AI in commodity trading: a perspective

Artificial intelligence is neither magic nor hype in commodity trading; it is a set of tools that help in specific places and mislead in others. This paper takes a measured view of where machine learning and, increasingly, agentic AI add genuine value to a trading and risk operation, where they should be kept away from the numbers of record, and what everything useful about them depends on. Its central argument is that the value of AI on a trading desk is bounded above by the quality and governance of the data beneath it: a model trained or prompted on inconsistent data inherits the inconsistency, and no amount of algorithmic sophistication repairs a broken foundation.

The paper frames what AI is and is not for a trading desk, examines the data foundation that everything depends on, surveys where machine learning genuinely helps and where caution and explainability must lead, considers the emerging role of agentic AI and copilots in operations, sets out the governance and model-risk discipline that responsible adoption requires, describes the architecture and MLOps that make AI safe to run, and closes with a measured adoption roadmap. It is written to be useful to trading and risk leadership, data leaders, quants, and operations managers who must separate the real from the oversold.

The paper is written for the people who must decide where, and whether, to apply AI in a trading operation, and who are accountable for the consequences. Trading and risk leadership will find a frame for judging AI proposals on their merits rather than their marketing. Data leaders will find the argument, useful internally, that data governance is the precondition for AI value, not a competing priority. Quants and model owners will find the boundary drawn between where statistical learning helps and where transparent models must remain. And operations managers will find the most immediately practical applications, in data operations and anomaly detection, laid out.

How the paper is organised

1. Framing. What AI is and is not for a trading desk.
2. The data foundation. Why governed data is the precondition for everything.
3. Where AI genuinely helps. Forecasting, anomaly detection, data operations, language.
4. Where caution must lead. Pricing, limits, and regulatory numbers.
5. Agentic AI and copilots. The emerging operational role and its guardrails.
6. Governance and model risk. Validation, explainability, and regulation.
7. Building it safely. Architecture, BYOK, and MLOps.
8. A measured adoption roadmap. Where to start and how to sequence.

The posture of this paper is deliberately measured. It is neither an AI-skeptic tract nor a booster document. The aim is to help a serious operation apply AI where it genuinely helps, insist on governed data underneath it, and keep humans accountable for consequential decisions.

Before asking where AI helps, it is worth being clear about what is meant by AI in this context, because the term spans very different things with very different risk profiles. At one end sit well-understood statistical and machine-learning models, regression, classification, clustering, time-series methods, that have been used in trading for decades. In the middle sit modern deep-learning models for forecasting and pattern recognition. At the far end sit large language models and the agentic systems built on them, which can read, summarise, draft, and increasingly act. Each has a different balance of capability, transparency, and risk, and lumping them together is the first mistake.

What AI is good at

The genuine strengths of machine learning are pattern recognition in large, noisy datasets; automation of tasks that are repetitive but require judgment; and, for language models, the processing of unstructured text and the generation of fluent summaries and drafts. These are real capabilities, and where a trading operation has a problem shaped like one of them, AI can add substantial value.

What AI is not good at

Equally important is what AI is not good at, or not to be trusted with. It is not good at producing numbers that must be exactly right and fully explainable, because most powerful models are statistical and opaque. It is not good at guaranteeing behaviour, because it generalises from data rather than following rules. And it is not a substitute for accountability: a model cannot be held responsible for a decision, so a human must be. On a trading desk, where some numbers are of record and some decisions are consequential and regulated, these limits are not abstract; they define where AI belongs and where it does not.

The right mental model

The most useful way to think about AI on a trading desk is as an assistant and an accelerant, not an oracle. It excels at the tasks around the edges of the consequential decisions, cleaning the data, surfacing the anomalies, drafting the summaries, answering the questions, and it thereby frees human experts to focus their judgment where it matters. It is at its most dangerous when treated as a source of truth for numbers that must be exactly right. Holding this distinction clearly is what separates a productive AI strategy from an expensive and risky one.

Every useful application of AI in this paper rests on the same foundation, and it is worth stating the principle before the applications, because it governs all of them: the value of AI is bounded by the quality and governance of the data beneath it. This is the single most important idea in the paper, and the one most often ignored in the rush to adopt.

Garbage in, garbage out, at scale

A machine-learning model learns the patterns in its training data, including the errors. If the data is inconsistent, if the same trade appears differently in different systems, if reference data is stale, if prices come from an ungoverned source, the model learns the inconsistency and reproduces it, often amplified and harder to detect because it is now buried in a statistical model rather than visible in a report. The old maxim, garbage in, garbage out, applies with special force to AI, because AI industrialises whatever it is fed. A model built on a fragmented, unreconciled data landscape does not rise above that landscape; it inherits it.

Why governed data matters more, not less

It is sometimes assumed that AI reduces the need for clean data, that a sufficiently clever model can see through the noise. The opposite is true. Because AI acts on data at scale and with reduced human oversight, the consequences of bad data are larger and less visible, not smaller. A human analyst reading a report will often notice that a number looks wrong; a model consuming millions of records will not. This is why a single governed model and a trustworthy analytical layer matter more in an AI-enabled operation, not less: they are what make the data foundation sound enough to build on.

Lineage and features

Two properties of governed data are especially important for AI. The first is lineage: knowing where every data point came from, so that a model’s inputs can be traced and its behaviour explained. A model whose inputs cannot be traced cannot be validated or trusted for anything consequential. The second is the discipline of features: the derived inputs a model consumes, computed consistently and reproducibly from the governed data, so that the same feature means the same thing in training and in production. An ungoverned feature pipeline is a common and hidden source of model failure, where a model trained on one definition of a quantity is fed a subtly different one in production.

The foundation is the strategy

The practical conclusion is that the first step of an AI strategy is not to acquire models but to govern data. An organisation with a governed model and a trustworthy analytical layer is ready to apply AI wherever it genuinely helps; an organisation without one will find that its AI initiatives stall, mislead, or quietly fail, not because the models are inadequate but because the foundation is. Data governance is not a competing priority to AI; it is the precondition for it.

With the framing and the foundation established, the applications where AI adds real value on a trading desk come into focus. They share a common shape: they are tasks where pattern recognition or language processing at scale genuinely helps, and where the output assists a human rather than becoming an unchecked number of record.

Data operations

The most immediate and least glamorous application is in data operations: classifying, matching, and cleaning trade and reference data. Matching confirmations to trades, resolving counterparty names across sources, classifying instruments, spotting and correcting data errors, these are tasks that are repetitive, high-volume, and require judgment, exactly the shape machine learning suits. The value is direct: less manual operations effort, fewer breaks, cleaner data, which in turn improves everything downstream. Because the output feeds a governed process with human oversight of exceptions, the risk is contained.

Anomaly detection

Closely related is anomaly detection: surfacing unusual trades, prices, positions, or settlement breaks that merit a human look. A model that has learned the normal patterns of a desk’s activity can flag the abnormal, a price far from where it should be, a trade unlike the desk’s usual business, a settlement that does not fit, far faster and more consistently than manual review. This is valuable across the operation: in surveillance, in data quality, in risk, in operations. Again, the model surfaces candidates for human attention rather than making final decisions, which is the safe and correct division of labour.

Forecasting

Forecasting is where AI meets the trading decision most directly, and where it must be handled with the most care. Machine learning can improve forecasts of the fundamental drivers of commodity prices, load, generation, renewable output, demand, where there is abundant data and genuine pattern. A better load or wind forecast has real trading value. But a forecast is an input to a human decision, not a decision itself, and the temptation to let a model trade directly on its own forecasts is where forecasting shades into danger. The value is in better inputs to human judgment; the risk is in abdicating the judgment.

Language and knowledge

The newest application is language: using large language models to summarise, draft, and answer questions over the organisation’s own governed data. A trader or analyst asking a question in plain language and receiving an answer computed from the governed model, a risk manager getting a drafted commentary on the day’s moves, an operations analyst summarising a batch of confirmations, these are genuine productivity gains. The crucial qualifier, developed in the next module, is that the language model must answer from governed data with its sources traceable, not from its own opaque memory, or it becomes a confident source of plausible errors.

For every application where AI helps, there is a domain where it must be kept away from the numbers of record, and drawing this line clearly is as important as finding the opportunities. The principle is simple: where a number must be exactly right, fully explainable, and auditable, it should come from a transparent model on governed data, with human accountability, not from an opaque statistical one.

Pricing

The prices and values of record, the marks that drive P&L, margin, and reporting, must be produced by transparent, validated pricing models whose every output can be explained and reproduced. A black-box model that produces a value no one can explain is unacceptable as a mark of record, however accurate it may appear, because when it is wrong, and every model is sometimes wrong, no one can see why or correct it. AI may assist around pricing, suggesting, checking, flagging, but the number of record comes from a transparent model.

Risk limits and controls

Risk limits and the controls that enforce them are, by their nature, rules that must behave predictably and be explainable to regulators and boards. A limit that is enforced by an opaque model, whose behaviour cannot be fully predicted or explained, is not a control; it is a new risk. The measurement of risk may use sophisticated methods, but the limits and controls built on it must be transparent and deterministic.

Regulatory reporting

Regulatory numbers demand explainability and auditability by law. A regulator asking how a reported figure was produced expects a traceable answer, not a statement that a model produced it. Regulatory reporting must therefore run on transparent logic over governed, lineage-tracked data, with AI confined at most to assisting the process, never to producing the numbers themselves.

The explainability principle

Underlying all three is a single principle: the more consequential and regulated a number, the more it must be explainable, and explainability and model opacity are in tension. This does not banish AI from the regulated core; it positions it correctly, around the transparent numbers of record rather than in place of them. The mature stance applies AI where its opacity is tolerable because a human checks its output, and keeps it away from where its opacity is not, because the number stands on its own and must be defended.

python# Numbers of record come from transparent models; AI assists, never decides.
def price_of_record(instrument, market):
    return transparent_pricer.value(instrument, market)   # auditable, reproducible

def ai_assist(instrument, market, context):
    suggestion = llm.summarize(context)      # helps a human
    anomaly    = anomaly_model.score(market) # flags for review
    return Assist(suggestion=suggestion, anomaly=anomaly, decides=False)

The most recent development, and the one attracting the most attention and the most hype, is agentic AI: systems built on large language models that do not only answer questions but take actions, calling tools, executing multi-step workflows, and operating with a degree of autonomy. In a trading operation the promise is significant and the risk is equally so, and a measured view requires separating the genuine opportunity from the dangerous overreach.

The copilot pattern

The safest and most immediately valuable form is the copilot: an AI assistant that works alongside a human, proposing and drafting but not deciding. A trading copilot might assemble the context for a decision, pulling together positions, risk, market data, and relevant history in response to a plain-language request. An operations copilot might draft the resolution of a settlement break for a human to approve. A risk copilot might generate a first draft of the daily commentary. In each case the human remains the decision-maker, and the AI compresses the time from question to informed action. Because the human approves every consequential step, the copilot pattern keeps accountability where it belongs.

Agentic workflows in operations

Beyond the copilot, agentic systems can automate multi-step operational workflows that are currently manual: triaging exceptions, gathering the information needed to resolve a break, preparing a reconciliation, routing an item to the right team. These back-office and middle-office workflows are attractive first targets for agentic AI precisely because they are structured, high-volume, and lower-stakes than the trading decision itself, and because a human can review the outcome. The value is in removing operational drudgery; the guardrail is that consequential actions require human confirmation.

The guardrails agentic AI requires

Autonomy demands guardrails proportional to the stakes. An agentic system that can act must operate within explicit boundaries: what it may do without human approval, what it must escalate, and what it may never do. It must act on and through governed data and supported interfaces, so that its actions are traceable and reversible, not reach into systems through unsupported back doors. Every action it takes must be logged in the same audit trail as a human’s, so that the operation can reconstruct what the agent did and why. And the more consequential the action, the more certainly a human must remain in the loop. Agentic AI without these guardrails is not an efficiency; it is an uncontrolled actor in a regulated business.

BYOK and keeping control

A practical architectural point deserves emphasis: an operation adopting AI, and especially agentic AI, should retain control of its models and its data. A bring-your-own-key approach, where the organisation supplies and controls the AI model credentials rather than surrendering its data to an opaque embedded service, keeps the organisation in command of which model is used, what data it sees, and how it is governed. Control of the model and the data is not a detail; it is what allows the governance the rest of this paper insists on.

AI models are models, and the discipline of model risk management, developed over decades for the quantitative models of finance, applies to them directly and with additional force, because AI models are often more opaque and more data-dependent than the models that discipline was built for. Responsible AI adoption is, in large part, the extension of existing model governance to a new and more demanding class of model.

Model risk management

The core of model risk management is that every model in consequential use is inventoried, validated independently, monitored in production, and owned by an accountable person. This applies to an AI model exactly as to a pricing model: it must be documented, its assumptions and limitations understood, its performance monitored, and its use bounded to what it has been validated for. The temptation with AI is to treat it as different, as too novel or too complex for the existing discipline, but that is precisely backwards: its novelty and complexity make the discipline more necessary, not less.

Explainability and validation

AI models pose a specific challenge to validation: many are opaque, and an opaque model is hard to validate and harder to explain. This is not a reason to abandon validation but a reason to demand explainability where the stakes require it, and to confine the most opaque models to the least consequential uses. A model whose decisions cannot be explained has no place producing numbers of record, however well it backtests, because a model that cannot be explained cannot be defended when it fails, and it will sometimes fail.

Bias, drift, and monitoring

Two failure modes deserve particular attention. Bias is the risk that a model has learned a systematic distortion from its training data, producing skewed outputs that may be hard to detect. Drift is the risk that the world changes so that the patterns a model learned no longer hold, and its performance silently degrades. Both require ongoing monitoring in production, comparing the model’s behaviour against reality and against expectation, so that a model that has gone wrong is caught rather than trusted. A model deployed and forgotten is a liability accumulating quietly.

The regulatory horizon

The regulation of AI is developing, and a responsible operation anticipates it rather than waiting for it. Frameworks such as the European Union’s AI Act introduce obligations that scale with the risk of the application, and financial-sector model-risk guidance increasingly encompasses AI. The prudent posture is to govern AI to a standard that will satisfy the regulation that is clearly coming: documented, validated, monitored, explainable where it matters, and with human accountability throughout. An operation that governs AI well is not only safer; it is ready for the regulatory environment that is arriving.

The principles of this paper translate into concrete architectural and operational requirements. Building AI into a trading operation safely is not primarily a matter of choosing the cleverest model; it is a matter of putting the model in a structure that governs its data, controls its actions, and monitors its behaviour. This module sets out that structure.

AI on the governed model

The first requirement follows from the whole paper: AI must sit on the governed data model and the trustworthy analytical layer, consuming governed, lineage-tracked data through supported interfaces. This is what makes its inputs traceable, its behaviour explainable, and its outputs consistent with the rest of the system. An AI capability bolted onto a fragmented data landscape inherits the fragmentation; one built on a governed model inherits the governance. The architecture of the companion cloud-native paper is not incidental to AI; it is its precondition.

The MLOps discipline

Running machine-learning models in production is its own discipline, often called MLOps, and it extends the ideas of reproducibility and versioning to models and their data. A production model needs versioning, of the model, its training data, and its features, so that any prediction can be reproduced and any regression diagnosed. It needs a pipeline that retrains and redeploys under control, with validation gates, rather than ad hoc updates. It needs monitoring of performance and drift in production. These are the same disciplines of reproducibility and control that govern the rest of a serious trading platform, applied to models.

Control of models and data

The bring-your-own-key principle introduced earlier is worth restating as an architectural requirement: the organisation controls which models are used and what data they see. This keeps the choice of model, and the governance of the data it consumes, in the organisation’s hands rather than an opaque vendor’s. It also allows the organisation to change models as the fast-moving AI landscape evolves, without being locked into one provider’s embedded and ungoverned service.

Human accountability by design

Finally, the architecture must encode human accountability rather than leaving it to policy. Consequential actions require human approval as a matter of system design, not good intentions; the boundaries of autonomous action are enforced, not merely documented; and every AI action is logged in the audit trail alongside human ones. Accountability that depends on people remembering to check is not accountability; accountability built into the system is.

The paper closes with the practical question its audience will most want answered: given all of this, where should an operation actually start, and how should it sequence its adoption of AI? The answer follows directly from the argument, and it is deliberately measured.

Start with the foundation

The first step is not an AI project at all; it is the governed data model and trustworthy analytical layer that everything else depends on. An operation without this foundation should build it before reaching for models, because AI applied to ungoverned data will disappoint at best and mislead at worst. An operation that already has it is ready to proceed. This ordering frustrates those who want to start with the exciting part, but it is the difference between AI that compounds value and AI that quietly fails.

Then the low-risk, high-value applications

With the foundation in place, the right first applications are those where value is high and risk is contained: data operations and anomaly detection, where AI relieves operational burden and surfaces issues for human attention without producing numbers of record. These build capability, demonstrate value, and develop the organisation’s judgment about AI, all while the stakes are moderate and the human remains firmly in control.

Then assistance and forecasting, carefully

Next come the copilots and the forecasting applications: AI that assists human decisions with summaries, drafts, answers, and better inputs. These deliver broad productivity gains and bring AI into the daily work of the desk, still as an assistant to human judgment rather than a replacement for it. Forecasting in particular should be introduced as an input to decisions, with the discipline to keep it there.

Agentic automation last, and bounded

Agentic automation of operational workflows comes last and most carefully, bounded by the guardrails set out earlier: explicit limits on autonomous action, action only through governed interfaces, full logging, and human confirmation of consequential steps. It is introduced where the workflows are structured and the stakes are contained, and expanded only as the organisation’s confidence and controls mature.

The measured stance, restated

The through-line of the roadmap, and of the whole paper, is a measured stance: apply AI where it genuinely helps, insist on governed data underneath it, keep humans accountable for consequential decisions, and sequence adoption from the foundation outward rather than from the hype inward. An operation that adopts AI this way captures its real value, in productivity, in cleaner data, in faster insight, without taking on the risks that undisciplined adoption invites. That is the opportunity, correctly understood: not a revolution that replaces human judgment, but a powerful set of tools that, on a governed foundation and under human accountability, make a trading operation faster, cleaner, and sharper.

AI in commodity trading is real, valuable, and easy to get wrong. Its genuine strengths, pattern recognition at scale, automation of judgment-laden tasks, processing of language, map onto real opportunities on a trading desk: cleaner data operations, faster anomaly detection, better fundamental forecasts, and productivity gains from copilots and assistants. But every one of these depends on a foundation of governed, lineage-tracked data, and every one is safest when it assists a human rather than replacing the judgment behind a consequential, regulated decision.

The measured stance this paper has argued for is not caution for its own sake. It is the recognition that AI amplifies whatever it is built on: sound data and good governance, or fragmentation and opacity. Applied on a governed model, around transparent numbers of record, under real human accountability, and adopted from the foundation outward, AI is a genuine and compounding advantage. Applied to ungoverned data, in place of transparent models, without accountability, and chased for its novelty, it is an expensive and dangerous distraction. The difference between the two is not the sophistication of the model but the discipline of the operation that deploys it.

This paper draws on the established disciplines of model risk management, data governance, and machine-learning operations, and on the developing landscape of AI regulation. Specific regulatory texts should be consulted in their current published form.

• Supervisory guidance on model risk management, for the validation, monitoring, and accountability principles of Module 6.
• The European Union’s AI Act and comparable emerging frameworks, for the risk-based regulatory obligations referenced in Module 6.
• The literature and practice of machine-learning operations (MLOps), for the reproducibility and versioning disciplines of Module 7.
• Companion Gravitas whitepapers: Cloud-native ETRM architecture, for the governed model and analytical layer this paper treats as AI’s precondition, and Quantitative risk in commodities, for the transparent risk models AI must sit alongside rather than replace.

The Gravitas ETRM and CTRM data dictionary and the overview of how AI is applied in the platform are useful practical companions to this paper.