For decades, the ambition of artificial intelligence has been to make machines increasingly autonomous, to reduce the need for continuous human supervision and enable systems to perceive, reason, decide and act independently. From the earliest expert systems of the 1970s to today’s AI agents capable of planning, coding, negotiating and executing complex workflows, the trajectory has been remarkably consistent: to move machines progressively from being passive tools to becoming active participants in human activities. This pursuit of autonomy has already transformed many aspects of modern life. Industrial robots now assemble vehicles with minimal human intervention. Autonomous drones inspect power lines, pipelines and offshore wind farms without requiring pilots. AI systems monitor financial markets and execute trades within milliseconds based on predefined strategies. Self-driving vehicles are being tested on public roads whilst autonomous agricultural machines plant, irrigate and harvest crops with extraordinary precision. In healthcare, AI can analyse medical images, prioritise patient cases and recommend treatment pathways faster than many traditional workflows. Even in software development, AI agents are increasingly capable of writing code, testing applications, fixing bugs and deploying updates with only high-level human instructions.
Perhaps the clearest recent example is the rapid rise of agentic AI. Indeed, we are already seeing AI agents becoming economic participants - Coinbase’s x402 protocol enables AI agents to authenticate, negotiate and pay for APIs, datasets and digital services using stablecoins without human intervention. Unlike traditional chatbots that simply respond to prompts, agentic systems can pursue goals over extended periods, break complex objectives into smaller tasks, call external tools, access databases, collaborate with other AI agents and adapt their strategies as circumstances change. Rather than merely answering questions, they can complete projects. The commercial momentum behind this vision is immense. Companies such as OpenAI, Google, Microsoft, Anthropic, Tesla and many others are investing billions of dollars to develop increasingly capable autonomous systems. Governments are similarly exploring autonomous AI for public services, defence, scientific research, infrastructure management and economic planning. The race is no longer simply to build smarter machines; it is to build machines that require less and less human intervention. At first glance, this appears to be an entirely logical progression. Greater autonomy promises greater efficiency. Machines do not tire, forget, become distracted or require holidays. They can operate continuously, process enormous quantities of information and often respond far faster than human operators. In environments where speed, precision and scale are paramount, increasing autonomy can produce extraordinary economic and societal benefits. Yet, I believe, this pursuit raises an important question that has received far less attention than the engineering itself. Should every decision be made autonomous simply because it can be?
History suggests otherwise. Humanity has repeatedly developed technologies that are technically capable of complete autonomy, yet deliberately chooses to keep humans within the decision loop whenever the consequences are significant. Commercial aircraft can fly for most of a journey using autopilot, but qualified pilots remain responsible for critical decisions. Modern hospitals rely heavily on AI-assisted diagnostics, yet physicians remain accountable for diagnoses and treatment decisions. Even biological systems such as highly trained police dogs, despite years of specialised training, are never regarded as autonomous decision-makers; they remain extensions of their handlers’ judgement and authority. Corporate treasury systems such as SAP and Oracle already use approval workflows too. Payments.com recently wrote: “Agentic commerce turns corporate API calls, software usage, data checks, AI tasks and back-office workflows into payable events, making finance and procurement the first serious test cases of AI agents in a wholesale, not retail, setting.”
These examples point to a distinction that may become one of the defining challenges of the AI age. The objective should perhaps not be to maximise machine autonomy, but to determine where autonomy is appropriate and where it must remain guided by human authority, accountability, and values. I believe payment systems and financial transaction fall into this category. I believe that the next chapter of artificial intelligence may not be about creating machines that operate entirely on their own. It may instead be about designing machines whose autonomy is explicitly bounded, governed and accountable, a philosophy that could be described as Guided Autonomous AI Systems, where machines are empowered to act independently within carefully defined limits whilst ultimate responsibility always remains with the humans they serve. Today the narrative is becoming familiar. We celebrate every milestone where AI requires less human intervention, makes more decisions on its own and carries out increasingly complex tasks without supervision. But perhaps we have been heading in the wrong direction. The issue is not:
“How autonomous should AI become?” The more important question is: “Over what decisions should humans never relinquish sovereignty?”
There is a profound difference between an AI acting for a human and an AI acting instead of a human. Unfortunately, this distinction is becoming increasingly blurred as we dive into the age of agentic AI. Autonomy is not the problem. I believe autonomy itself is not dangerous. In fact, autonomy is exactly what makes AI valuable. An AI agent that can organise meetings, optimise supply chains, negotiate prices, analyse medical images, draft legal contracts, monitor industrial systems or manage logistics without constant supervision can dramatically improve productivity. Autonomy is a capability. It enables machines to perform tasks faster, more consistently and often more accurately than humans. The problem begins when capability is mistaken for authority. In this regard there is a big difference between capability and authority. Consider a highly trained police dog, an example I referred to earlier. It can search buildings, detect explosives, track suspects and make split-second operational decisions. It displays remarkable autonomy within its area of expertise. Yet nobody considers the dog an independent authority. Why? Because every action ultimately remains under the authority of its handler. The moment the dog ceases to obey lawful commands, it can no longer serve in that role. Its operational autonomy becomes a liability because the chain of accountability has been broken.
The same principle should apply to AI. Indeed, back in 2022, Anthropic proposed that AI ought to operate under predefined principles. In the EU, the AI Act is proposing, “high-risk transactions” using AI requires meaningful human oversight and the US Government has already recognised use of AI needs governance, accountability, human oversight, risk controls and verification. In business, Visa and Mastercard believe that if AI Agents can have authority to transact, they must do so within predefined parameters. So, whilst an AI agent may independently perform thousands of operational decisions every hour, those decisions should always remain within a clearly defined framework of delegated human authority. We need to move away from this illusion of AI becoming an independent economic actor, something a number of commentators are suggesting. In this scenario AI agents will soon become economic actors in their own right - negotiating contracts, making purchases, investing capital and settling disputes with little or no human intervention.
Technically, this may become possible. But should it?
When it comes to economic value, history has taught us that every financial transaction ultimately carries responsibility. Every payment has someone behind it. Every contract has someone accountable. Every mistake has someone liable. The danger is not that AI becomes capable of spending money. The danger is that society gradually accepts systems that spend money without a clearly identifiable human sovereign. And this is not merely a technological shift. It is a constitutional one. I believe when it comes to financial and other equally important transactions, guided autonomous AI is what is needed. Rather than pursuing completely autonomous AI, we should exercise a degree of control. Under the model of “Guided Autonomous AI”, the system enjoys extensive operational freedom whilst remaining strategically accountable to an identifiable human principle. AI is free to decide how to achieve delegated objectives. It is not free to redefine who is responsible. This distinction may appear subtle. In reality, it changes everything.
As such, in my view, when it comes to financial transactions, every AI agent will need a schedule of authority. Today’s AI systems are largely designed around prompts. We tell them what to do. We give them access to tools. We connect them to databases and APIs. But we rarely define what they are authorised to decide. I believe every significant AI agent should contain a mandatory architectural component that I call a “Schedule of Authority”.
Just as organisations define financial approval limits, signing authority and delegated responsibilities for employees, AI systems should operate within clearly defined boundaries established by their human principals. A Schedule of Authority might specify the financial limits AI may authorise, decisions requiring explicit human approval. It should specify actions that must never be performed autonomously. And the rules for mandatory escalation. Today companies pass Board Resolutions. Tomorrow Boards may approve:
Resolution 2028-14 - the Chief Procurement AI Agent is authorised to enter contracts up to £5 million under English law with investment-grade counterparties, provided settlement occurs through approved regulated payment rails and all transactions are recorded within the firm’s AI Authority Register. Artificial intelligence is rapidly becoming capable of negotiating contracts, allocating capital and initiating payments without continuous human intervention. The defining question is no longer what AI can do, but what it should be authorised to do. As AI evolves from an assistant into an economic actor, capability must never be confused with authority. Every significant AI system should therefore operate within a machine-readable guided autonomous AI schedule, ensuring that important legal, financial and strategic decisions always remain accountable to an identifiable human principal. Before any material action is executed, AI should first ask, not “Can I do this?” but “Am I authorised to do this?”
This is not simply a technical challenge - it is a constitutional one. Just as modern organisations are governed through delegated authority and clear accountability, the AI economy will require equivalent governance embedded into its architecture. The future will not be won by creating the most autonomous AI, but by creating the most trusted AI. In the age of agentic commerce, autonomy is a capability, accountability is a necessity - and human sovereignty must remain the foundation upon which artificial intelligence is built.
This article first appeared in Digital Bytes (21st of July, 2026), a weekly newsletter by Jonny Fry of Team Blockchain.