Microsoft used its Build conference in San Francisco to show how it plans to compete more aggressively in enterprise AI, pairing new reasoning and multimodal models with tools for agents and security.

At the center of the announcement was MAI-Thinking-1, Microsoft’s first reasoning model. The company said the model has 35 billion parameters and a 128K context window, and is designed to help reduce token usage, a cost concern for companies deploying large numbers of AI agents. Microsoft also introduced MAI-Image-2.5 for image generation, MAI-Transcribe-1.5 and MAI-Voice-2 for speech, and MAI-Code-1, a coding model tuned for GitHub.

The model release was part of a broader push around agents, which have become one of the main themes in the AI industry. Microsoft unveiled what it describes as a Copilot superapp, combining coding, chat and the Copilot Cowork agent. CEO Satya Nadella also described a new class of enterprise agents he called autopilots, long-running systems that can use connectors, context and memory. The first example is Microsoft Scout, a personal agent built on OpenClaw that can help with meeting preparation, scheduling and routine tasks.

Nadella said Microsoft plans to expand that into a larger digital team of agents over time. That approach reflects the company’s effort to make AI more deeply embedded in workplace software, especially for enterprise customers already using Microsoft products.

Microsoft also announced several tools aimed at giving agents access to information and helping companies manage them. Microsoft IQ, the company’s enterprise knowledge and intelligence system for agents, is now generally available. WebIQ, meanwhile, is a new platform that lets agents use real-time web search.

Security and governance were another major part of the launch. Microsoft introduced ASSERT, an open-source tool for automating AI safety evaluations, along with the Agent Control Specification, an open-source standard for managing agent controls. The company also announced Codename MDASH, an agentic bug-hunting system, and Frontier Tuning, which applies reinforcement learning within a company’s compliance rules so agents can learn from organizational policies.

Sarah Bird, Microsoft’s chief product officer for responsible AI, said the company sees these pieces as interconnected rather than standalone features. In an interview with The Deep View, she said Microsoft needs an ecosystem where agents can work effectively inside organizations while remaining safe and secure.

Microsoft’s announcements suggest a strategy aimed less at headline-grabbing breakthroughs and more at building a practical enterprise platform around AI. The company is leaning on its longstanding presence in corporate software and its security credentials as it tries to catch up in a market where other model developers and cloud providers have already established strong positions.

That leaves Microsoft with a familiar playbook, but one that may resonate with enterprise buyers. By combining models, agent tooling and security controls, the company is trying to make itself a central platform for businesses that want to deploy AI without losing oversight of cost, access or compliance.