A profile of Microsoft EVP of Copilot Charles Lamanna, a potential Satya Nadella successor; a source says Microsoft is on track to pay Paradox $2B+ per year (Aaron Holmes/The Information)

Microsoft Copilot is an AI assistant built into Microsoft’s productivity software. It is meant to help users automate work rather than perform every step manually. In the source, Microsoft specifically connects Copilot with work in Excel and PowerPoint. That makes it important because these applications are central to everyday business tasks.

In Excel, Copilot might help organize information, analyze figures, or create formulas from a natural-language request. In PowerPoint, it might turn notes into slides, summarize material, or suggest presentation content. The basic mechanism is an AI model interpreting the request and taking actions within the application.

The promise has not always matched customer experience. In summer 2025, Microsoft customers complained that Copilot was buggy and not good enough at automating work in Excel or PowerPoint. Improving reliability will matter if Microsoft wants Copilot to become a trusted workplace tool rather than an experimental assistant.

In summer 2025, Microsoft faced customer complaints about Copilot’s performance. Users said the assistant was buggy. They also said it was not good enough at automating work in Excel or PowerPoint. These complaints challenged the product’s central promise: that AI could make familiar office tasks easier and faster.

The problem was not simply that Copilot made occasional mistakes. Customers were dissatisfied with its ability to carry out useful work inside major Microsoft applications. For example, an assistant that struggles with spreadsheet work or presentation creation may save little time and require users to check or repair its output.

The complaints created pressure for Microsoft to improve Copilot. The product is positioned as an important AI assistant, but customers’ experiences showed a gap between that ambition and practical reliability. That gap also helps explain why Microsoft’s model choices and leadership around Copilot matter.

Charles Lamanna is Microsoft’s executive vice president of Copilot. That places him at the center of the company’s effort to develop and improve its AI assistant. Copilot is also facing customer complaints, making Lamanna’s role especially visible as Microsoft works to turn the product into a dependable workplace tool.

The source identifies Lamanna as a potential successor to Satya Nadella. It does not provide a detailed succession plan or explain exactly why he is being considered. The clearest stated connection is his senior position overseeing Copilot, one of Microsoft’s most prominent AI efforts.

His possible succession matters because Copilot’s performance could shape views of Microsoft’s AI strategy and leadership bench. For now, the article presents Lamanna as a potential successor, not as Nadella’s confirmed replacement. It also places his leadership against a backdrop of product complaints and major model-supply decisions.

A source cited in the article says Microsoft is on track to pay Paradox more than $2 billion per year. That is an exceptionally large recurring technology expense, showing how valuable access to advanced AI models can become. It also suggests that model supply is a major part of Microsoft’s broader Copilot strategy.

The payment would go to Paradox, an AI company whose models Microsoft could use. The article gives one clear figure: $2 billion or more annually. It does not state the cost of developing or operating any particular major AI system, so a precise comparison cannot be made from the source.

In general, large AI systems require substantial computing, research, and operating resources, but those costs vary widely by system and usage. The reported payment therefore signals scale without proving that Paradox access is cheaper or more expensive than building everything internally. Microsoft’s commercial choices remain tied to Copilot’s performance and customer demand.

Microsoft might pay Paradox to broaden the models available for Copilot and other AI products. Relying on an outside provider can give a company access to capabilities, improvements, or capacity that its internal models may not offer at the same time. That flexibility can matter when customers expect fast progress and dependable performance.

The practical mechanism is model access through a commercial relationship. Microsoft could use Paradox’s models for some tasks while using Microsoft-developed models for others. Different models may perform better on different kinds of work, including reasoning, writing, coding, or handling requests inside office software. The source does not specify which Paradox models or tasks are involved.

This approach also creates a trade-off. External access can accelerate product development, but it introduces a supplier relationship and a recurring bill. The reported $2 billion-plus annual figure shows that Microsoft may see outside models as strategically valuable, even while Copilot faces complaints about bugs and weak automation.

If Microsoft depends heavily on Paradox’s models, Copilot could gain access to useful capabilities or improvements. Better models might help address the complaints that Copilot was buggy and weak at automating work in Excel and PowerPoint. Product quality would then depend partly on how well Paradox’s models fit Microsoft’s software.

The key mechanism is shared dependence. Microsoft would pay for model access and build products around an outside provider’s technology. If usage grows, the recurring bill could grow too. Microsoft might also need to adapt its products when Paradox changes pricing, availability, performance, or technical terms. These are general business implications, not outcomes stated directly in the source.

The relationship could therefore produce both benefits and risks. Stronger models could make Copilot more useful, while supplier dependence could raise costs and reduce Microsoft’s leverage. Microsoft’s own models, multiple suppliers, or alternative systems could preserve more negotiating power, but the source does not say which path it will choose.

Large language models generate answers by processing a user’s request and predicting a sequence of likely words or tokens. When connected to software, they can also choose structured actions, such as creating a formula, summarizing data, or drafting slides. The model uses the request and available context to produce an answer or suggest an operation.

Inside Excel or PowerPoint, a system typically combines the model with software tools and the user’s files. The model interprets what the user wants, while the application performs approved actions. For example, a request to analyze a table could lead to generated calculations or a written explanation. The model’s prediction is not the same as guaranteed verification.

That is why Copilot can sound confident while producing errors. It may misunderstand the request, use the wrong context, make an incorrect calculation, or trigger an unsuitable action. The source’s summer 2025 complaints show the practical result: customers found Copilot buggy and not good enough at automating work in Excel and PowerPoint.