent of operational decisions, where guardrails can be codified, to be made by AI without human intervention. Today that figure sits at 25 percent.

And the stat that should stop the room. Only 25 percent of the workforce uses AI regularly. 86 percent of CEOs believe their employees already have the skills to do it.

Read those last two together. A title is being installed faster than a function being built. Decision authority gets pushed outward without the connective tissue to carry it. Strategic confidence in AI inputs keeps rising while the workforce barely touches the tools. Capability is present. Deployment isn't. The gap widens as the org chart races ahead of the operating reality.

Every one of these numbers is a coordination problem dressed up as a technology problem.

Where Cohn's argument stops

Gary Cohn writes in the foreword of the IBM report, “Decision cycles will compress. Boundaries between functions will dissolve. Advantage will accrue to those who can learn, adapt, and execute faster than their competitors.”

He's correct. IBM’s Vice Chairman is also describing the conditions under which an unprepared institution falls apart faster, not the conditions under which it wins.

Compressed decision cycles only produce advantage if the institution can receive the right information, directed to the right person, at the right time to decide. Dissolved functional boundaries only produce coherence if signals can actually route across them. Neither is a software problem.

Cohn calls CEOs "architects of intelligence." Great phrase — expect it to be reused on stage for the next eighteen months. I'd push it one notch further. The CEO is the architect of what the enterprise pays attention to. That's the job underneath the job. Model selection, agent design, workflow embedding, all of it sits underneath that single question.

The resource being allocated

Every operating model that's ever existed, industrial, digital, AI, is fundamentally a way of allocating one finite resource across the enterprise.

Attention.

Where leaders look. What gets escalated. Which signals trigger a decision and which decay into noise before anyone who could act on them ever sees them.

This isn't a metaphor. It's the actual mechanics of institutional response time.

Forty years of enterprise software has been organized around the opposite premise, that the bottleneck is data. So we built systems of record to store it. Systems of engagement to surface it. Dashboards to organize it. Then AI to interpret it. Each layer was sold as the answer. Each layer made the underlying problem worse, because each new tool produced more outputs the institution had to act on without changing the institution's capacity to act.

That's the layer the IBM study circles without naming. When 79 percent of executives say they're "decentralizing decision-making," what they're actually saying, whether they realize it or not, is that the center can no longer process the volume.

Decentralization isn't a strategic choice here. It's an overload response.

What an AI-first enterprise actually looks like

The IBM study makes the gap legible. 85 percent of CEOs say every functional leader must become a technology expert in their domain. 77 percent say talent and technology leadership roles are converging. 83 percent say AI success depends more on people's adoption than on the technology itself.

Translate that out of CEO-speak and you get something messier. The work is now happening at the seams. Between functions. Between humans and agents. Between the moment a signal appears and the moment someone acts on it. The functions themselves aren't the unit of competition anymore. The handoffs are.

A real AI-first enterprise isn't one with the most agents. It's one where the seams have been engineered. Where attention is treated the way capital is treated. Measurable. Allocable. Governed. Where a signal that matters can reach a decision-maker before it decays. Where the institution itself has memory, so the same problem doesn't get rediscovered every quarter by a different team.

We call this Systems of Attention (SoA). A new category of technology, SoA is the fourth layer of enterprise software, following Record, Engagement, and Insight. Attention Infrastructure is what Signal Labs builds within that category.

What the layer actually does

Underneath Systems of Attention sits a technical architecture we call Attention Infrastructure. Inside it, and well-explained in our white paper, Trust Zones handle the part of the problem the IBM study gestures at but doesn't solve. How confidence travels with a decision.

When 64 percent of CEOs say they're comfortable making strategic calls on AI-generated input, the next question is the one nobody is asking. Comfortable based on what? An AI output without provenance is a hunch with better packaging. An AI output with traceable confidence and audit lineage is a decision the board can defend.

Trust Zones is how the second one becomes possible at scale. The white paper lays out the mechanics. What I want to flag here is the implication.

The 48 percent autonomy figure IBM projects for 2030 is not achievable, in any institution operating under regulatory scrutiny, without SoA. You cannot codify guardrails into autonomy without a substrate that carries trust along with the signal. The institutions that try will either pull back when the first audit lands. Or they won't pull back, and they'll find themselves in a much worse conversation.

The stat that sits with me

Of all the numbers in the IBM study, one keeps me up. 86 percent of CEOs believe their employees have the skills to collaborate with AI. Only 25 percent of those employees are actually doing it.

That isn't a training gap. That's an institution where capability exists, tools exist, willingness exists, and the work still happens the old way because the seams between people, agents, and decisions were never engineered. The signal isn't reaching the worker. The decision isn't reaching the surface. The output of one agent doesn't reach the next one downstream.

A Chief AI Officer can't fix that. Neither can a new operating model on a slide. You fix it by treating attention as the thing you're actually building and a key driver for the right person making the right decisions at the right time.

What Krishna got right, completed

Krishna is right that AI is the business. He's right that enterprises treating it as a layer of technology will lose to those treating it as a new way of operating.

The unfinished half: the operating model only matters when the institution can act on what it produces. That's not an algorithm problem. It's an architecture problem. It has been the architecture problem for thirty years. AI didn't create it. AI has just been stacked upon it, making the challenge more impossible to ignore.

Every CEO worth their salt knows something is missing. They're reaching for the right pieces. New roles, decentralized decisions, and redesigned workflows. The piece they keep reaching past is the one underneath. The layer that decides which signals get through, which decisions get made in time, which outputs the institution can actually absorb.

Build that layer, and the operating model writes itself.

Skip it, and three years from now we'll be reading the next CEO study, asking why the agents didn't scale.

I know which one I'd rather be reading.