AI in Commodity Risk Management – Getting the Data Right Comes First
by Gary Vasey, CTRM Center
There is no shortage of interest and discussion around artificial intelligence in the CTRM and risk management software markets at the moment. Almost every vendor now seems to have an AI story, and increasingly, end users appear under pressure from their boards and senior management to demonstrate that they are doing something with the technology. However, there is a considerable difference between experimenting with AI and actually deploying it in a way that delivers measurable business value. That was one of the topics I discussed recently with Ofir Gefen, CEO of CubeLogic.

Ofir told me that virtually every customer he had met during his first few months at CubeLogic had some form of AI initiative underway. In some cases, the mandate was coming directly from the boardroom. Yet enthusiasm for AI doesn’t necessarily translate into enthusiasm for allowing it to make decisions. Indeed, many appear to prefer a rather more cautious and pragmatic approach, he told me.
According to Ofir, most credit and risk managers are interested in AI as a means of improving productivity and decision support rather than replacing human judgment. “For example, a credit manager might welcome an AI assistant that reviews exposures overnight and presents the most pressing issues each morning. Similarly, an agent capable of preparing margin call communications, gathering supporting information, and identifying potential problems could save considerable time”, he said. But actually sending those margin calls or making credit decisions without human approval remains problemmatic for many in the industry,
Ofir cited industry survey findings suggesting that just 2% of respondents wanted complete AI automation. This reflects what we at ComTech have been hearing elsewhere in the industry. While there is considerable interest in using AI to eliminate repetitive tasks, improve access to information, and identify potential risks earlier, there is much less appetite for handing over responsibility for decisions to AI with human oversight and inolvement.
We also discussed the obstacles that are preventing firms from adopting greater automation. Ofir cited survey results from industry events in San Diego and Valencia, where 39% and 49% of respondents, respectively, identified rigid, hard-coded technology as the biggest barrier to automation.
We have spent many years discussing the challenges of integrating CTRM systems, consolidating data, and eliminating manual processes. Yet despite considerable investment in technology, many firms still operate environments characterized by disconnected applications, spreadsheets, custom interfaces, and fragmented data. AI does not magically make those problems disappear. In fact, it may actually expose them.
An AI system can only produce useful analysis if it has access to sufficiently complete, accurate, and timely information. If exposures are scattered across multiple applications, collateral records are incomplete, or counterparty information is inconsistent, the output may be unreliable regardless of how sophisticated the underlying model happens to be. Ofir made the point that “AI is generally more effective at extracting useful signals from well-organized information than it is at somehow resolving every inconsistency in disparate data sources.”
There is a tendency perhaps to view AI as a shortcut around some of the longstanding data management problems in commodity trading. But we suspect that firms taking this approach may eventually discover that they have simply added another layer of technology to an already complicated environment.
This thinking is also reflected in CubeLogic’s product roadmap. Ofir described the company’s V12 release as the foundation for a broader, multi-phase strategy built around improving data accessibility and processing performance before introducing more sophisticated AI capabilities. The architecture uses data lake technology and is designed to move away from traditional batch-oriented processing, toward much more timely risk calculations.
According to Ofir, “CubeLogic can now process around one billion exposure records in approximately eight seconds, compared with processing times previously measured in hours.” That is an eye opening improvement, although actual performance will likely depend on the nature of the calculations, data, and operating environment.
If a risk management platform can access and process large volumes of reliable data quickly, it becomes possible to support more responsive, event-driven risk management. That sought after agility in systems becomes possible. Rather than waiting for scheduled overnight calculations, firms could potentially recalculate exposures in response to new trades, significant price movements, changes in collateral, or other material events. This is something we have discussed at ComTech for some time.
In fact, a survey we conducted a few years ago suggested that many market participants were not necessarily looking for continuously updated risk calculations. Instead, they wanted calculations to be triggered by relevant events. The objective is not simply to calculate risk more frequently. It is to provide decision-makers with updated information when it is actually needed.
CubeLogic’s approach is to build AI capabilities progressively on top of this data lake foundation. Ofir described near-term plans for agents capable of ingesting and interpreting unstructured information, such as letters of credit and other collateral documentation, and making that information available within the platform. Anyone familiar with credit risk management in commodity trading will know how much important information can be contained in documents that are not readily accessible to conventional systems. Extracting relevant details from those documents, validating them, and making them available for analysis could deliver meaningful operational benefits without requiring firms to surrender control over critical decisions. Later phases of the roadmap envisage more sophisticated co-pilot capabilities and eventually event-driven agents able to respond to changes in underlying risk data.
Ofir also emphasized that CubeLogic is using privately hosted AI models within its own controlled environment rather than relying on public-facing models. Given the sensitivity of trading positions, counterparty exposures, and credit information, data security and containment will clearly be important considerations for prospective users.
Another topic that emerged was whether large commodity firms should develop their own AI solutions or rely on specialist technology providers. Ofir argued that the traditional economies-of-scale advantages of software vendors remain relevant in the AI era. If several major energy companies each develop their own risk management AI capabilities, they must individually bear the costs of development, maintenance, testing, and ongoing adaptation. A specialist vendor, on the other hand, can potentially spread those costs across its customer base and incorporate lessons learned from a variety of users in the industry..
Some of the larger trading organizations possess considerable internal technology expertise, and proprietary data and processes can provide real competitive differentiation. For these firms, developing certain AI capabilities internally may make good sense. The more likely outcome is a combination of approaches, with specialist vendors providing core capabilities and customers developing additional functionality around their particular requirements.
One final point worth making is that all of this is happening against a backdrop of growing geopolitical uncertainty and increasingly complex commodity markets. As Ofir observed, geopolitics has become a major driver of demand for better risk management technology. The events of 2022 demonstrated just how quickly energy market conditions can change and how rapidly liquidity and collateral requirements can escalate. More recent geopolitical developments have only reinforced the need for timely information and a clearer understanding of exposures.
AI may eventually play an important role in helping firms identify emerging risks, prioritize actions, and automate routine processes. Firms still need accurate data, appropriate controls, effective processes, and people who understand the risks they are managing. Indeed, the more automation that is introduced, the more important those foundations may become.
One thing is clear and that is that Ofir sees CubeLogic’s V12 as an important step forward in developing real solutions to some of these longstanding industry challenges. What makes V12 different is its underlying architecture. Rather than simply adding new functionality to an existing risk management platform, CubeLogic has adopted a data-centric approach built around data lake technology, designed to consolidate and process enormous volumes of exposure data far more rapidly than traditional batch-based systems. Ofir cited the ability to process a billion exposure records in around eight seconds, compared with hours previously. That is a potentially significant change, not just in performance but in what becomes possible in terms of consolidated counterparty views, timely risk analysis, and eventually event-driven risk management. It also provides the foundation for CubeLogic’s planned AI capabilities, rather than attempting to bolt AI onto fragmented data and legacy processes.