What’s next AI

 It feels as if every week brings another generative AI announcement. A new model. A new agent. A new promise to transform the way we work.

But the most important shift in 2026 is not simply that AI can produce better answers. It is that AI can increasingly take action: gather information, use tools, move between applications, and complete complex steps in a workflow.

That is an exciting change, but it calls for perspective. It does not mean we have reached Artificial General Intelligence (AGI), or that agents can reliably run entire businesses on their own. The question for leaders is more immediate: Where can AI execute valuable work today, and how do we remain accountable for what it does?

From Features to Workflows

Google’s Pixel ecosystem offers a clear glimpse of this shift in everyday technology. Driven by Pixel hardware updates and Gemini Intelligence on Android, devices are moving beyond answering passive questions toward proactive suggestions and multi-step execution across applications. Some capabilities run locally on-device; others draw on broader Gemini services. The direction is clear: our personal technology is becoming context-aware and inherently actionable.

Microsoft is pursuing the same shift in the enterprise. With Claude models generally available in Microsoft Foundry on Azure, enterprise customers gain access to frontier reasoning models within Azure’s established identity, billing, and governance framework. For leaders, the story isn't just about model options—it is about deploying actionable AI systems inside controls their organizations already trust.

Adobe is similarly turning documents and creative software into active workspaces. Following the launch of conversational PDF Spaces in Acrobat Studio, Adobe introduced the Productivity Agent in Acrobat to automatically transform complex, lengthy reports into audio summaries, interactive presentations, and visual reports. Paired with its expanding Creative Agent ecosystem, Adobe is coordinating multi-app creative execution across its suite.

There are even more ambitious experiments occurring at the frontier. Elon Musk’s vision for "Macrohard"—also referred to as "Digital Optimus"—combines Grok’s reasoning with an agent capable of navigating computer interfaces natively, as detailed in reports on the Tesla-xAI joint project. It serves as an illustration of where the market wants to go, rather than proof that AI can run an autonomous enterprise today. The broader strategic context has also changed, with SpaceX acquiring xAI and subsequently completing a record-setting IPO.

The ROI Reality Check

The capability curve is moving quickly, but the value curve is proving harder to bend.

In McKinsey’s 2026 State of AI survey, 40% of respondents at organizations generating over $1 billion in annual revenue reported scaling AI agents. Yet only 37% of respondents overall attributed any positive impact on their organization’s EBIT to AI—a figure essentially unchanged year-over-year. Conversely, at the individual level, 80% of respondents reported improvements in their individual productivity.

That gap between individual productivity and enterprise-wide financial returns may be the most important AI management challenge of 2026.

It reveals a critical lesson: Adding an agent to an existing process is not the same as improving the process. McKinsey found that nearly three-quarters of AI "high performers" fundamentally redesigned workflows to accommodate AI, compared with only about one-quarter of standard respondents. Sustainable returns come from rethinking how work gets done, not from accumulating isolated software pilots.

Consider an insurance claims process. An agent might retrieve relevant documents, highlight missing information, and prepare a preliminary draft for a specialist. The primary value isn't the draft itself—it is the streamlined, accelerated workflow that allows a human professional to retain oversight over coverage decisions, customer communications, and exceptions. That is the fundamental difference between automating a task and redesigning an operating model.

What Leaders Should Watch

Several core factors will determine whether the shift from assistance to execution delivers on its operational promise:

Agent Governance: As agents gain permission to use external tools and execute tasks, safety expands beyond model outputs. Following industry frameworks like OpenAI's agent builder safety guidelines and learnings from ChatGPT agent actions, organizations require clear permission boundaries, approval checkpoints, runtime monitoring, and rapid intervention and shutdown mechanisms.

Interoperability: Platforms highlighted at events like Google Cloud Next '26 are vying to manage agent deployments. Protocols such as MCP and Agent2Agent (A2A) Protocol—originally developed by Google and now governed under the Linux Foundation—are emerging as important building blocks for connecting tools and agents across platforms.

Cost and Infrastructure: Multi-step agentic execution requires significant computing capacity. Leaders must measure cost per useful outcome, rather than cost per prompt. Furthermore, macro infrastructure constraints—such as the International Energy Agency's energy projections on data center power demands—could increasingly influence long-term cost structures and infrastructure decisions.

Build vs. Buy: Nearly one-third of organizations report choosing to build at least one software product or feature rather than buy it, a shift being enabled in part by increasingly capable agentic coding tools. This shift will redefine enterprise software procurement—provided internal teams account for long-term maintenance, security, and operational overhead.

Workforce & Compliance: Expectations around AI-driven workforce adjustments remain high. At the same time, regulatory mandates—such as the EU AI Act and voluntary safety benchmarks like Anthropic's Transparency Hub commitments—make governance and transparency a matter of operational compliance.

The transition from assistance to execution is well underway. However, the defining question for leaders is not, "How much work can we hand over to AI?"

Instead, it is: "How can we redesign our work so that AI removes operational friction, our people exercise better judgment, and the organization delivers measurable results?"

That is where this revolution becomes real: not when an agent completes a compelling demo, but when it helps people deliver better outcomes—reliably, responsibly, and at scale.

Views expressed are strictly my own and do not represent my organization.