Unlocking Wall Street’s Dark Data: How AI Agents are Transforming Trading Floor Chat

For decades, some of the most valuable information in financial markets has been hiding in plain sight. Client intent, actionable orders, and vital market colour have been locked within the unstructured, transient streams of human-to-human chat. On trading floors worldwide – particularly in over-the-counter (OTC) markets – this conversational data represents a multi-trillion-dollar blind spot: a continuous flow of ‘digital exhaust’, rich in potential but historically impossible to capture, structure, or leverage at scale. This is the front office’s dark data problem.

But how are leading institutions navigating this shift in practice? What strategic drivers are compelling them to act? What hurdles must be overcome? And how is the vision of an AI-augmented trading desk starting to materialise?

The Strategic Imperative: Why Now?

The sudden acceleration in this space reflects two powerful forces reshaping the industry.

First, market structure. To avoid the information leakage and market impact associated with transparent central limit order books (CLOBs), institutions rely on private, bilateral chat for large, complex trades. This need for discretion has long defined OTC risk transfer and is now driving the increasing use of non-lit liquidity in equities, as firms across all markets seek to protect their execution intent.

Additionally, the ‘financialisation’ of niche commodity markets has attracted quantitative and systematic funds used to the data-rich ecosystems of equities and bonds. These players have an insatiable appetite for structured, historical data to support their models. Yet much of this data has historically lived only in brokers’ conversations, leaving vast market segments under-analysed.

Chris Hudson, COO of Freight Investor Services (FIS), confirms the trend: “We’ve been having more and more conversations about data collection and the increasing financialisation of some of our key markets,” he explains. “Freight is a great example. When a market starts, you typically have participants with physical exposure and operations who use those markets to hedge risk or trade for their own exposure. As volume increases, it attracts financial players. Iron ore is a good example; it’s now a very large market where the volume of derivatives and futures traded is eight or nine times the physical volume. In the dry bulk freight market, derivatives are now over double the size of the physical market”

He continues: “That type of participation brings an increasing demand for data and an understanding of how to use it. This trend led us to release our own platform in 2019, which enabled customers to view indicative pricing across our markets. That platform has since gone through multiple iterations, and we’re now at the stage of having conversations about NLP models and the use of AI agents, both internally and externally.”

The other factor is the increasing openness of chat platforms. The longstanding open API philosophy of platform providers such as Symphony has now been joined by dominant incumbents such as Bloomberg opening up their systems. This unlocks conversational data streams, turning previously closed systems into fertile ground for automation and analytics.

Michael Lynch, Symphony’s COO, highlights the company’s foundational role: “We’ve always had an open API. We have thousands of bots on the platform today being used for everything from workflows on the sales and trading desk to operations and risk. Because of our end-to-end encrypted nature, we don’t necessarily know what they all do; but we do know that our customers make heavy use of them.”

Colin, Duggin Global Head of Equities Product at TP ICAP, sees this openness as a key catalyst: “Traditionally, the challenge with conducting business over messaging and voice channels has been capturing and utilising the data effectively,” he notes. “Until recently, much of this information was locked within closed systems. Now, with platforms increasingly supporting API integration, we’re seeing a surge in activity. The opportunity lies in using this previously inaccessible data to enhance client service. We can now extract, structure, and analyse it – and we’re only beginning to uncover its full potential.”

From Information to Automation to Control

Capturing the opportunity is not as simple as plugging AI into chat streams. The firms leading this transition are adopting a staged, disciplined approach best understood through a framework of Information, Automation, and Control.

Information is the foundation: capturing raw chat and transforming it into structured, machine-readable data. Matthew Cheung, CEO of ipushpull, explains: “In its simplest form, from a data flow perspective, you have either inbound or outbound flows. Typically, in trading and broking, the outbound flow consists of sending out prices, axes, and runs. In OTC markets, particularly on a bank’s fixed income desk, you’ll see people pasting a whole bunch of axes and runs into a chat at eight o’clock in the morning. This data might also be distributed to venues like Tradeweb or Neptune, but for some products, like many swaps axes, there is no predefined data structure, so they don’t go anywhere apart from chat.”

He continues: “The other side is inbound, where a customer or counterparty is grabbing something from the chat. If I’m sending out my axes, I’m hoping to get an RFQ back based on that information. This leads to a multi-step, negotiation-type process: a client might bid 8, I’m offering 10, we go back and forth and agree on 9, and then we say ‘Done’ in the chat. At that point, that trade data needs to get into two or three other systems for processing.”

Automation is the next layer. Once conversational data is structured, firms can automate the low-value but time-intensive tasks that burden traders. For IT leaders like Timothy Flynn at Koch Ag & Energy Services, the primary inefficiencies in these manual workflows have always been clear: “The biggest inefficiencies were timeliness and accuracy. Traders previously hand-entered deals into our system of record at the end of the day. This increased the chances of making an error or worse, missing a trade, with the associated costs and downstream impacts. From an IT infrastructure perspective, there weren’t any real inefficiencies as deals were entered manually, but the process was cumbersome, and traders might miss opportunities while they were entering deals and reconciling any errors.”

Julien Dugat, Head of Fixed Income Product Management at NatWest Markets, adds: “The goal is to augment the entire client sales workflow. A lot of our business, especially in rates, takes place in chat, with clients asking for prices all day on everything from a simple 10-year gilt to more complex structures. Many of these tasks are not super high-value-add; reading that a client wants a price for a vanilla product, looking at a screen, and copy-pasting a response back is a workflow that can be relatively easily automated.”

The aim is to reduce operational risk and free up skilled professionals, says Hudson. “My focus has been on what AI can and can’t do. And much of what it can do well is the mundane stuff that we don’t like doing. That’s where we’ve concentrated on practical implementation that delivers something useful.”

Control is the essential safeguard. In a regulated, high-stakes environment, AI cannot operate unchecked. “The system interprets the transactions, inferring the fields needed to book that transaction in the system of record,” explains Flynn. “Our traders copy the trade into the chatbot and, when they’re ready, review the inference, make adjustments, and submit. For higher-risk trades, those are identified, and the trader is alerted to double-check the deal’s accuracy.”

Lynch underscores Symphony’s approach: “We already have the necessary data loss prevention, compliance controls, and guardrails for secure conversations. Those same controls can be applied to our API-driven workflows when executed by an agent. For example, our data loss controls can block a user from sending a file marked ‘confidential internal use only’ to an external party. We can do the same for an agent.”

He adds: “The way we’ve built our ecosystem is with trusted entities. Our identity platform is key. We don’t have any public participants; everyone on the platform is vouched for by their enterprise. This guarantee of genuine identity is something we want to see translate into the agentic world. With our platform, you have a higher level of trust that what you’re communicating with is genuinely representative of the firm that published it.”

In Practice: Augmenting, Not Replacing

A consistent theme across the industry is that this technology is about augmentation, not replacement. The co-pilot metaphor is widely used: AI enhances human capabilities but doesn’t supplant them. The downstream impact of structuring trade data instantly is transformative, as Darnell Bortz, director of natural gas trading at Koch Energy Services, explains: “Our traders save time, reduce errors, and are now naturally adapted to the platform’s efficiency. When the trader double-checks the inference in the trade listing blotter, the details are viewed next to the conversation they copied down. This significantly reduces errors and provides a more accurate position and P&L. Additionally, because we store the chat and booked deal together, we can easily pull the conversation between counterparties in the event any questions arise. Finally, the traders learn what a good message looks like, vastly increasing the accuracy of the inference, improving the communication with the counterparty and minimizing discrepancies.”

At TP ICAP, success has hinged on persuading a sceptical broker workforce. “I envision our brokers having more capacity to focus on client-facing, high-value activities,” says Duggin. “A key activity is recognising the human tasks that the new technology is ideally suited to – the time spent rekeying and copy/pasting, not meaningful work but necessary wherever the correct tooling is not available. New methodology can form part of the solution to unlock this sunk time. The goal is to create the right tools so they can engage more clients and spend more time uncovering insights into market movements.”

Hudson echoes this experience. FIS has used agentic technology to expand its business without a proportional increase in headcount, creating leverage without diminishing expertise. Yet, the human broker remains indispensable. “We’re at a stage where there’s such a disconnect between the average user and an understanding of the technology that’s been developed,” he notes. “The logical result of that is reticence when you try to implement these systems. So we spent a lot of time internally thinking about the need to show the outputs and understand them in the context of what we’re trying to do, while absolutely keeping a human within that element.”

He adds: “Within data collection, we have multiple validation stages, which are tech-enabled to an extent, but ultimately it is a human who makes the final ‘yes’ or ‘no’ decision. In volatile markets, our internal system might flag a potential erroneous price, for example. That query goes to people internally who are aware of the markets, and they can confirm if it’s correct. So it still keeps a human in the system.”

NatWest Markets is taking a broader approach, building a foundational automation toolkit for the entire enterprise rather than solving single workflow issues. This raises a nuanced challenge: creating higher-level agents intelligent enough to know when not to intervene. “The first challenge for the framework is to determine, ‘Is my input required here, and what value can I add to this conversation?’” notes Dugat.

This ambitious goal, however, immediately confronts the practical limitations of the underlying technology. Vendors and internal teams alike note that LLMs, trained primarily on textual corpora. The answer lies in orchestration frameworks and a relentless focus on accuracy. As Dugat points out, “When you’re taking a piece of text and turning it into a structured inquiry, it has to be correct so that we price the right thing. We need something like 98–99% accuracy for this.”

The Agentic Leap: From Bots to Agents

As adoption grows, industry consensus is forming around a key distinction between bots and agents – a spectrum of autonomy that reflects how much independence a system has in interpreting context, making decisions, and taking action.

Traditional bots are deterministic tools: they follow explicit instructions. A user issues a command in a predefined syntax – often little more than a shortcut into a larger system – and the bot executes a single programmed task. Examples include a chat command that pulls a price from a market data feed, or a bot that posts trade confirmations to a channel at set intervals. These bots are functional but rigid: they require precision, cannot infer intent, and can’t adapt to new situations without code changes.

Agentic workflows occupy the middle ground and are where most of today’s innovation lies. These systems understand natural language, interpret intent, and choose the right tool from a toolkit of functions. An agent might recognise that “Show me yesterday’s cleared swap trades above €50 million” is a request to query multiple datasets, apply filters, and return a structured view – all without the user knowing the underlying syntax.

Cheung explains how this unlocks sophisticated applications: “One powerful use case is the creation of a shared blotter. On a trading desk consuming axes from dozens of different banks, that information typically exists only within disparate chat windows. An agentic workflow can now monitor all of those rooms, grab the unstructured data, and consolidate it into a single, structured grid. Once digitised, this data can feed other downstream systems, transforming isolated conversations into a unified market view.”

He provides another example: “Accessing a sophisticated risk margining tool at a bank for example, would normally require a six-to-twelve-month installation and approval process for a new desktop application. Now, an access-controlled agent can act as a secure gateway. A trader can drop a portfolio of positions into a chat; the agent structures the data, passes it to the firm’s risk engine, and returns the calculated output directly into the conversation. This model dramatically accelerates the deployment of new tools by leveraging the existing, approved interface of the chat client which is ubiquitous on traders screens.”

Beyond reactive tasks, agents are enabling proactive intelligence. In a large bank with siloed desks, flow coming into the equity derivatives desk (for example) might be valuable to the prime broking team, but such connections are rarely made. Cheung elaborates: “An agent with a ‘hivemind’ view can scan every relevant chat across the organisation in real time. When a client inquiry comes into one desk, the agent can instantly identify a potential matching interest on another, uncovering opportunities that would have been impossible for a human to find. The same capability can be used for real-time sentiment analysis: ten minutes before a major economic data release, a trader could ask, ‘What’s the sentiment on the floor?’ and receive a summary of rumours, client views, and research opinions from across the network.”

Agents are also being deployed for cross-lifecycle orchestration, automating routine but critical processes like pre-trade compliance and sanctions checks, performing them as soon as a conversation about a specific instrument begins. For institutions with fragmented, legacy tech stacks, chat agents act as a unifying layer, pulling data from disparate sales, trading, risk, and back-office systems and presenting it in one interface.

Agentic workflows such as these operate within strict guardrails. Permissions, entitlements, and workflows are enforced externally, ensuring they cannot exceed what the human user is authorised to do. Human approval is still required for critical actions like trade booking and order submission.

At the far end of the spectrum sit fully autonomous agents – systems capable of acting without human prompting, making decisions dynamically, and even adapting workflows. In theory, such agents could negotiate, execute, and manage trades end-to-end. In practice, financial markets are still far from adopting this model; operational, legal, and reputational risks remain high, and regulators are still in the early ‘horizon scanning’ phase of exploring oversight frameworks.

The Future: Towards a Networked Conversation

For now, most efforts focus on optimising internal workflows, but platform providers see a longer-term evolution: a secure, agent-to-agent marketplace.

“You will increasingly see agent-to-agent interaction,” predicts Lynch. “We have bot-to-bot conversations happening on Symphony today, for sure. The world of agent-to-agent will happen, likely starting in areas where it’s relatively low-risk and in markets that are highly deterministic. The agents will have flexibility, but guardrails will ensure they only have flexibility within lanes that are well-defined by both providers and technology.”

This evolution promises to unlock even greater value, says Duggin. “Our long-term ambition is to understand how one market influences another – how an event in one area triggers movement elsewhere. That’s where we can add value to our clients. Access to large volumes of pre-trade and trade data across asset classes must be a quant’s dream.”

For human professionals, the shift points to a future where their expertise is applied at a higher, more strategic level. Insiders believe the most transformative use cases will emerge not from top-down mandates but from the desks themselves. “The real breakthrough will come when trading desks take ownership of data and begin using it in ways we haven’t even imagined,” Duggin adds. “That’s where innovation will flourish – not from technologists prescribing use cases, but from practitioners experimenting and discovering new value.”

The message across the industry is clear: trading firms are sitting on a vast, untapped data asset hidden in their chat streams. With platforms now open and AI tools maturing, the opportunity to unlock this value has arrived.

The question is no longer if firms should act, but how. And the competitive risk is stark: those who fail to embrace agentic workflows will face increasing inefficiencies and data blind spots, while peers move ahead.

As Matthew Cheung of ipushpull concludes, the fear that AI will eliminate jobs is misplaced. “Chat is not going away. It’s on everyone’s desktop, and you can use it for more than just chatting; you can use it for true workflow automation, data connectivity, and digitising workflows. You need to be investing in this now, because otherwise, you will get left behind. The efficiency and savings it will unlock make it a no-brainer. People need to get past the idea that it’s going to put them out of a job. The reality is, if you’re not using it, that’s when you’re at risk of being out of a job.”