For most of the history of artificial intelligence, the direction of travel has been clear. We have tried to make machines more capable, more independent and less reliant on continuous human supervision.
That ambition made sense. A machine that constantly needs a human looking over its shoulder is not particularly useful. From industrial robots and autopilot systems to algorithmic trading and automated logistics, we have progressively taught machines to do more on their own.
But AI is now crossing an important threshold. The chatbot era was largely about machines that could answer questions, generate content and offer advice. The emerging era of Agentic AI is different. AI agents can increasingly pursue objectives, break them into tasks, use external tools, communicate with other systems, access databases and execute transactions. AI is moving from answering to acting.
This shifts the debate from autonomy to authority. The question is not simply how much AI can do on its own, but how much power we are prepared to place in its hands.
There is a crucial distinction between capability and authority. Much of my career has been spent at the intersection of technology research and the practical realities of building and managing business systems. In academia, we naturally ask what technology can do. In business, an equally important question follows: who has the authority to do it?
A finance director may have considerable expertise and operational independence, but that does not necessarily give unlimited authority to commit the company to a £100 million transaction. An employee may negotiate a contract but not have the authority to sign it. A fund manager may make investment decisions within an agreed mandate but cannot simply rewrite that mandate.
We separate competence from authority because authority ultimately carries responsibility. So why should Autonomous AI be any different?
These thoughts led me to a concept I call Guided Autonomous AI. The principle is simple: AI should be given the freedom to operate autonomously, but only within boundaries explicitly defined by human authority. An AI agent may determine how best to achieve a delegated objective, but it should never determine the extent of its own authority. In short, AI may exercise autonomy, but humans must define its authority.
This becomes particularly important as AI agents begin interacting with the world. We are already hearing predictions of AI agents becoming independent economic actors: negotiating with other agents, purchasing goods and services, investing money, entering agreements and eventually settling transactions autonomously.
Technically, much of this may be possible. But technical possibility is not the same as societal desirability. Imagine giving an AI agent control of a corporate function. It might examine thousands of suppliers, negotiate prices, select the optimum contract and make payment in seconds. That could be extraordinarily useful. But should it be permitted to spend £10,000? £1 million? £100 million? Could it enter a five-year contractual commitment? Transfer funds overseas? Borrow money? And if something goes catastrophically wrong, who is responsible? Every significant transaction ultimately needs an accountable human principal.
This is why I believe advanced AI agents should contain a Schedule of Authority. The concept is borrowed from something businesses have used for generations. Organisations routinely define delegated authorities: who may approve expenditure, who can sign contracts, which decisions require board approval and when matters must be escalated. An AI agent should have an easily traceable, digital equivalent embedded into its architecture.
Its Schedule of Authority could specify which decisions it may take autonomously, the financial limits within which it may operate, which actions require human approval, which decisions it must never make independently, when escalation is mandatory and, crucially, which human remains accountable. Before executing a consequential action, the AI agent would therefore ask a question subtly but profoundly different from the one we tend to design around today. Not simply: “Can I do this?” But: “Am I authorised to do this?”
That distinction could become one of the most important architectural principles of the Agentic AI era. Consider commercial aviation. Modern aircraft can perform much of a flight automatically. Yet we have not concluded that because an aircraft can fly autonomously, human authority has become unnecessary.
Or consider medicine. AI may become extraordinarily good at interpreting scans, identifying patterns and recommending treatments. But capability alone does not automatically confer the authority to make every consequential decision affecting a patient’s life. The same principle applies across financial services, defence, critical infrastructure, law and government. Some decisions should remain human not necessarily because humans will always make them better, but because societies depend upon identifiable responsibility.
This is why I think the AI governance debate needs to move beyond its current emphasis on safety, explainability and transparency. These are important, but there is a deeper issue emerging: sovereignty. As AI becomes capable of acting rather than merely advising, we need to determine where decision-making sovereignty ultimately resides. We should resist the two extremes. AI does not need to remain a passive tool requiring human permission for every action. That would destroy much of its value. Equally, we should not assume that technological progress requires removing humans from every decision loop.
The objective should instead be maximum autonomy within human authority. We already know how to govern powerful human agents inside corporations, governments and institutions. We give them mandates, responsibilities, limits and escalation procedures. We now need to apply the same thinking to artificial agents. The defining question of the Agentic AI era may therefore not be how intelligent our machines become, nor even how autonomous they become. It may be whether we remember the difference between autonomy and authority. Autonomy is a capability. Authority is delegated. Accountability is human. Preserve those distinctions and AI agents could become extraordinarily powerful extensions of human capability without becoming substitutes for human sovereignty.
Fail to preserve them and we risk something more troubling: consequential decisions being made by autonomous systems without meaningful human responsibility. That is not merely an AI safety problem. It risks creating a form of digital tyranny, authority exercised without accountability. The future, therefore, should not simply be autonomous AI. It should be Guided Autonomous AI.
Jamil El-Imad is a special situations advisor at Astraea, a legal and professional services firm & Honorary Senior Research Fellow at the Institute of Biomedical Engineering, Imperial College.
