Microsoft Needs To Up its AI Game
A recent report from The Information documenting how Microsoft's internal AI efforts are doing, and it's not a pretty picture. According to the article (paywalled), Microsoft salespeople are cutting forecasts and reducing sales goals for its Azure AI products across the board.
It's not a total wash, as The Information noted AI has been "a major boon" to Microsoft's business, largely due to new spending by AI firms such as OpenAI, which has projected it would rent about $15 billion worth of Azure cloud service this year. Microsoft also gets sales of AI software such as its 365 Copilot workplace tools and GitHub Copilot coding agent.
But those are falling short of targets. The Information said that Microsoft salespeople are struggling to meet goals, due to a lack of demand. There has been some evidence of this. This past June, Bloomberg reported that workers preferred using generic OpenAI to Copilot, which was cutting into Microsoft's ability to sell its Copilot.
Microsoft attempted damage control, with a spokesperson telling Bloomberg in a statement that "The Information's story inaccurately combines the concepts of growth and sales quotas," and that "aggregate sales quotas for AI products have not been lowered."
One of the problems with agentic AI, at least in its current state, is that it requires far too much human intervention -- but it's supposed to be autonomous. In a recent experiment by Carnegie Mellon University, researchers at the school created a fake software company entirely run by AI agents from Google, OpenAI, Anthropic and Meta.
The simulation, called TheAgentCompany, used agents in roles such as financial analysts, software engineers, and project managers, working alongside simulated coworkers like a HR department and a chief technical officer. The researchers set the agents to do tasks based on the day-to-day work of a real software company.
The results were something akin to the clones skit in "Iron Man 2." The best performing model was Anthropic's Claude 3.5 Sonnet, which could only finish just 24 percent of the jobs assigned to it, and each task required an average of nearly 30 steps to complete and a cost of over $6 per task, making it both inefficient and too expensive.
The Information also cited a real-world example of agentic AI falling short. A private equity fund called Carlyle Private Equities last year started using Copilot Studio to automate tasks like summarizing meetings or drafting financial models based on Excel spreadsheets, without needing to write any code.
But a source within the company told The Information that after a few months, Carlyle representatives told Microsoft they were having trouble getting the AI to reliably tap data from other applications, which was necessary for some of Carlyle's automations. This fall, Carlyle reduced the amount it was spending on the tools, the source said.
That could be why office workers prefer to use plain old ChatGPT instead of Copilot. I find Copilot to be a little on the clumsy side, but then again, that's par for the course with Microsoft. The first few versions of any piece of software from that company have always been dodgy at best.
All of this illustrates one thing: agentive AI is not ready for prime time. Microsoft, OpenAI, and everyone else in this business have a lot of work ahead of them and given the gravity of what they are attempting, they need to get it right.
Posted by Andy Patrizio on 01/09/2026