Amplified, Not Replaced

Every few months, a new headline promises that an artificial intelligence will soon run a trading desk on its own. Anyone who has actually worked on an energy desk tends to smile at that. Energy is physical. A vessel is crossing an ocean, a pipeline is full, a heat wave is about to change demand. Numbers on a screen are the end of a long chain of real events, and the people responsible for them have to explain what they did and why.

AI in Energy Trading: The Real Opportunity

That is the starting point for a more realistic conversation about AI in energy and commodity trading, and the evidence suggests the industry is having it. In a survey of more than 400 respondents across over ten industries, published by Molecule in its 2025 ETRM and CTRM modernisation report, 91% said they expect AI to significantly affect their trading operations within five years. The use cases they named most often were automated reporting at 18.7%, algorithmic trading at 18.2%, predictive forecasting at 16.8% and fraud detection at 15.1%. Notice what is missing from the top of that list. Nobody is asking for a robot to take the desk.

The more immediate opportunity sits in the work around the trade. A modern trading business generates a constant stream of unstructured material: contracts to read, broker conversations to interpret, confirmations to match, exceptions to chase and reports to assemble before the market opens. People spend a surprising part of their day connecting information that already exists. Large language models are good at precisely this kind of reading and summarising, and they open up material that once needed a person’s eyes on every page. A trader who no longer waits hours for a position summary, or a risk team that stops reconciling three spreadsheets by hand, gains something very practical: time to think.

Why Explainability Beats Autonomy

There is a catch, and the industry knows it. Energy markets have physical constraints, extreme events and regulators who ask awkward questions. When a decision goes wrong, someone has to be able to explain why it was made. A model that produces a confident answer without a traceable reason is a poor fit for that environment. This is why many experienced practitioners prefer the phrase human expertise amplified by AI, rather than humans against machines. The system proposes, summarises and flags. The person decides and remains accountable.

Explainability, then, becomes a design requirement and not a nice extra. Firms need to know which data a recommendation used, which assumptions sat behind it, and how it compares with what an experienced trader would have done. That favours AI built into existing controls over AI bolted on at the end. It also favours teams who understand trading well enough to ask whether a process should exist at all before anyone automates it.

The Platform Constraint: Modernising Underneath

Which brings us to the unglamorous part. Most AI ambitions in this sector run straight into the platform underneath. In the same survey, 64% of respondents said their current ETRM system does not support all of their processes or traded commodities, and fewer than four in ten said it handles diverse trade types effectively. The share saying it cannot handle all of their commodities rose from 21% to 31% in a year. Operational visibility, at 31%, has now overtaken agility, at 23%, as the top long term risk priority.

The response has been rapid. Seventy per cent of respondents reported modernisation projects underway, compared with 24% the previous year, and 94% had launched or planned some kind of initiative. That jump says a lot. Companies have realised that a clever model sitting on top of fragmented data and manual workarounds will only produce faster confusion. Clean data flows, consistent position aggregation and trusted risk processing come first.

Fixing Friction Over Total Replacement

Interestingly, the best results often come from fixing friction rather than replacing everything. The textbook answer to an ageing platform used to be a multi year rebuild, with all the risk and cost that implies. A growing number of firms are looking for the specific points where work is slow, then adding targeted automation, better data flows and clearer controls around what already works. Nothing has to be ripped out. When it works, the technology becomes almost invisible, and people notice only that their day has become easier.

For leaders, the practical lessons are fairly simple. Start where the pain is measurable, such as reporting, reconciliation and exception handling. Build testing and reconciliation into the architecture, so that checks do not depend on someone remembering to run them. Keep a named human accountable for every material decision. And treat the data layer as a strategic asset, because every future use of AI will depend on it.

The industry will probably get its more ambitious tools in time. Until then, the firms that gain the most will be the ones that treat AI as a colleague with specific strengths, supervised by people who understand the business well enough to know when it is wrong.