AI agents may soon become more independent, but Microsoft’s CEO of AI Mustafa Suleyman is drawing a sharp line between useful autonomy and dependable autonomy. In a podcast with investor Seth Rosenberg, he said models could act mostly autonomously within two years in narrow settings, while truly reliable action may still require a major technical jump.
The core issue is not whether AI can produce impressive results. It is whether a model can produce the exact right output when that output must trigger a real action.
Why autonomy is not the same as reliability
Suleyman expects AI agents to soon operate with less constant user oversight in limited use cases. That matters because many current AI systems still rely on people to check, steer, or correct the model before anything important happens.
But the standard changes when an AI system is expected to act. In that setting, the output is not just an answer for a person to read. It may need to match a function call exactly, with little room for variation.
According to Suleyman, current models still show variance across correct solutions for most answers. That makes them useful for tasks where flexibility is acceptable, but harder to trust when consistency is the whole point.
He described today’s 80 percent accuracy as insufficient for reliable action in new applications. For end users to maintain trust, he said, accuracy needs to reach 99 percent.
The GPT-6-level gap
Suleyman’s estimate is that models may need two more model generations and up to 100 times more computing power to become consistently reliable. He places that target at the GPT-6 level, likely OpenAI’s model after GPT-4 and GPT-5.
He said this may take another two years. That timeline suggests a near future in which AI agents become more capable and more visible, but still remain bounded by where mistakes are tolerable.
Importantly, Suleyman framed the challenge as precision rather than a search for entirely new abilities. The goal is a more exact mapping between prompt and output, so that the model’s response consistently lines up with the intended action.
That distinction matters for businesses considering AI agents. A system can be powerful and still not be reliable enough for high-stakes deployment. The question is not only what the model can do, but how often it does the right thing without intervention.
Where AI agents may work first
Suleyman sees the strongest near-term prospects in areas where some inaccuracy is acceptable. Legal research is one example he gave, because there can be multiple reasonable answers.
That does not mean the work is unimportant. It means the output can often be reviewed, compared, or treated as part of a broader research process rather than as a final action with no margin for error.
Medicine sits on the other side of the line in Suleyman’s view. Because it can require life-critical precision, he sees it as too demanding for the current level of reliability.
The practical split is clear:
- AI agents are more suitable where answers can be checked or where multiple reasonable paths exist.
- They are less suitable where a single wrong action could create severe consequences.
- Reliability depends on consistent prompt-to-output behavior, not only on broad model capability.
Suleyman also warned that complete autonomy is dangerous. By that he means systems that make their own plans and procure their own resources. He said that kind of autonomy needs to be regulated.
Why data quality may matter more than size
Suleyman also argued that model size is becoming less important to AI success. He pointed to Microsoft’s open-source Phi 3 model, which is 100 times smaller than top models in terms of inference effort.
He said Phi 3 may not be as good as the top models, but it matches or beats GPT-3.5. For Suleyman, the lesson is that data quality is becoming critical, not just the number of parameters.
OpenAI CEO Sam Altman recently made a similar point, saying the future of AI training is about learning more from less data. In this framing, better training material can matter as much as, or more than, simply scaling model size.
That creates an opening for startups. Suleyman advised young companies to focus on collecting high-quality training data for their applications, rather than assuming that only the largest possible model can compete.
He also said this includes training human experts to give proper feedback for model training. His AI company Inflection, somewhat acquired by Microsoft, used strict selection for "AI teachers" when developing the Pi chatbot.
The assistant that follows your digital life
Suleyman’s longer-term view is that personalized AI assistants will accompany users throughout their digital lives. He predicts they will remember everything users do and say, while proactively making suggestions.
That vision depends on the same reliability problem. An assistant that remembers context and suggests actions becomes more useful as it becomes more dependable. But the more it acts on behalf of a user, the more trust it must earn.
For now, Suleyman’s message is restrained but consequential. AI agents may become more autonomous soon, especially in narrow use cases. The harder milestone is getting from impressive performance to consistent action, and he expects that may require GPT-6-level compute, better data, and a much tighter connection between instruction and output.