A new generation of artificial intelligence systems is crossing a threshold that researchers have long theorised about but never quite reached: the ability to set goals, make plans and take extended sequences of actions in the world — without waiting for a human to approve each step.

These so-called 'agentic' systems can already book a table at a restaurant, execute a multi-step legal research project, write and deploy software, or manage a complex calendar schedule. They do this not by following rigid scripts, but by reasoning dynamically about the best way to achieve a stated objective.

The implications for the economy are staggering. Goldman Sachs estimates that agentic AI systems could automate tasks that currently occupy the equivalent of 300 million full-time jobs globally — not by replacing workers outright, but by handling the cognitive-but-routine components of their work.

"We are moving from AI as a tool to AI as a colleague," said Dr. Fei-Fei Li, co-director of Stanford Human-Centred AI Institute. "The question is not whether this transition happens, but whether we design it to benefit everyone."

The safety challenges are equally profound. When an AI system can take actions that affect the real world — sending emails, executing financial transactions, controlling physical infrastructure — the consequences of errors or misalignment become far more serious than a chatbot producing an inaccurate response.

OpenAI, Google DeepMind and Anthropic have each published safety frameworks for agentic systems, but critics argue that the industry is moving faster than the guardrails can keep up with.

Regulation is struggling to catch up. The EU's AI Act, which came into force in 2025, contains provisions for 'high-risk AI systems', but agentic AI was not explicitly anticipated in the original drafting.

What seems certain is that the transition to agentic AI will be rapid. Goldman Sachs predicts that 40% of enterprise software spending will shift toward agentic systems by 2028, as companies realise the competitive advantage of AI that can operate continuously, without supervision.