The full hour recording “From Concept to Competitive Edge: AI in 2026” panel from CTW Europe, is available above. The exchanges between panellists, the follow-up questions, and the tone of the room are worth the full listen. Or read on below, where we’ve pulled out the parts of the conversation with the most actionable and thought-provoking insights.
Commodity Trading Week Europe put five voices on stage to talk about AI in 2026, and the range of experiences is what made the session worth an hour of anyone’s day. A JLR technology leader walked through raw material forecasts that jumped from roughly 50% confidence to above 90% once supply chain data was properly connected. A JDE Peet’s leader told a candid story about a frost prediction project the team gave up on too early, and regretted it six months later. A platform founder talked through how fast agentic AI has moved from document extraction to end-to-end workflows in just the last year. Each of those perspectives brought something different to the table, and hearing them build on each other in real time is exactly why the recording is worth listening to in full.
Within that mix, Gen10 CEO Richard Williamson’s answers gave us a few specific, practical ideas for implementing a practical AI strategy. Here’s what stood out for us.
A simple test for whether AI is the right tool
Asked how to approach an AI strategy, Richard didn’t start with tools or vendors. He started with a distinction: some problems are better-solved by deterministic computing, and some genuinely benefit from a probabilistic model. Confusing the two can be a very expensive mistake.
His example: don’t ask AI to revalue your trading position every time someone wants an answer. It can technically do it. It will also cost you every time you ask, for a task a good Commodity Management System should already be doing reliably, in real-time, and for free. This is the kind of test any commodity trading firm should apply to every AI idea before spending a token on them.
The data lake trap
When the panel discussed AI projects that fail quietly, Richard raised a common scenario. A firm wants to connect its data with minimal changes to its operational systems, so it builds a data lake, then a data warehouse on top, and eventually gets to a point where it can ask that warehouse all sorts of questions. It is good progress in the moment, but the problem comes later, when the firm tries to move from asking questions to running agentic workflows, and discovers it can’t. That kind of project hasn’t failed, but its original goal was both limited and limiting.
Agentic AI is the shift from a system that tells you something to one that goes and does something about it, without a person triggering each step. Agentic AI only works if the AI has a route back into the systems where the actions happen. A data warehouse can be a superb source of answers, but if your AI isn’t connected to your operational systems, there is nowhere for an agent to act.
Corporate memory won’t look after itself
The panel’s most human moment came when they discussed what skills people will need as agents take on more of the day-to-day work. Richard pointed out that this is also linked to one of today’s major challenges: commodity trading firms lose institutional knowledge every time someone changes jobs. If agents now do the tasks that used to teach new starters how the business actually works, how will we ever rebuild that knowledge?
His answer was to treat corporate memory as something to be deliberately captured in a knowledge base that understands your workflows and systems, rather than something that simply lives in people’s heads. It’s a practical answer to a problem most AI strategies skip past entirely.
Introducing Gen10’s agentic AI
Richard also used the session to introduce NaNi, the agentic AI platform Gen10 has launched to incorporate AI directly into CommOS or other CTRMs. NaNi is connected to that single source of truth, able to act directly on operational data and workflows, and with the governance, audit trails, guardrails, and cost management that keep an AI strategy accountable.
Where to start
If you’re looking to get more serious about agentic AI, the points the panel made form a handy checklist for getting started:
- Listen to the full panel discussion: the recording above covers the anecdotes, follow-up questions, and back-and-forth between all five panellists that a written summary can’t fully capture.
- Run the deterministic-vs-probabilistic test: before committing budget to an AI idea, check whether it’s solving a problem that a good CTRM should already handle reliably and for free.
- Trace one workflow end to end: pick a single process and follow it from your data platform through to the operational system that would need to act on it. Where would AI add genuine value? And are there any data traps that would stop AI taking action?
- Start capturing corporate memory deliberately: identify the knowledge that currently only lives with your most experienced people, and build a plan to get it into a system.
- Check your governance layer: make sure whatever you’re building has the audit trails, guardrails, and cost visibility your other systems all live by.
None of this needs to happen at once, and it doesn’t need a finished strategy to get going, it all starts with an honest look at your own data, and a conversation.
