Ask two people in commodity trading whether AI is delivering and you may hear two different applications. One technology leader described taking measured P&L uplift from tools they had built themselves to their heads of desk running billion-dollar books. Whereas a consultant who works across several firms described a market where nearly everyone has tried prototyping and almost nobody has reached production.

We kept hearing both versions. Rather than assume one was right and one was wrong, we built a report based on anonymised conversations with leaders, Cititec’s own live market activity, and attributed public research. Pockets not Platforms is the result, and this is a preview of what it found.

Why both stories are true

In most of the conversations we had, capability sat in pockets: a capable individual builds something genuinely useful, it runs its course locally, and when that person moves on all progress goes with them.

Public research confirms what we’re witnessing. McKinsey’s March 2026 survey of more than 150 commodity traders found about 40% had entered implementation and 30% were piloting. Yet, more than half reported that AI had moved EBIT by less than 2% thus far. 

Louis Dreyfus Company has publicly described roughly 2,000 personal agents and 20 scaled AI-enabled products. All the while, its chief information officer has acknowledged that finding a product owner with the accountability to drive change is hard. Scale and unresolved ownership, it seems, can coexist.

Where the work is landing

We mapped five workflow areas. Two show AI running and being depended on. The first is coding and software delivery, where the most notable reported example was an ETRM-scale migration. Originally quoted at 12 to 24 months, a small team delivered the migration in approximately 12 weeks.

The second is the work around core trading systems, where agents and lightweight applications take on manual tasks while ETRM or CTRM remains the system of record.

Research and decision support, risk and compliance monitoring, and operations were still live experiments. While operations are a common starting point, implementation was a common falling off point. 

What separates firms that repeat results

Access to capable models no longer appears to be the main differentiator. Rather, the difference lies in the conditions around the model: whether the data is available, whether anyone in the business can shift working systems, whether a governed path into production exists, and whether the cost comparison has been assessed honestly. In several organisations, the data work became the AI programme itself.

Three further patterns recur in the report. 

1. How controls were built into delivery was a better indicator of whether the tool reached supported production, rather than the firm’s choice in “federated” or “centralised” rollout. 
2. Someone has to carry the prototype that last mile, and firms are still working out who that person is and how to hire them.
3. Some AI work may compound, but it comes down to whether a firm has agreed what a cargo, a counterparty, and a position are before the build.

The first two of these are our interpretation of the material rather than established fact. More details on these findings are inside the report.

The people question

Roles are being created and redefined faster than job descriptions and pay scales can follow. Our reading from live mandates is that the vulnerability is internal before it is external: someone who has spent eighteen months becoming capable with these tools may now be worth materially more than the role they were hired into. They’re being head hunted, and firms that can’t close the gap may lose their best talent. This report sets out questions for testing internal calibration, and Cititec’s current compensation data by role and location is available on request.he cutting-edge tech shaping the sector by subscribing to our newsletter, Commodities Tech Insider.