Microsoft AI CEO: Human-Level AI Performance Imminent
Microsoft’s AI CEO, Mustafa Suleiman, predicted that within 12 to 18 months, AI will achieve human-level performance in most professional white-collar tasks like accounting, legal, and marketing. However, a related survey found that while 70% of professionals feel AI progress is accelerating faster than they can keep up, organizational adoption remains slow, with only 0.5% of knowledge workers having built a custom GPT.
- Mustafa Suleiman co-founded two significant AI companies, DeepMind and Inflection AI, before becoming the CEO of Microsoft AI. At DeepMind, which Google acquired for a reported £400 million, he was the head of applied AI. His work there included reducing Google's data center cooling costs by up to 40% using AI and launching DeepMind Health to develop technology for the UK's National Health Service. - Suleiman's prediction is part of a broader trend of experts shortening their timelines for when artificial general intelligence (AGI) will be achieved. For example, the median estimate on the forecasting platform Metaculus for AGI's arrival has dropped from 50 years away in 2020 to much sooner, with a 50% chance projected by 2033. Other leaders in the field, such as the CEOs of OpenAI, Tesla, and Anthropic, have predicted AGI could emerge in the next two to five years. - The term "human-level performance" is often measured by AI's ability to pass complex, multi-step reasoning benchmarks designed for humans. Key metrics also include human evaluation of an AI's output for quality and coherence, as well as its impact on business key performance indicators (KPIs) like productivity, cost savings, and customer satisfaction. - Despite the optimism from AI leaders, enterprise adoption faces significant hurdles. A 2025 survey highlighted a scarcity of skilled professionals as a primary barrier for 55% of respondents. Other major challenges include concerns about data privacy, security, and bias (cited by 45% of businesses), and a lack of sufficient proprietary data to customize models. - In fields like accounting and law, AI is already being used to automate tasks such as contract review, eDiscovery, and anomaly detection in audits. For instance, AI can now automate data extraction from various document formats and help ensure compliance with accounting standards like ASC 606. - Suleiman has indicated that achieving reliable, autonomous AI actions will likely require future model generations, potentially at the level of GPT-6, and a hundredfold increase in computing power. He suggests that until then, the best applications are in areas where some level of inaccuracy is acceptable, like legal research, as opposed to life-critical fields such as medicine. - There is a notable difference in AGI timeline predictions based on geography. In a 2017 survey, Asian AI experts expected AGI to arrive in 30 years, while North American experts predicted it would take 74 years. This highlights the global variance in optimism and expectations regarding AI development. - The slow adoption rate of custom AI tools, like the 0.5% of knowledge workers who have built a custom GPT, can be partly attributed to the difficulty in proving a clear return on investment (ROI). Many companies struggle to move AI projects from the pilot stage to full-scale enterprise adoption due to a lack of standardized processes and bottlenecks in IT and data alignment.
