As banks and investment firms look to bring agentic AI into trading operations, the challenge is increasingly one of architecture: how to give AI access to the applications and context it needs while maintaining control over its actions.
HERE, the financial desktop technology provider formerly known as OpenFin, recently held sessions with tier-one banks and a major hedge fund to explore how AI should work on trading desks. According to CEO Mazy Dar, the discussions revealed differing approaches to deployment, from proprietary AI frameworks to enterprise platforms such as Microsoft Copilot, but also common concerns around security, governance and return on investment.
The Desktop Context Gap
Technologies such as the Model Context Protocol (MCP) enable AI systems to connect with enterprise data sources and external tools, but the information available through those connections does not necessarily reflect the user’s immediate working environment.“Context is maybe the biggest concept in AI,” Dar tells TradingTech Insight. “The more context you have, and the better context you have, the better the AI gets. One of the big themes we heard was that there’s all this context on the desktop, across all the different applications an end user is using, and for the most part AI has no access to that context. All the AI systems everyone’s setting up are connected into data sources and all kinds of things, but they just don’t know what’s happening where the end user is actually doing their work.”
This need for richer context is also shaping how established market data providers are developing their AI offerings. Bloomberg, for example, has expanded its ASKB assistant to incorporate client conversations, worksheets and files. But with traders working across multiple vendor and internally developed applications, providing AI with a comprehensive view of their activities remains a broader integration challenge.
AI’s potential contribution extends beyond information retrieval to supporting operational tasks. Dar gives the example of allocating a newly booked trade across multiple funds. An AI system could propose allocations and populate the relevant fields within an existing application, leaving the trader to review, adjust and submit them. This could reduce manual processing while preserving human oversight.
“We’ve got all these core applications that our business runs on – trading systems, order management systems, risk systems – and there’s one camp of people who think all those systems are going to just go away, that it’s all going to be replaced by an AI or a chat interface,” he says. “But I think there are more people who are realistic about the fact that wiping away all those systems is way more complex than it might appear. Certainly in the near term, the more likely and the better outcome for end users is if those same applications could be in the AI workflow.”
FDC3 and Controlled Agentic Workflows
FDC3, the financial desktop interoperability standard originally developed through OpenFin and subsequently established under FINOS, offers one potential foundation for connecting AI with established trading systems.
Dar sees particular relevance in the standard’s existing mechanisms for application interaction as institutions explore agentic workflows.
“One of the things FDC3 defines is the idea of an intent – an application sending a workflow instruction to another application with context. The really interesting thing now is that AI is, in a way, unlocking the power of FDC3 and enabling it to be used in new ways. If you take the example of allocating a trade, there’s an application that knows how to do that within your organisation. That application needs to say, first, I’m okay being invoked in this way, and here’s the interface. Second, you can put additional guardrails around that to say whether AI can invoke me or not. That’s very different from having some AI just clicking on any button that’s available to it in the application. That surface area is huge, versus a very specific workflow instruction that I’m happy to have AI invoke me with.”
Implementation requirements differ depending on the level of functionality involved. According to Dar, firms can connect their preferred enterprise AI system to HERE’s workspace and provide it with desktop context without additional engineering. Enabling AI to invoke application workflows requires the relevant application owners to implement FDC3 intents, although existing implementations can provide a starting point.
Institutional approaches also vary. Larger banks and asset managers with substantial engineering resources are developing proprietary orchestration frameworks, sometimes incorporating FDC3, while smaller firms can adopt turnkey workspace platforms with their preferred AI systems.
Governance and the Limits of Autonomy
Security and governance remain significant barriers to adoption, particularly as institutions determine where AI can operate independently and where existing approval processes must be maintained.
“There are a whole host of activities that AI can be trusted to go and do on its own, where it can be fully agentic and you don’t need the human in the loop,” observes Dar. “Doing a trade is an example of something you wouldn’t be comfortable today with AI going and doing on its own. Then there are all the things where you really need a human in the loop. Those humans are sitting in front of a desk, and they have an existing workflow that uses all these different applications. For all those activities where you want the human in the loop, you’d still like the AI to be helpful to that human, to make it easier for them to take the actions they’re trying to take and synthesise the information they’re seeing on screen faster.”
Recent scrutiny of trader-developed applications, including reported restrictions at Goldman Sachs, highlights the difficulties institutions face in balancing experimentation with centralised technology controls. Dar identifies security and governance as the principal barriers to broader deployment, rather than a lack of demand for AI capabilities.
Current best practice in securing AI-enabled workflows encompasses granular access controls, data loss prevention, restrictions on information exposed to AI, detection of unauthorised AI applications and measures to prevent the transfer of personally identifiable information. These protections form part of the broader governance infrastructure required to support AI-enabled workflows, complementing the application interoperability mechanisms provided by FDC3.
For financial institutions, the emerging challenge is therefore both architectural and operational. AI capabilities are appearing across market data platforms, productivity applications and proprietary trading systems, while banks continue to develop their own frameworks for connecting and governing them. The extent to which these different environments can share context and support controlled actions will help determine how far agentic AI moves beyond standalone assistance into the day-to-day operation of trading desks.
Subscribe to our newsletter
