When it comes to AI, 2024 was about possibility. In 2025, AI moved into enterprise workflows. So far, 2026 has been about real-world value.
Global investment in AI infrastructure has continued at a staggering pace, with estimates putting total spending on AI-related technology at $700 billion for 2026. But the phase of deploying AI to generate excitement is giving way to a much more challenging phase: deploying AI to generate measurable value.
The Pragmatism Era
As AI moves into everyday business operations, expectations are changing. Impressive demonstrations are no longer enough. Organizations want systems that can reduce costs, improve productivity, and demonstrate clear returns.
The most telling signal: High-profile demo-first products have stalled or been abandoned. OpenAI shut down Sora, its AI video-generation tool, after failing to convert technical ambition into sustainable revenue. In January, OpenAI introduced ads to ChatGPT, a significant pivot signaling real monetization pressure.
Pragmatism is also emerging on the buyer side. Microsoft canceled most of its Claude Code licenses over cost concerns. AT&T is limiting employee access to GitHub Copilot for the same reason. And Uber's COO said AI costs are getting "harder to justify." Some organizations are actively reducing their bills by routing work to cheaper models depending on the task complexity and restricting the priciest models to the employees who actually need them. Ensemble Health Partners moved its insurance-appeal-letter tool to a cheaper model and projects $700,000 in annual savings from that change alone.
Across the industry, organizations that rushed into experimentation are increasingly scrutinizing ROI, narrowing deployments, and focusing on systems that can deliver measurable operational value.
Embedded Intelligence
Another pattern of the past six months: AI is disappearing into the background. Intelligence is being embedded directly into workflows across the software people already use every day: email, productivity suites, development environments, and business applications.
Google's I/O conference in May offered one example: The company announced AI features embedded within Gmail, Docs, YouTube, and Search itself, treating AI less as a product and more as a layer running through everything it already offers.
When AI is a discrete tool, adoption is a choice. When it is embedded in the workflow, it becomes infrastructure.
From Scale to Specialization
Over the past six months, we’ve seen increasing emphasis on domain-specific AI systems.
In healthcare, OpenAI, Anthropic, Perplexity, and Microsoft all launched domain-specific products. In finance, Perplexity launched Perplexity Computer for Professional Finance, allowing finance teams to bring licensed data from providers like Morningstar and PitchBook into Computer, while Anthropic released AI agents for financial services firms. OpenAI announced a “personal finance experience” within ChatGPT, allowing users to connect their financial accounts to the platform via Plaid. In June, AlphaSense announced SuperAnalyst, an always-on AI agent purpose-built for financial and strategic workflows, while OpenAI launched six new Codex plugins covering data analytics, sales, creative production, equity investing, and investment banking.
The pattern is consistent: Generalized intelligence is losing ground to contextual intelligence. But as organizations connect AI systems to more data sources through an expanding ecosystem of integrations, many are discovering that connected is not the same as intelligent. A system can connect to dozens of sources and still miss important evidence. In enterprise workflows, access alone is not enough. Outputs need to be grounded in the right evidence and context.
The Rise of Agentic AI (and Its Limits)
Agentic AI has been one of the most talked-about developments of 2026. The ambition is significant and early enterprise adoption is widespread. Projects like OpenClaw and broader experimentation across the ecosystem point to a future where AI systems take actions, not just generate outputs. In May, Google announced Gemini Spark, a 24/7 agentic assistant, its response to Anthropic’s Claude Cowork and OpenAI’s ChatGPT agent.
But the last six months have also exposed the gap between promise and practice. In April, Claude-powered AI coding agent Cursor deleted a company's entire production database and backups in nine seconds.
And the risks aren't limited to operational errors. In April, Anthropic restricted access to its Claude Mythos Preview model specifically because of its exceptional ability to identify software vulnerabilities, a direct acknowledgment that the same capability that makes AI useful for security teams makes it useful for attackers.
Organizations evaluating agentic AI today should be clear-eyed: These systems can work well for well-defined, bounded tasks, often with human review in the loop. They are significantly riskier for open-ended tasks where the cost of an error is high.
AI Meets the Real World
Moving AI deeper into real operations has proven more complicated than expected.
The AI jobs apocalypse narrative is already being revised by some of the most prominent voices behind it. Both OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei recently acknowledged that they were wrong about AI’s impact on jobs. Amodei now says he expects automation to actually expand job responsibilities. Not everyone in the industry agrees. Anthropic co-founder Chris Olah, speaking at the Vatican's AI ethics conference in May, said the risk of large-scale labor displacement remains very real.
And as AI becomes more visible and more consequential, public anxiety about its effects is intensifying.
Commencement speakers were booed at graduation ceremonies across the country this past spring. Pope Leo XIV devoted his first papal encyclical to warning about the pace of AI adoption. In April, a man was charged with attempted murder after throwing a Molotov cocktail at OpenAI CEO Sam Altman's home. Earlier that month, someone fired shots at the home of an Indianapolis city councilman after he voiced support for a proposed data center, leaving behind a note that read: "No data centers."
The US government’s relationship with AI has become significantly more complex as well. The Trump administration blacklisted Anthropic, was subsequently sued over it, and then entered discussions to deploy its models within the federal government. In June, the administration ordered Anthropic to suspend access to its most powerful models for all foreign nationals, including some of Anthropic’s own employees.
What Comes Next
The first half of 2026 marked a turning point in how AI is being built, deployed, and evaluated.
The gap between frontier models is narrowing, as the underlying technology becomes increasingly similar across providers. Across industries, systems are being integrated into real workflows, and products are being evaluated based on cost, reliability, and impact.
At the same time, new challenges are surfacing around trust, security, workforce dynamics, and control. As AI becomes more embedded in decision-making and operations, the ability to measure performance, risk, and return will become critical.
Looking ahead, several trends are likely to shape the second half of the year:
- Continued movement toward specialized AI systems
- More proactive and agent-like systems
- Deeper integration into core business software and workflows
- Increased focus on governance, risk management, and auditability
- Growing pressure to demonstrate measurable ROI
AI is becoming more embedded, more consequential, and more difficult to evaluate with existing frameworks. The “wow phase” is ending. What comes next is far more complex and far more important.
Read the full Mid-Year AI Checkpoint report for a deeper look at the trends, risks, and strategic implications shaping AI in 2026.
Discover how you can transform your research process with AlphaSense’s Generative Search. Start your free 2-week trial of AlphaSense today.
