At Energy Trading Week London 2026 we surveyed trading, risk and compliance professionals about AI. More than half of respondents (58%) expect the industry to deploy AI at scale within two years. Three quarters (75%) are not yet confident that AI outputs in risk and compliance can be trusted. So what is holding it back, and where will it help first.
For every entry submitted, BroadPeak made a £5 donation to the Fairtrade Foundation, which works to ensure farmers and producers in developing countries receive fair prices, safe working conditions, and better rights for the commodities they grow.
Key findings
- Strong views on timing, same barrier: data Every respondent who said AI is ready now or not ready (question 1) named data quality and fragmented architecture as the biggest barrier (question 2). Those who expect scale in one to five years almost all named governance, skills or integration instead.
- Data first, governance next. Asked for the biggest barrier to AI delivering value (question 2), half chose data quality and 25% chose governance. Asked about trust (question 4), 75% said they are not yet confident in AI outputs in risk and compliance because governance frameworks are not mature enough. Data is the first barrier people name, and governance comes up in both questions.
- The workflow that benefits most is where the data problem starts. Half chose trade capture and post trade processing as the workflow that will benefit most from agentic AI (question 3). Most of that group named data quality as the biggest barrier in question 2.
- Controls come next. After trade capture (50%), the workflows seen as most likely to benefit from agentic AI (question 3) were risk limit monitoring and exception routing (25%) and credit checks and counterparty approvals (17%). Scheduling drew 8%.
- Surveillance leads post trade. Position limits drew nothing. Asked where AI will have the greatest impact on post trade operations (question 5), half chose trade surveillance. Reconciliation drew 17% and nobody chose position limit monitoring.
Q1: How close is energy trading to deploying AI at scale?
Just over half (58%) expect AI at scale within two years, and a quarter say adoption is already accelerating. The rest see a longer road. A quarter (25%) expect three to five years because significant barriers remain, and 17% say the industry is not ready because the fundamentals need fixing first.
What separates the two groups is how much they think must be fixed before AI can run at scale. Those who expect AI soon are still cautious. Most of those who said AI is ready now are not yet confident in AI outputs in risk and compliance (question 4). For them, ready means ready to use with a person checking the result.
The optimism has some evidence behind it. BCG found in July 2026 that power and pipeline gas traders lead AI adoption in energy, because their work is already numeric [1]. Oliver Wyman reported in April that tech native trading houses run at 25% of the seat cost of traditional setups [2]. It also found that productivity gains above 20% come from building AI into workflows and controls, and that isolated pilots fall short of that.
Agents are already live in narrow back office tasks. One US gas trading operation cut monthly pipeline invoice processing from 18 to 20 days down to one day, with an agent that extracts, standardises and reconciles invoices across more than 300 pipelines [3].
Regulation is catching up slowly. In May 2026 Ofgem updated its guidance on ethical AI use in the energy sector [4]. It is voluntary and sits alongside existing rules, including REMIT. It asks firms to set clear accountability for AI across its life cycle, build in human oversight from the start, and explain how their AI works in proportion to the risk. Under Article 5a of revised REMIT, firms that trade algorithmically must notify their national regulator and ACER, and keep effective systems and risk controls that are tested and monitored [5]. There is still no AI rulebook for energy trading. Firms have to work out for themselves how existing obligations apply to AI, and that gap shows up in the trust results below.
The early wins are narrow, back office and easy to check. What the market has yet to show is AI at scale in risk and compliance, where every output has to stand up to an auditor.
Q2: What is the biggest barrier to AI delivering value in energy trading operations?
Half of respondents (50%) named data quality and fragmented architecture. Governance and controls around AI outputs followed with 25%, skills with 17% and integration with 8%. The table groups respondents by how close they think the industry is to AI at scale (question 1), then shows which barrier each group named in this question.
Those who say AI is ready now think the models work and the data is the only thing left to fix. Those who say it is not ready think the data will not be fixed soon enough. The middle group sees an operating model problem. Who approves a model for production use? Who in the team can run and challenge it? How does it connect to the E/CTRM and the downstream control systems?
Firms have to fix both. Fix governance but not the data, and you have an approved model working from bad data. Fix the data but not governance, and you have clean data that no model is allowed to use. Solve data quality without governance and you have clean data that no model is allowed to use.
Q3: Which workflow will benefit most from agentic AI?
Trade capture and post-trade processing came first with 50%. Risk limit monitoring and exception routing drew 25%, credit checks and counterparty approvals 17% and scheduling 8%.
Trade capture. Most of those who chose trade capture also named data quality as the biggest barrier in question 2. They want the agent at the point where the data problem starts. That is where a mis keyed volume, a wrong delivery point or a missing broker fee enters the book. Confirmation, reconciliation, position calculation and regulatory reporting all inherit it.
There is a dependency here that is easy to miss. An agent that reads a broker confirm and books the trade checks it against reference data: products, delivery periods, counterparties, venues, fee schedules. If that static data is wrong, the agent books the wrong trade with full confidence. The first use case and the first barrier are the same problem.
Risk limits and credit. Risk limit monitoring drew 25% and credit checks 17%. The wording of the option matters here: risk limit monitoring and exception routing. The agent watches utilisation and exposure and routes the exception. A person makes the call. Limit utilisation and credit exposure are continuous, rule bounded and have a clear breach definition, which makes them good ground for that split.
Q4: How confident are firms that AI outputs in risk and compliance can be trusted?
Three quarters (75%) are not yet confident, because governance frameworks are not mature enough. The remaining quarter split evenly across the other three answers: fully confident, confident for routine cases with human review for exceptions, and not confident with human oversight on every decision.
These answers sit alongside question 2, which asked for one barrier only. There, 50% chose data quality and governance came second with 25%. Yet even among those who named data quality in question 2, most said here that they are not yet confident because governance frameworks are not mature enough.
Data is the first problem most people name. Governance is the next one. A programme that fixes the data and leaves governance alone will still end with a person checking every output.
Expecting AI to reach scale soon makes little difference to trust. The 58% who said in question 1 that the industry is ready now or will deploy AI at scale within one to two years answered this question much like everyone else. Most are not yet confident in AI outputs in risk and compliance.
That has a cost. When every output is checked, review effort grows with volume. The work moves from doing the task to checking it, and much of the saving goes with it.
The firms that get real value will be the ones that can prove which outputs need review. That takes three things, and together they are what a mature governance framework for AI outputs is made of:
- Deterministic validation of every AI output against the source record, so a person only sees what fails.
- Confidence thresholds that route low confidence results to a person and let the rest through.
- Lineage from every output back to the trade, the feed and the rule that produced it.
Lineage is also what a regulator will ask for. If a surveillance alert was closed or a transaction report was corrected with AI involved, the firm has to show how that decision was reached.
Q5: Where will AI have the greatest impact on post-trade operations?
Trade surveillance and anomaly detection came first with 50%. Automating regulatory reporting workflows drew 33% and data reconciliation and break resolution 17%. Real time position limit monitoring and breach detection drew no votes.
Surveillance. Energy surveillance is a volume and false positive problem. Rules based scenarios for spoofing, layering, wash trades and marking the close generate alerts that analysts clear by hand. Cross market behaviour, such as a physical position used to move a financial one, or gas trading used to move power, is hard to write as a fixed rule. That is where pattern detection earns its place.
Most of those who chose surveillance named data quality as the biggest barrier in question 2. That is consistent. A surveillance model needs orders and trades from every venue, OTC broker data and physical positions, all on a common timestamp. Missing order data from one venue leaves the model blind to the behaviour it is meant to find.
Regulatory reporting. Every respondent who named governance as the biggest barrier in question 2 chose regulatory reporting here. Most of those who chose regulatory reporting named governance in question 2. People who worry most about governance want AI in the most governed process.
That makes sense. REMIT and EMIR reports have fixed schemas and published validation rules. An AI output can be checked field by field against a known standard, and it either passes or it fails. The governance question is easier to answer where the answer is binary.
Position limits. Nobody chose real time position limit monitoring, although 25% picked risk limit monitoring as the workflow that will benefit most from agentic AI (question 3). We read that as a view on where the real work sits. A position limit check is a calculation: aggregate positions by contract, compare against the limit, flag the breach. The hard part is the reference data underneath it. Spot month, accountability and delivery limits, step downs as expiry approaches and venue specific rules all change, and keeping them up to date takes constant effort. The calculation itself follows fixed rules and has to be exact every time, so there is little for AI to add.
Reconciliation. Only 17% chose it. We think that undervalues it. Matching confirmed trades against the book is largely solved with rules. The exceptions are not: breaks caused by timing differences, unit conversions, price formula mismatches and late amendments. That is where analysts spend their time, and where AI would help.
What should energy trading firms do before they scale AI?
The survey points to a clear order of work.
- Fix the data at capture first. Validate every trade against reference data at the point of entry, before any agent is asked to book one.
- Build lineage before you ask for trust. Every output needs a path back to its source trade, feed and rule. Without it, a person has to check everything, and scale stays out of reach.
- Start where outputs can be tested. Regulatory reporting and limit monitoring have hard pass or fail checks. They are the fastest route to evidence that a model works.
A firm that scales AI on unvalidated trade data will produce errors faster, and a person will still have to find every one.
This is the post trade foundation BroadPeak has built for energy and commodity trading firms. It covers the trade lifecycle from capture to regulatory reporting.
- Capture and validate. BroadPeak Trade Capture connects directly to exchanges and brokers and applies validation and correction rules before trades reach the E/CTRM. BroadPeak Data Integration normalises, enriches and validates data from E/CTRM systems, broker and exchange feeds and trade repositories.
- Trace every trade. Each trade and data payload carries a permanent, time stamped audit trail from ingestion to delivery.
- Control and report from one record. Trade Allocation, Trade Reconciliation, Position Limits, Trade Surveillance and Regulatory Reporting all work from the same validated trade data. Surveillance already uses a machine learning model to prioritise high risk alerts, and Regulatory Reporting returns real time confirmations and error alerts from the trade repository.
When a firm adds AI on top, the model starts from checked data, and every trade it touches can be traced back to its source.
Source
- BCG, How AI Is Transforming Energy Trading, July 2026
- Oliver Wyman, How to thrive in commodity trading’s AI future, April 2026
- Risk.net, The spread of agentic AI in energy trading, April 2026
- Ofgem, Ethical AI use in the energy sector (version 2), May 2026
- Bird & Bird, Are you ready for REMIT II? Essential updates for algorithmic traders, 2024
