Satya Nadella Says He Interrogates AI Agents Rather Than Trusting Their Output

Microsoft CEO Redefines Executive AI Workflow Through Active Interrogation

Satya Nadella, Microsoft CEO, has outlined a fundamentally different approach to working with AI agents, emphasizing active interrogation over passive acceptance of outputs1

1

. Speaking on the Sources Podcast, Nadella explained that he doesn't simply view a model output and accept it. Instead, he wants to manipulate it, interrogate AI models, and reason over the results. This hands-on methodology represents a shift in how executives are leveraging AI agents for critical business decisions, particularly when analyzing complex data like return on invested capital (ROIC) metrics and SEC filings1

1

.

The Microsoft CEO specifically highlighted his use of the company's Excel agent and Cowork platform within Copilot for operational analysis. When his data center team sent him a complicated spreadsheet, Nadella opened it and directed the Excel agent to examine a specific cell's formula and create five sheets with different scenarios1

1

. This capability to perform multi-step tasks on existing data, rather than just generating new models, marks what Nadella considers a critical advancement in AI agents.

Causal Reasoning and Enterprise Context Drive AI Agent Value

Nadella identified two major advancements that make current AI agents valuable for enterprise applications. The first is the ability to analyze existing models in depth, which he describes as causal reasoning1

1

. The Excel agent's capacity to mirror this type of reasoning allows him to continuously evaluate ROIC across different business layers. Rather than accepting surface-level outputs, this approach enables deeper analytical work that directly addresses concerns about whether massive AI infrastructure investments are generating actual value.

The second advancement combines enterprise context with global information. Nadella maintains a data runner in Fabric that aggregates all SEC data from cloud providers and transforms it into a semantic model1

1

. His coding agent then analyzes this data and creates a dashboard that merges every SEC filing with internal analysis, providing real-time ROIC metrics. This integration of external market data with internal operations exemplifies how AI agents can deliver strategic insights when properly configured and questioned.

Trust Emerges as Central Challenge for Autonomous Agents

As Microsoft launched its redesigned Copilot platform on September 25, consolidating chat, coding, and agentic capabilities into a unified product, Satya Nadella identified trust in AI as the biggest challenge facing the technology2

2

. The fundamental question he keeps returning to is whether users would hand AI agents their login credentials and allow autonomous activity without close oversight at each step. This concern intensifies when AI agents operate inside businesses, where the systems they access and actions they take carry far greater consequences than personal tasks.

The shift from AI assistants that wait for requests to AI agents that receive goals and work through sequences independently creates new exposure. When a system operates with meaningful autonomy across several steps, errors or unauthorized actions may be several layers deep before detection2

2

. Unlike chatbots that deliver wrong answers to direct questions, autonomous agents making intermediate decisions throughout multi-step tasks present a different risk profile that enterprises must carefully evaluate.

Microsoft Consolidates Copilot Platform with Multi-Model Approach

The updated Microsoft Copilot platform routes each request automatically to what it determines is the best available model, choosing between AI systems from OpenAI and Anthropic2

2

. Users can also select models manually, and Microsoft plans to add more providers over time. Each agent runs continuously in the cloud within a company's Microsoft 365 environment and carries its own identity, memory, and email address, allowing users to tag agents directly in Teams or Outlook.

Microsoft 365 Copilot reached 30 million paid seats in July 2024, with net additions doubling quarter-over-quarter2

2

. The September launch positioned Microsoft against rivals like OpenAI, which introduced its own unified AI application in July, and Meta, which has been developing its own agent platform. Nadella emphasized that companies should run their own performance evaluations and switch between vendors, treating model selection as a decision they control.

New Pricing Model Balances Access with Infrastructure Costs

Microsoft is implementing a per-seat-plus-consumption billing structure for its Copilot platform. A standard monthly subscription covers everyday AI work, including conversational tasks and standard Copilot features, with token limits the company says most users won't reach2

2

. Customers with heavier workloads can switch to consumption billing, which covers additional capabilities and longer, more demanding agentic tasks. Administrators can set spending limits to keep costs within defined budgets.

Nadella explained this combination is designed to offer broad access while tying the cost of advanced AI work to the value each customer derives from the platform2

2

. He noted that consumption billing will grow in importance as AI companies pull back from subsidized rates used to drive early adoption. Running AI infrastructure is expensive, and costs climb when tasks require multiple model calls or sustained processing time, making usage-based billing a way to recover costs from customers generating them while keeping flat subscriptions for standard use.