Microsoft And DeepSeek Are Launching New AI Together

Key Concepts

• Agentic AI: AI systems capable of autonomous, multi-step task execution, including tool usage, file analysis, and long-running cloud processes.
• Multi-model Strategy: An architectural approach where an enterprise platform routes tasks to different AI models based on cost, performance, and task complexity.
• Usage-Based Billing (Pay-as-you-go): A pricing model for AI agents where costs are calculated based on model calls, context retrieval, tool usage, and runtime.
• Grounding: The process of connecting AI models to real-time, verified data (e.g., Web IQ) to improve accuracy and reduce hallucinations.
• Distillation Concerns: The risk of companies using advanced AI models to train or improve their own competing proprietary models.

---

1. Microsoft Copilot Co-work: Overview and Capabilities

Microsoft has launched Copilot Co-work, an "agentic" version of its Copilot product designed for complex, long-running tasks. Unlike standard chat interfaces, Co-work can:

• Break down jobs into multi-step workflows.
• Access company data and internal systems.
• Execute tasks across thousands of files and spreadsheets.
• Operate autonomously in the cloud to deliver a final result.

Real-world Applications:

• Engineering: Automating the editing of batch job spreadsheets and generating dependency flowcharts.
• Data Analysis: Comparing thousands of files across product versions in minutes—a task previously taking weeks.
• Sales: Analyzing stalled pipelines to identify "at-risk" opportunities and cold follow-ups.

2. The Economics of Agentic AI

Microsoft has transitioned Co-work to a usage-based billing model because agentic workflows are computationally expensive.

• The Cost Problem: Because agents perform multiple model calls, retrieval steps, and tool interactions, costs can "explode" compared to simple prompt-response interactions.
• Pricing Structure:
  • Copilot Credits: The unit of measurement for usage.
  • Factors: Price is determined by model use, context retrieval, tool calls, and runtime.
  • Tiers: Microsoft categorizes tasks as Light (low reasoning/output), Medium (structured reasoning), and Heavy (broad aggregation/deep reasoning).
  • Payment Options: A standard pay-as-you-go rate (1 cent per credit) or a "P3" volume-commitment discount.

3. The Multi-Model Strategy and DeepSeek Integration

Microsoft is evolving into an enterprise AI platform that routes tasks to the most efficient model.

• Model Diversity: Co-work currently utilizes Anthropic (Opus 4.8, Sonnet 4.6) and OpenAI (GPT 5.5) models.
• DeepSeek Integration: Microsoft is exploring the use of a fine-tuned DeepSeek V4 model as a cost-effective option for specific workloads.
• Co-work 1: Microsoft’s own secure, fine-tuned model, designed for everyday, cost-sensitive tasks, expected to launch in the coming weeks.
• Security: If implemented, DeepSeek would be fully hosted on Azure, ensuring data remains within Microsoft’s enterprise security and compliance perimeter.

4. Microsoft’s Geopolitical AI Strategy

Microsoft occupies a unique position as a bridge between Western AI innovation and the Chinese tech ecosystem.

• Market Presence: Despite geopolitical tensions, Microsoft’s Azure AI revenue in China has seen massive growth (tripling in fiscal year 2025).
• Key Customers: Major Chinese firms like ByteDance, Ant Group, Meituan, and Tencent utilize Microsoft’s cloud and AI services.
• Operational Constraints: To mitigate IP theft, Microsoft does not host OpenAI models in Chinese data centers; instead, Chinese customers access them via international facilities (e.g., Singapore).
• The "Distillation" Challenge: There is ongoing concern regarding whether Chinese firms use Western model outputs to improve their own proprietary models. Microsoft employs automated monitoring, though enforcement remains complex.

5. Web IQ: Grounding for Agents

Microsoft introduced Web IQ, a Bing-powered grounding system specifically architected for AI agents rather than human users.

• Technical Shift: Unlike traditional search engines optimized for human readability (links/ads), Web IQ is optimized for machine consumption, providing fresh data with low latency.
• Performance: Microsoft claims Web IQ is 2.5 times faster than alternatives, which is critical for agents that may perform hundreds of search calls per task.
• Strategic Goal: By owning the search layer (Web IQ), the model layer (multi-model routing), and the cloud runtime, Microsoft aims to control the entire "agent stack."

---

Synthesis and Conclusion

Microsoft is shifting from a single-model assistant provider to a comprehensive enterprise AI agent platform. By introducing usage-based billing, multi-model routing (including potential integration of Chinese models like DeepSeek), and specialized infrastructure like Web IQ, Microsoft is prioritizing efficiency and scalability. The strategy reflects a pragmatic approach to the "agent economy," where the primary challenge is balancing the high computational cost of autonomous work with the need for enterprise-grade security and performance.