From trading terminal to AI agent: who owns the decision?

For most of the history of electronic markets the division of labour was clear. The terminal delivered information, the software calculated, and a human decided. The contributed piece below argues that generative and agentic AI compress those three stages into one interaction, so that a system which ingests market data, interprets news, assesses exposure and suggests an action is a single step from executing it.

The author's concern is not that such a system will produce an obviously irrational trade. It is that a plausible recommendation for the wrong reason, approved by a human who has neither the time nor the information to challenge it, leaves nobody clearly accountable when the loss arrives. His proposal is architectural: decide in advance what an AI agent may see, recommend and execute, and when its authority is withdrawn.

Amir Naser Hojati is a futures trader and fintech practitioner with more than 15 years of experience in financial markets, including CME futures, and the founder of a company focused on automated trading technology. His work focuses on translating discretionary trading concepts, market behaviour and risk decisions into systematic and automated trading frameworks. The article that follows is a contributed piece and presents his opinion.

For decades, the relationship between traders and technology was relatively clear. The terminal delivered information. The software calculated. The trader decided.

Artificial intelligence is beginning to blur that division. As brokers, exchanges and financial platforms connect AI systems to live market data and trading workflows, investors can increasingly move from asking what is happening in a market to asking what action should come next.

That may look like a natural evolution of the trading terminal. It is more significant than that. The important question is no longer simply whether AI can analyse financial markets. It is what happens when the system providing the analysis becomes part of the chain that produces the investment decision. At that point, firms need to answer a deceptively simple question: who actually owns the trade?

From information to authority

Traditional trading technology largely separated information from judgement. A charting platform could identify a price breakout. A risk system could calculate exposure. An execution algorithm could divide an order into smaller pieces. But the investment decision generally came from somewhere else: a trader, portfolio manager or predefined systematic strategy.

Generative and agentic AI compress those stages. An AI assistant can ingest market data, interpret news, compare instruments, assess portfolio exposure and suggest an action in the same interaction. The next step is obvious: connect that assistant more closely to execution.

The efficiency case is powerful. But the closer AI moves towards the trade itself, the less useful it becomes to describe it simply as another analytical tool. A system that says "volatility has increased" is providing information. A system that says "reduce this position" is influencing capital allocation. A system that can carry out that instruction has crossed another boundary entirely. Those three activities should not be governed as though they were the same.

The final click may not mean control

There is a temptation to draw the line at execution: as long as a human presses the final button, the human remains in control. In practice, the boundary is less comfortable.

If an AI system selects the relevant information, frames the market conditions, ranks the alternatives and proposes a specific action, much of the investment decision may already have taken place before the human approves it. The final click can create an illusion of control.

This matters because machine-generated recommendations can become increasingly persuasive as they become faster, more consistent and apparently more data-driven. Over time, independent human judgement can quietly turn into routine confirmation. The difficult risk is not necessarily that AI produces an obviously irrational recommendation. It is that it produces a plausible recommendation for the wrong reason. In markets, that distinction can be expensive.

Markets are not just data

The same observable signal can mean very different things under different market conditions. A rapid price move might reflect genuine information entering the market, temporary liquidity withdrawal, forced positioning, a reaction to another asset or a short-lived order imbalance. A model may recognise the pattern without fully understanding why the pattern exists.

That distinction becomes particularly important when discretionary trading decisions are translated into automated rules. After more than 15 years in futures markets, one of the harder lessons to encode is that experienced traders do not respond only to signals. They respond to context. Sometimes the right decision is not to act on a technically valid setup because the surrounding market has changed: liquidity is behaving differently, participation is unusual, or the normal relationship between price and volume has broken down.

Turning that judgement into code requires making explicit what was previously implicit. Giving an AI agent greater freedom does not remove that problem. It can simply hide it behind a more sophisticated interface.

The accountability gap

The deeper challenge appears when something goes wrong. Imagine an AI assistant recommends reducing a futures position. The portfolio manager accepts the recommendation and the trade is executed automatically. The result is a substantial loss. Where did the decision originate? Was it the model that interpreted the data, the manager who approved the action, or the institution that allowed the recommendation to become an executable trade?

Firms may eventually settle these questions legally. Operationally, they need answers much sooner. A useful principle is that delegating analysis should not automatically mean delegating authority.

Financial institutions already understand this distinction elsewhere. A risk model can influence a decision without possessing authority to make it. An analyst can recommend a position without being permitted to execute it. AI should not erase those boundaries simply because one interface can technically perform several functions.

The architecture matters more than the chatbot

The debate around AI in finance often focuses on model capability: which system reasons better, produces fewer errors or has access to more data. For trading, the more important question may eventually be architecture. What can the agent see? What can it recommend? What can it execute? And under what conditions is its authority removed?

These are not minor implementation details. They determine the difference between an AI assistant and an autonomous investment actor.

A robust architecture could separate the process into layers. AI might have broad freedom to analyse markets and challenge existing positions. It could have narrower freedom to propose changes. Execution authority could then be constrained by position size, instrument, volatility, risk budget or other predefined limits. Certain conditions, such as abnormal liquidity, unusual volatility, large model disagreement or breaches of risk limits, could automatically return the decision to a human.

The objective should not be to keep humans involved in every routine action forever. It should be to make clear where machine authority begins and where it ends.

Human oversight must be meaningful

"Human in the loop" is often presented as the answer to AI risk. But human oversight means little if the person involved has neither the time nor the information necessary to challenge the system. Real oversight requires the ability to understand why an action is being proposed, reject it without operational friction and recognise when the model is operating outside the conditions for which it was designed.

This matters especially in fast-moving markets. A human approval step added to a process operating at machine speed can become ceremonial rather than meaningful. The better question is not whether a human remains somewhere in the workflow. It is whether human judgement is preserved at the points where judgement actually matters.

The next trading interface

The traditional trading terminal was built around a simple assumption: technology would present the market to a human decision-maker. The emerging AI interface is built around a different possibility. Technology may increasingly interpret the market, narrow the choices, recommend the action and eventually execute it.

That could make trading more efficient. It could also make the location of the investment decision increasingly difficult to see. Financial firms should define that boundary before convenience defines it for them.

Because once AI moves from answering "What is happening?" to answering "What should I do?", the central question is no longer how intelligent the assistant is. It is how much authority we are prepared to give it, and who remains accountable when it is wrong.