2025 AI Yearbook: The Year AI Became Enterprise Infrastructure

2025 was not the year of a single breakthrough model. It was the year AI moved from assistive software into enterprise infrastructure. Across platforms and enterprises, the focus shifted away from chat interfaces toward execution, governance, and operational control.

This 2025 AI yearbook documents that shift.

What follows is a view of how AI entered production systems, how enterprises began operating it, and why governance became the central challenge heading into 2026.

I. From copilots to agents: AI crossed the execution boundary

Early 2025 still looked like the “copilot era”: assistive tools, chat interfaces, productivity boosts. By year-end, the most important announcements had nothing to do with chat. The most consequential platform announcements were about AI agents that act.

• Anthropic launched Enterprise Agent Skills and opened it as a standard, turning agent capabilities into reusable, portable components across platforms.

• OpenAI expanded Codex Agents, enabling multi-step repository access, CI/CD execution, and tool orchestration inside real engineering environments.

• Amazon Web Services (AWS) announced Bedrock AgentCore at re:Invent - covering governance, evaluation, memory, and runtime control, not just model hosting.

• Databricks shipped the Mosaic AI Agent Framework, tightly integrated with enterprise data lakes.

• Snowflake expanded Cortex AI, embedding LLMs and agentic workflows directly inside data warehouses.

• Oracle embedded GenAI agents directly into ERP, HCM, and supply-chain products - where AI is no longer assisting work, it is the workflow.

• IBM embedded watsonx Orchestrate into S&P Global’s enterprise systems, marking the adoption of agent-based automation in regulated, data-intensive environments.

• Microsoft doubled down on “agents as product”: at Build 2025, Copilot Studio highlighted multi-agent orchestration; and Microsoft later pushed Foundry Agent Service as a managed way to build/deploy “trusted” agents

• Google expanded Gemini Enterprise, its agent-centric AI platform that lets organizations create, deploy, and manage AI agents across enterprise workflows with built-in tool integrations and governance.

II. MCP went mainstream and quietly changed the risk model

One of the most consequential shifts of 2025 was the normalization of MCP as shared infrastructure for AI Agents.

A key signal was MCP’s movement toward open governance. The protocol and its ecosystem were placed under the Linux Foundation AI & Data umbrella, signaling that MCP was being treated as foundational interface infrastructure, similar to how Kubernetes, OpenTelemetry, and Envoy evolved beyond single-vendor control.

At the same time, major platforms began operationalizing MCP: When Google moved to managed MCP servers, agents gained the ability to connect to tools via a URL. That single change effectively turned IDEs, browsers, developer laptops into integration platforms.

Once agents can dynamically connect to external MCP servers, the core question stops being: “Which model are we using?” It becomes: “What systems can this agent reach and who approved that access?”

III. Fortune 500s didn’t debate AI — they operationalized it

The strongest signal in 2025 came from how large enterprises adopted AI. Enterprise spend on generative AI surged (~3.2x YoY). Companies spent roughly $37B on generative AI in 2025, with strong growth in application layer investments.

• Walmart made “agentic commerce” mainstream, partnered with OpenAI to enable AI-first shopping and checkout flows, while rolling out internal AI tools for associates and merchants.

• JPMorganChase disclosed internal platform “LLM Suite” which claims LLM usage across internal workflows, spanning contract intelligence, analytics, and operations.

• Goldman Sachs rolled out its generative AI platform (GS AI Assistant) to a large share of employees and is expanding usage across functions for tasks such as document summarization, technical analysis, and workflow automation

• Visa partnered with major tech companies (including Microsoft, OpenAI, Anthropic, IBM, Mistral and Stripe) to launch Visa Intelligent Commerce, a platform that enables delegated shopping tasks via AI agents with controlled spending limits.

• Siemens deployed industrial AI copilots across manufacturing and operational technology (OT) environments.

• GE Aerospace deployed a new AI-enabled blade inspection tool in 2025 to increase inspection accuracy and consistency for narrowbody aircraft engines.

• SAP reported 400+ AI business use cases across its enterprise suite.

While some of them were pilots, many were production deployments in enterprise environments.

IV. Vertical AI crossed into core systems

Across industries, AI moved into mission-critical workflows:

• HR: ADP embedded AI - ADP Assist into payroll and workforce intelligence

• Enterprise SaaS: Salesforce expanded Einstein Copilot into agent-like automation

• Contracts & Compliance: Workday launched a Custom AI Model Library for contract automation, leveraging Evisort acquisition capabilities.

• Retail: Amazon launched Rufus, an AI shopping assistant operating at consumer scale.

• Payments: Stripe engineered a payments-focused foundation model in 2025 that was designed to improve fraud detection and transaction decisioning compared to legacy approaches.

Each vertical adopted AI differently, but all crossed the same line. Across sectors, AI was integrated directly into decision-making and execution workflows.

V. AI became measurable, and boards took ownership

2025 was also the year AI stopped being anecdotal.

• Fortune and ServiceNow launched the inaugural AIQ 50, ranking Fortune 500 companies by measurable AI impact.

• Enterprises began referencing AI ROI directly in earnings calls. Example: Microsoft stated, “more than 90% of the Fortune 500 use Microsoft 365 Copilot".

• AI maturity moved from innovation teams to board-level oversight.

• In payments, AI adoption began appearing in risk committee and audit committee discussions.

Once AI is measurable, it becomes governable, and once it’s governable, it becomes mandatory.

VI. Infrastructure and energy became the real bottlenecks for AI

As AI scaled, the constraint shifted. Not talent. Not models. Energy and compute.

• Alphabet Inc. made a $4.75B acquisition of Intersect, an energy company focused on powering AI data centers.

• NVIDIA secured inference technology and talent from Groq, reinforcing the shift from training to serving at scale.

• AMD, Intel and Cerebras pushed new accelerators optimized for inference efficiency.

• Cerebras expanded wafer-scale inference and partnered with Meta for large-scale deployment.

• DeepSeek, low-cost open AI models reduced training and inference cost assumptions, forcing incumbents to re-evaluate compute efficiency and pricing strategies.

AI infrastructure planning increasingly resembled industrial capacity and energy planning, rather than traditional cloud service expansion.

What changed is that major platforms started packaging “AI factory” language and agent runtime + governance controls together indicating infrastructure is being sold as an end-to-end production stack.

VII. Global and sovereign AI entered the equation

AI also became geopolitical infrastructure:

• The UAE expanded sovereign AI infrastructure and national model programs.

• The EU clarified enforcement timelines for the AI Act, forcing enterprises to operationalize governance.

• Major Chinese enterprises accelerated domestic LLM deployment across manufacturing and public services.

• SoftBank + OpenAI formed a joint venture in Japan to deliver enterprise AI at a national scale.

AI strategy is now inseparable from national and regulatory strategy.

The macro signal of 2025 for AI

The defining shift of 2025 was not faster models or broader experimentation, but

where AI was placed in the enterprise stack

Across all announcements, one pattern dominated:

1. AI moved closer to execution

2. Capabilities became modular (agents, skills, MCP)

3. Access expanded across trusted environments

4. Measurement became mandatory

5. Infrastructure became strategic

In short, AI became infrastructure.

What will 2026 be defined by in AI Governance?

The defining questions in 2026 will no longer be about access or experimentation. They will be about control, reliability, and accountability. In 2026, enterprises will be forced to answer questions they could defer in 2025:

• Visibility: What agents exist across the organization? Where do they run? What systems, data, and tools can they access?

• Control: Which actions are allowed to execute autonomously? Which require approval? How are permissions scoped, reviewed, and revoked?

• Governance: What policies apply before an agent acts-not after an incident occurs? How are those policies enforced consistently across environments?

• Reliability and cost: How are agent behavior, inference cost, latency, and failure modes monitored at scale? What happens when agents fail, loop, or degrade silently?

• Third-party risk: How are external skills, tools, MCP servers, and agent capabilities vetted, versioned, and governed as part of the software supply chain?

The gap in 2026 will not be between companies that “use AI” and those that do not. It will be between organizations that can see, control, and govern autonomous systems in real time, and those that are reacting after those systems have already acted.

In 2026, AI governance becomes the operating system for enterprise AI.