What Is the Difference Between Narrow AI and AGI? And Should Boards Care About AGI?

Narrow AI, sometimes called weak AI, refers to AI systems designed and capable of performing specific, defined tasks. All of the AI systems in commercial use today are narrow AI: they are excellent at their designated tasks and useless or non-existent at tasks they were not designed for.

A chess-playing AI beats world champions but cannot drive a car. A speech recognition system transcribes audio with high accuracy but cannot diagnose diseases. GPT-4, Claude, and Gemini are extraordinary at language tasks: writing, summarising, translating, explaining, and reasoning over text. They are not general intelligences; they are specialised tools for a particular domain of tasks.

The characterisation of modern large language models as 'narrow' is contentious: they can perform a very wide range of language and reasoning tasks, far broader than earlier narrow AI. But they remain fundamentally task-constrained: they work on language inputs and produce language (or code, or structured data) outputs. They do not have goals, persistent memory across sessions by default, or the ability to act in the physical world.

Artificial General Intelligence refers to a hypothetical AI system capable of performing any intellectual task that a human can perform, with equivalent flexibility and the ability to transfer learning across domains. An AGI could be given any intellectual task, learn what it needed to perform that task, and execute it at or above human level.

AGI does not currently exist. The definition is aspirational and contested: different researchers define AGI differently, and some argue the term is ill-defined. OpenAI, Anthropic, Google DeepMind, and other frontier AI labs describe the development of AGI as part of their long-term mission, but none have claimed to have achieved it.

The timeline for AGI is genuinely uncertain. Some researchers believe it is years away; others believe it is decades away; others question whether it is achievable at all with current architectural approaches. This uncertainty is honest, not evasive: nobody knows.

For most boards, the primary AI consideration should be current, deployable, narrow AI capabilities, not AGI. The business decisions at hand are about deploying Microsoft Copilot, governing AI use by employees, managing model risk in data-driven decisions, and navigating AI regulation. None of these require AGI to be on the immediate horizon.

However, boards should have a basic awareness of AGI-related risks for two reasons. First, frontier AI capabilities are advancing rapidly, and the boundary between narrow AI and proto-general capabilities is blurring. Systems that can autonomously plan and take actions across a wide range of tasks (agentic AI) represent a qualitative shift from earlier narrow AI, even if they are not AGI. Second, regulatory and insurance environments are beginning to engage with AGI-related risk scenarios, and boards of large enterprises will increasingly be asked about their AI risk horizon awareness.

The appropriate posture is awareness without preoccupation: understand that AGI is a long-horizon consideration, keep track of how the frontier is evolving, and focus governance energy on the AI capabilities that are deployed today and will be deployed in the next two to three years.

The most practically significant development between current narrow AI and hypothetical AGI is the emergence of agentic AI: systems that can plan and execute multi-step tasks, use tools, and take actions in digital environments with limited human oversight.

Current agentic AI systems (Microsoft Copilot agents, autonomous AI assistants) are still narrow in important ways: they operate within defined environments, have limited ability to acquire new capabilities, and require significant human design and oversight. But they represent a material increase in AI autonomy relative to earlier systems.

Boards should ensure that their AI governance frameworks address agentic AI specifically. The governance requirements for an AI that can send emails, create records, execute transactions, and modify documents on behalf of employees are substantially more demanding than for an AI that generates text for human review.