Would you trust an AI agent on your trading desk?
Would you trust an AI agent on your trading desk?
On a trading desk, nobody gets full system access on day one. Access is earned, limited to the job, and every action is logged. An AI agent shouldn't be any different.
That's the idea our team took to Copenhagen last month, where we spent two days at Microsoft's invite-only hackathon, working directly with Microsoft engineers. Trips like this are part of how we invest in AI at Previse: we send our developers to build alongside the best in the field, so the AI and MCP layer in Coral is built to a standard our customers can rely on.
Our question for the week: can we run AI agents inside Coral safely, without losing tenant isolation or cost control?
The short answer is yes. The longer answer is that the hard part isn't the AI.
Permissions, not prompts
Model quality wasn't the constraint. Identity, permissions, auditability and cost management were.
An AI agent that touches customer data has to meet the same standards as the people using Coral. In practice that means:
Scoped credentials, so an agent can only see and do what its user can
Tenant isolation, so each customer's data stays separate
A full audit trail of every action
A hard spend ceiling, so costs stay under control
Those are architecture problems, not prompt problems. They're also problems we know well, because trading software has always had to get permissions and audit right. It's the foundation we're building Coral's MCP layer and AI features on.
Side by side with Microsoft
We went with one clear goal: AI enablement in the Coral product itself. Focused building, daily check-ins and expert sessions with Microsoft gave us the time and access to make real progress on it.
Our Microsoft sponsor, Raphaël Bickel, helped us find our way through a fast-moving space and connected us with senior Microsoft engineers. Their advice on how to build AI into ETRM software properly was worth the trip on its own.
For our customers, that means the AI in Coral isn't being figured out alone. It's being shaped with input from the engineers behind some of the most widely used enterprise platforms in the world.
What aerospace taught us about trading
Some of the most useful ideas came from other industries. Teams in aerospace, healthcare, logistics and manufacturing were solving problems that look nothing like ours. Yet the way they handle specialist knowledge, human approval steps and hand-offs between agents maps directly onto trading. A four-eyes check is just a human approval step by another name.
You don't get that kind of insight by staying inside your own industry.
Putting it to work on a complex PPA
To test all this, we built a proof of concept where an agent takes a complex PPA from contract through modelling to deal capture inside Coral. The agent acts on behalf of a user, with exactly the information and configuration that user has access to. It ran.
What's next
We're now turning the proof of concept into something we can ship inside Coral, ideally together with one of our customers. If you're exploring how AI could fit into your own trading operations, feel free to reach out.
Thanks to the Microsoft engineers and fellow teams who pushed back on our assumptions, and to our team for building something that worked.