What Every CEO Needs To Know About AI In May 2026

Agents Arrived. Immediate Use Cases and Security Did Not.

Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, compared with fewer than 5% at the start of this year.

Global spending on agentic AI is growing exponentially, and the low-hanging use cases are largely harvested by now. Summarizing documents, drafting emails, building slides has been the early harvest, and the productivity gains from these applications are real but also well understood and increasingly commoditized.

The genuine returns from agentic AI sit in deeper process automation, where multi-step workflows run across systems with minimal human intervention.

Those use cases are also precisely where the security infrastructure at most organizations significantly lags.

The numbers from the security side deserve serious attention.

Forty-eight percent of cybersecurity professionals now identify agentic AI as the single most dangerous attack vector in their organizations, according to recent surveys.

The average organization manages around 40 deployed agents according to Gravitee’s survey report, and more than half of those agents run without security oversight or logging.

The gap worth understanding is not between capability and cost.

What we need to keep a closer eye on is the gap between what agents can do and what your security teams can observe or govern. Eighty-two percent of executives report confidence that their existing policies adequately protect against unauthorized agent actions, which would be reassuring if it were connected to the actual state of controls. Based on the Gravitee 2026 survey data, it almost universally isn’t.

The real agent use cases, the ones worth building toward, are getting deeper and more domain-specific: customer support with full system access, supply chain decision-making, financial controls, legal document review.

Each of those requires genuine governance and integration work before the returns justify the investment. The executives who are doing that foundational work now are building an advantage over those who are waiting for the ecosystem to mature before engaging seriously.

The ecosystem, for reference, has already matured.

The Bill Is Coming Due

In 2024, Klarna announced that its AI system had replaced 700 customer service representatives and was saving an estimated $40 million annually. The announcement circulated widely as confirmation that large-scale AI workforce replacement had arrived.

By mid-2025, Klarna was quietly rehiring, because customer satisfaction had collapsed and complex issues were going unresolved at scale. The CEO acknowledged the company had “focused too much on efficiency.” The cost of unwinding the decision, accounting for severance packages and retraining, came to at least $15 million, which meaningfully reduced the net gain from the original move.

Klarna is the most publicly documented example of a pattern that is much broader.

Two in three employers that conducted AI-driven layoffs are already rehiring, in many cases within months of the original cuts, according to a Careerminds survey of 600 HR professionals conducted in February 2026. Fifty-five percent of employers who made AI-driven cuts now report regretting the decision. Forrester’s 2026 Future of Work report estimates that half of all AI-attributed layoffs will ultimately be quietly reversed, often at lower wages than the original roles commanded.

On the cost side, the economics are becoming harder to ignore in a different direction.

Uber’s chief technology officer burned through his entire 2026 AI budget from token costs before the year was half over.

Enterprises are now spending substantial amounts on AI tooling, while many still struggle to clearly measure what they are getting back from it. Across industries, leaders increasingly report that deployments are moving faster than the organizational systems needed to evaluate their actual value.

Part of the challenge is that the economics of AI are genuinely unusual. The underlying models keep getting dramatically cheaper and more capable at the same time. GPT-4 class performance, which once carried premium pricing, is now available at a fraction of the earlier cost.

On paper, that should reduce spending. In practice, lower costs tend to increase usage volume even faster. Teams experiment more broadly, deploy more agents, automate more workflows, and generate exponentially larger amounts of inference traffic once the constraint loosens.

The result is that falling unit costs and rising total spend can happen simultaneously, which is exactly what many organizations are now discovering. AI budgets that initially looked like contained pilot programs increasingly resemble cloud infrastructure curves from a decade ago, where usage expands quickly enough that efficiency gains fail to reduce aggregate spending.

At the same time, most organizations are still learning what meaningful ROI from AI even looks like. Traditional technology investments usually operate on relatively familiar timelines and metrics. AI deployments behave differently because they often reshape workflows, decision structures, and operating models rather than simply speeding up an existing process. Measuring value therefore becomes harder, slower, and more organizationally dependent than many executives initially expected.

The constructive interpretation is that the market is maturing.

The honeymoon phase where experimentation alone counted as success is ending, and companies are beginning to evaluate AI the same way they evaluate every other major investment category.

Efficiency claims increasingly need operational proof, pilot programs need measurable outcomes, and “AI strategy” is slowly giving way to the more difficult question of whether these systems genuinely change productivity, margins, or growth in durable ways.

Which takes us to our next trend.

The Tools Have Actually Arrived

Early AI models famously miscounted letters in words. Ask the original GPT how many Rs appeared in “strawberry” and it would tell you two.

This became a widely circulated marker of AI’s fundamental unreliability, a reason for sophisticated skepticism about what the technology was ever going to become.

The early errors were real but the inference from them was all wrong.

They were leading indicators of a steep improvement trajectory, the way the first photographs were blurry and the first automobiles were slower than horses, and the executives who concluded from those errors that the technology was fundamentally unserious are now operating with a mental model of AI capability that is several years out of date.

In May 2026, the frontier moves almost monthly.

Models now specialize rather than dominate universally. Some are better at coding, others at scientific reasoning, others at conversational fluency, speed, or cost efficiency. At the same time, open-weight and lower-cost competitors have compressed pricing so aggressively that capabilities once reserved for the largest enterprises are now deployable at mass scale.

What this means for executives is straightforward. There is no longer a single “best AI model,” and there is certainly no permanent answer to which provider an organization should standardize around indefinitely. The market now behaves more like cloud infrastructure or semiconductors than traditional enterprise software. Different tools perform better on different workloads, and the landscape changes fast enough that locking into a single-model worldview becomes strategically dangerous.

The deeper implication is organizational rather than technical. Companies that still approach AI as a one-time procurement decision are treating a rapidly evolving capability layer as though it were static enterprise software. It is not. The firms pulling ahead are increasingly the ones building adaptive systems around AI rather than betting the company on a single vendor, benchmark, or moment in time.

The question is whether your governance frameworks have caught up with what your developers are already doing.

The Fluency Gap Is Becoming a Performance Gap

The most consequential AI story receiving too little boardroom attention has very little to do with models, agents, or token costs.

It has to do with the widening internal gap between the people inside organizations who have learned to work seriously with AI and those who have not.

The productivity gains themselves are real, but they are unevenly distributed. People with strong underlying judgment, domain expertise, and technical fundamentals tend to get dramatically more leverage from AI systems because they can evaluate outputs, spot weaknesses, and integrate the tools into real workflows. Others see far smaller gains because the technology amplifies existing capability more effectively than it replaces it.

The result is that high performers inside organizations are often becoming substantially more productive while others remain relatively flat, and that divergence compounds over time.

This creates a problem many leadership teams still underestimate. AI adoption alone tells you almost nothing about organizational readiness. Most large companies now have employees using AI tools in some capacity. The more important question is who inside the organization is generating disproportionate value from them, whether leadership knows who those people are, and whether the company has any systematic way to spread those capabilities.

Right now, AI fluency is spreading through many organizations through individual curiosity rather than organizational design. The employees extracting the most value are frequently self-directed experimenters operating far ahead of official workflows, training programs, or governance structures. Meanwhile, formally designated AI initiatives sometimes produce far less operational impact than expected because the people leading them are not necessarily the ones pushing the tools hardest in practice.

From the outside, this can look like broad organizational adoption. Internally, it often produces widening capability gaps that eventually surface in productivity, execution speed, and strategic adaptability.

A CEO’s practical agenda in May 2026 comes down to four questions, one per trend.

Where are your agent deployments, and is the security infrastructure anywhere close to adequate?

Where are your AI investments against the ROI timeline, and are the ones that aren’t performing being treated honestly?

Which models are actually being used for which tasks, and does anyone in the organization know?

And how many of your people have built real fluency with these tools, versus performing adoption for an audience?

The window for serious engagement with this technology is open.

It is also the same window it was at the start of the year, and the start of last year, in the sense that every quarter that passes without clarity on those four questions is a quarter of compounding disadvantage.

The original Luddites were not wrong about what was changing.

They were wrong about whether they still had time to engage with it on their own terms.

The executives reading this still do.