Chief Technology Officer at Bloomberg By his own admission, a job that is half about building the future and half about preventing the future’s most embarrassing impulses from approaching production. “Half my job is to keep the crap out of the company,” Sean Edwards tells the Observer, and he’s not exactly kidding. The nonsense he refers to is the tidal wave of generative AI hype that has swept through every conference room with a Bloomberg terminal. His mission is to find the narrow cross-section between what its engineers can only dream of and what the people who are actually trading bonds, checking credit, and preparing for earnings calls are desperately trying to achieve — not what they are doing today, but what they are actually trying to achieve. In Edwards’ view, this distinction is the organizing principle behind ASKB, the interactive and conversational AI system that Bloomberg built directly into the station.
Ask Edwards about the use case, and he describes the old way of doing things. “Before ASKB, a user had to go to different places in the Bloomberg terminal to look at company fundamentals, to look at Street estimates, to look at different performance metrics, company KPIs, a different place to look at literature, peer analysis and alternative data for the performance of the quarter — reading a lot of news about the company and reading lots and lots of company documents and sales-related research.” Edwards paused. “and then They will have to compile all this information.
As of FebruaryBloomberg users can request ASKB to collect information. “It knows where to bring it and it knows where to find all this information and gives you a quite detailed analysis so you’re prepared.”
The machine does not replace the analyst, and Edwards was careful, almost insistent, on this point. “ASKB doesn’t do the full job of an analyst, but it does 80 percent of the work of gathering and synthesizing this information. It also frees them up to do value-added thinking—they really decide where to really dig.” Workflow varies by office. For stock analysts, it’s about setting up events and monitoring theses. For credit analysts, it is liquidity analysis and bond screening. The common thread is the research problem.
Trust is not an advantage
If there’s one obsession that runs through Edwards’ account in the past several years, it’s the reliability of an engineering system based on technology that was never designed primarily for finance purposes.
“A big focus, and essentially a focus on my team over the last couple of years, along with our core engineering team, has been how do we build AI that is worthy of our customers’ trust to make mission-critical decisions,” he says. One of the principles followed in developing ASKB is to refuse to let the model speak for itself. “We never want her to generate an answer from her global knowledge,” Edwards explains. Instead, ASKB is guided by decades of proprietary Bloomberg data, risk analytics, and pricing generators — what Edwards calls “sources of truth.”
Getting there means building validation tools into every step of the process — some check facts in real time (“You didn’t make up a fact, I can verify that… You summarized, and I can look at all the facts and compare the two”), others that pick up more subtle failures, like reading inverted sentiment. On top of that is a framework for continuous evaluation, “independent and manual evaluations to verify that we are indeed continuing, that nothing has changed, and that nothing has gone wrong in our system.”
The last layer is transparency. ASKB directs users to the source material — the paragraphs, out of millions of documents, that yielded the insight. Tells users which query did this. She shares the analytical call she made. Edwards was frank about how unappreciated this is from the outside. “They’re underestimating these different layers that have to work in tandem to get something trustworthy,” says Edwards, among a handful of clients who have tried to replicate what Bloomberg is doing with off-the-shelf AI models. “It’s not easy. In fact, it’s very difficult to direct an AI to do this. It wants to be very helpful, and sometimes it’s not helpful.”
Part of the difficulty lies at the bottom of the model, in the unattractive work of getting data sources to talk to each other. “How do you connect data? How do you rationalize all these different sources of information and rationalize a data model that you can join across disparate data sources?” Most importantly, Edwards argues, the people directing this work can’t just be AI engineers. “Increasingly, domain experts are building our systems. They are the ones who help guide the AI, and the ones who say, ‘No, you’ve got this wrong.'”
Even a CTO has a learning curve
For all the engineering rigor, Edwards acknowledges something surprisingly human. Using the tools is more difficult than anyone initially expected, including himself. “There were a lot of expectations at first, as it became very natural to use and have a conversation,” he says. “But I think we’ve all learned, actually, that there’s a learning curve. You have to put in the energy and the effort to learn how to use these tools well, and the more you put in, the more you get out. Maybe that was a little surprising from the beginning.”
Bloomberg is moving toward more customization to smooth out this curve, allowing users to tell the system their preferences, their coverage universe, and their habits, so that “over time the system gets smarter about how the system reacts to your questions.” But Edwards is frank that these are first innings, and that it raises a really unresolved question in the AI industry about how much memory is too much. “There’s a lot of research and a lot of tools coming out around memory — how do you compress memory, how do you use memory, how much prior interaction do you want to use? Right now, we’re very strict about no Retaining certain information, so it’s kind of a balance.
Bloomberg’s current CEO, Vlad Klyachko, started the job the same year as Edwards, and the two rose through engineering together — a detail Edwards offers as evidence of a company that “you either get really into it and stay in it for a long time, or you leave very quickly.”
This culture goes back to Michael Bloomberg’s deliberately open and flat work environment, where “anyone can come up with a big idea.” Edwards describes his contribution as protecting and extending this instinct, engineering the conditions necessary for cross-functional collisions to occur between a technologist, someone with a technical degree, a former portfolio manager, and an operations person, all on the same whiteboard. Creating ASKB would require tearing up this organizational chart, at least temporarily. Traditional Bloomberg products were created by separate teams with separate functions, with UX partners designing individual screens. An AI system that accesses every domain in the terminal does not fit this model. “The surface area of ASKB and how it works is different from how we build other systems,” Edwards says. “The way you work together, how you build this, is completely different… Applying our old structures to this new way of building the product didn’t work. We had to change our thinking.”
Employing the ability to explain
When asked what he looks for in the people he hires, Edwards didn’t mention their qualifications. It enlists communication, that is, the ability to take a complex idea and explain it at multiple levels, as a physicist would do to a child, college student, or doctoral student. This requires the ability to listen and be flexible with your own ideas, especially within cross-functional teams where the value is often not the new idea itself but the discipline it imposes. “There is some research that suggests that cross-functional teams work better with diverse backgrounds, not necessarily because there are just new ideas, but because it makes each player work harder to express their ideas, and therefore think about the problem better.”
As for why talent would want to work at Bloomberg in the first place, Edwards points to the sheer scale of data — weather data that fuels commodity models, document analytics that support research, and real-time streaming pricing — because “finance is the world” and “many, many different aspects of the world influence finance.” Edwards also points to the speed of Bloomberg’s influence. “You can create a feature, product, or capability and actually go see the customers using it, talk to them, and get feedback from them. That’s exciting.”
When pressed on what has shaped his thinking, Edwards cites the history of Bell Labs and its cross-functional glory days—unsurprising, given his obsession with crash-prone teams. Less expected was his other choice: that of Hermann Hesse Steppenwolfa novel that taught him that people limit themselves to a limited vision of who they are, when in reality “we can take on different forms and use different parts of our personalities and minds to grow, and we can pull off different abilities at once.” His career developed almost exclusively during periods of self-imposed discomfort. He calls them “uncomfortable challenges.”
What comes next for AI at Bloomberg? Edwards responds with something close to wonder. “We’re still scratching the surface of what generative AI and our approach will do,” he says. “There is so much more in our vision of what we can achieve. This technology has allowed us to dream bigger and address problems we dreamed of but couldn’t build. Now we are able to build it.”
