The Cutting Edge - episode 001 - featuring Gabriel Stengel, CEO and Co-Founder of Rogo

Hiring top AI talent is the first bottleneck for financial firms implementing generative AI; even mega-cap tech firms like Google struggle to hire leading-edge AI engineers who are competing with OpenAI and Anthropic for top talent at higher compensation, and most large banks and private equity firms cannot afford to compete at that salary level.

“The second thing we're we're actually seeing is that it's uniquely hard to hire technical talent at at these firms to build out these tools and you know it's even hard for Google to hire leading edge AI engineers and they're competing with Open AI and Anthropic paying crazy amounts of money and it's really really hard for large banks and private equity firms and then you know easier for folks like Citadel Jane Street and so on who can actually afford top dollar technical talent but for a lot of organizations step one is even just figuring out how to sort of staff the ranks for folks who can start to build with these sorts of technologies.”

The critical bottleneck for AI integration at financial firms is data infrastructure: integrating historic data, ensuring proper metadata tagging, connecting disparate sources (broker research, internal notes, expert interviews, Bloomberg data), and organizing unstructured content is more important than model capability and cannot yet be automated by AI.

“The second focus is the internal data integration and regardless of you know what pilots are running and what use cases are are working folks are getting very attuned to the idea that they need to shore up how they've integrated data historically to make sure it's leverageable in a world where you might have you know a super smart PhD level intelligence in your pocket but what it's missing is the ability to surf through your files look through your you know historic notes look through all the data you already have access to.”

Baselining is an effective AI-enhanced process where an analyst uploads multiple earnings call transcripts to identify management tone shifts and language deviations that signal future risks, using historical baselines from 6-12 prior transcripts as training data for recognizing when management communication patterns change.

“Baselining is a process that when constructing a thesis, right? We want to go through and we want to try and identify the the communication baseline from the company, right? What has been the language and tone for management discussed issues, how have they framed those issues? Fortune 500 CEOs can certainly be spin machines more often than not are spin machines. And so there's a it takes a finely trained ear and understanding of the nuance of communication to try to understand these. But in terms of execution, we generally recommend going back when you're initiating on a stock, go back to 6 to 12 transcripts earnings calls investor presentation.”

All financial services AI tools should have real, clickable citations that link to underlying documents, webpages, or presentation slides so users can audit answers immediately; time-to-auditability is as important as raw accuracy, and generic consumer tools like ChatGPT lack this product richness for financial use cases.

“That's why and I think you'll see this emerging especially in in, you know, tools for financial services, always having a citation. Uh and a real citation. And one that you can click on and you can go and audit. And you'll have seen in Rogo that, you know, the answers aren't always right, but if you click that citation, we are going to take you to an underlying document or a webpage or a presentation side by side so that you can audit it as quickly as possible. And a core product constraint for tools that you're going to be using aren't just the actual accuracy, but the time to auditability, right? Like if you produced work for a PM, but you had no underlying data, no model, if you wanted to jump into it, there was nothing there, how useful is it going to be?”

Large language models perform at 'intern quality' (80% accuracy) on analytical tasks today, and reaching 'analyst quality' (100% accuracy/reliability) does not require dramatically smarter foundational models but rather better data integration, domain-specific tooling, and integration of classic ML/statistical techniques.

“Right now maybe we're it's intern quality where 80% of it is right, but 20% of the time it doesn't work. And so, if it doesn't work or it's wrong 20% of the time, you know, do you want to use it? What are the best tasks and so on? And for me, analyst quality just means it's it's always right. The same way that if you're a PM and there's an analyst you trust and you would never question what they give you, that's how you feel about these tools for those sorts of tasks. And it's it's for it depends on exactly what we're talking about, but it's not that contingent on these sort of frontier models getting that much smarter. There's so much to be gained just in the plumbing, right? Connecting them to the right data sources, make sure you're just starting to measure the accuracy, making sure you can improve those with more classic AI/ML techniques, and it's just all the domain-specific use cases.”

Most public marketing and announcements about AI at financial firms are marketing buzz rather than indicative of real value-generating use cases; firms are incentivized to overstate AI adoption (to attract engineering talent and impress investors) regardless of whether it genuinely drives alpha or bottom-line impact.

“I think the reality with a lot of these announcements is that there's more marketing buzz than value at the moment and everyone and every firm is incentivized to talk about that what they're doing where they're going both for the sort of overarching big projects right if you have a a huge tool that you want to say that everyone's already using so that you can help hire more technical talent to work on that tool you're incentivized to do that all the way down to hey telling your investors how this is going to actually impact your bottom line or impact the way you're you're able to generate alpha.”

More money is spent on failed technology projects than successful ones historically; this suggests that optimization around a specific tool or framework carries significant risk, making data infrastructure preparation (LLM solubility) more important than building around current tools.

“Um but it's I guess it's a question, but it might sound like a statement. It sounds like the data structuring, metadata, tagging the data may be the most important thing to do to future-proof you internally. So that you're not you're not optimizing to a tool and next thing you know, it's not available or it's fallen behind. Yeah, it's almost like how do you make your data LLM soluble?”

Current generative AI is better characterized as 'augmented analysts' (not 'supercharged analysts'), 'accelerated due diligence' (not 'automated due diligence'), and 'advanced pattern recognition' (not 'real-time market intelligence'), and there is currently no AI that functions as a true thought partner or autonomous decision-maker.

“People want supercharged analysts, what what really have is augmented analysts. People want automated due diligence, what we have is accelerated due diligence. People want real-time market intelligence that's going to tell you exactly what to do and when to do it. Uh the reality is we just have advanced pattern recognition right now.”

Rogo's philosophy is to build custom proprietary tooling that incorporates a firm's specific investment processes, to remain model-agnostic and data-agnostic so tools don't become obsolete if a new frontier model (from OpenAI, Gemini, Anthropic) emerges or if a provider suddenly becomes unavailable, and to prioritize speed of iteration and rapid deployment of tools to investors.

“In terms of our philosophy for how we iterate and how we think, you know, the the best firms do it's it's about building that custom proprietary tooling. It's about building in the processes that are relevant to your work because that's actually what's going to give you an edge over time. The second thing is just building a strategy that is model, framework, and data agnostic. And so that if a better model comes out from OpenAI or from Gemini or from Anthropic, you're not sitting with a bunch of tools and a bunch of tooling that can't take advantage of those developments.”

Proprietary tooling should be built only around processes and workflows that embed the firm's unique competitive edge; generic research processes like tone analysis, transcription review, and earnings note synthesis are not proprietary and should be bought or partnered rather than built internally, freeing engineering resources for true differentiation.

“I think there's more that can be offloaded than should necessarily be proprietary. I think, you know, just to be very tactical and I'm going to go through with some of those tools you can just buy today, what would actually be a useful question, what would actually be, you know, useful ways to use them. Um and then I'm going to demo Roggo and and talk about how we work, especially relative to tools like Claude that that Brett mentioned. Um but how you can use ChatGPT, how you can use Claude today, they're really very similar, is those sort of file uploads as Brett mentioned.”

Andrew Carr tested ChatGPT by asking it to provide recent hospital transactions with multiple information; ChatGPT generated fake hospital names and fabricated URLs that sounded legitimate, demonstrating the hallucination problem with consumer tools for financial analysis.

“So, I tried using um ChatGPT and um you know, I I asked a question something like, you know, give me like hospital transactions, you know, recent transactions with multiple stuff like that. And it basically spat out hospitals that it made up. Like false information. And then when I asked for the sources, it made up URLs. Um and they they all sounded like legit.”

Models working at inference time (reasoning during generation rather than just pre-training/post-training) are getting smarter but will take significantly longer to reply (60-90 seconds versus near-instant), and organizations that built tools assuming fast response times will be disrupted if they don't anticipate this architectural shift.

“Uh just, you know, last week and over the last few months, uh we've really seen a an emphasis going from pre-training and post-training models to models working at inference time. And that's great and they've gotten a lot smarter and I won't go into the weeds on what that means, but the sort of tangible downstream consequence is that they're going to take a lot longer to reply to you. And if anyone's used the new ChatGPT, they'll realize, "Hey, it's now taking 60 seconds, 90 seconds before it replies to me." And that can be totally fine if it's giving you a way smarter answer, but if all your tooling has been built in a way that like you're reliant on it replying pretty quickly, you're going to be really upset when these models and these frameworks change.”

The LLM evaluation field is rapidly evolving with new metrics emerging (Elo-based preference systems for comparing model outputs, measuring breadth of data used in answers, accuracy metrics) beyond simple binary correctness, enabling more nuanced assessment of AI quality as tools mature.

“The second thing and the other metrics you can use is how much data was actually used in the answer. Very easy for us to say, oh well, in our answer we actually looked over the course of five years of transcripts and filings and so that's what's being fed in and we can measure that um in addition to accuracy. And so there's all sorts of sort of metrics and the field of LLM evaluation is one that's, you know, rapidly rapidly picking up speed um and I think we'll only get better and better at learning how to judge these tools.”

Idea generation is gaining mindshare at financial firms as a generative AI use case; new techniques around using large language models to develop theses and flag opportunities are emerging as alternatives to traditional quant alerting and signal processing systems.

“The final thing is that idea gen is gaining mindshare and we're seeing this you know both everyone knows with classical AI ML techniques things like alerting systems you know signal processing systems even for sort of traditional quant mental funds have always been popular but using gen AI for all of these things I'll caveat everything I say with you know AI has been used ubiquitously across Wall Street for a long time. I'm really focused on generative AI and so I'm going to leave out a lot of things that people are probably saying well you know don't we do this don't we do that shouldn't I be doing this. Likely all true I'm really focused on the sort of emergent use cases of large language models and to that end for idea generation there's a number of new techniques and sort of new interesting possibilities around developing a thesis flagging an opportunity you might not have been able to find before using these sorts of tools.”

The future of AI in investing will evolve toward AI agents: specialized intelligent systems that handle distinct sliced tasks (e.g., one agent that continuously monitors for tone shifts, another that tracks institutional ownership, another that compiles expert interview summaries), with the human investor directing the agents and integrating their outputs.

“What will be coming and what is actually on the horizon looks much more like these sort of far-fetched AI agents that everyone keeps talking about. Um and the way that we think about agents and and workflows is really that, you know, you're going to have a lot of different tools, a lot of different sort of analysts to do the different parts of your job, right? And maybe Brett will have an agent that, you know, does some of that preliminary work on looking for early signs of a tone shift because of the the margin compression. And he'll always be able to go to that agent and ask the question on the host of new names. And then 6 months later, he won't have to ask it, it'll just come to him, you know, every time earning season's coming up with those reports saying, "Hey, we could have seen this come." Um and that's really how we think the industry's going to evolve.”

When management language shifted from 'off-the-chair bullish' about 'unlimited demand' in July and October 2020 to merely 'strong demand' in January 2021 at Quidel, that tone shift was an important early warning trigger for a short position, even though the fundamental narrative remained positive to new observers.

“Every management meeting meeting I was in an incredibly off-the-chair bullish off-the-chair bullish off-the-chair bullish in January 20 2021. There was a I had a meeting and and he he was just bullish, the CEO. Like anyone who walked into that meeting for the first time was like, "Yeah, he sounds great, very optimistic." But in the context of how off-the-chair bullish he had been talking about unlimited demand in July and October to just strong demand in January, that deviation was an important tone shift.”

Humana has been a Medicare Advantage winner over 15 years as a 15-bagger compounder with historically consistent 4-6% operating margins (a quasi-public utility business), but recently margins have compressed to 2.6% in 2024 due to medical cost and utilization pressures, regulatory headwinds (two midnight rule, physician fee schedule), and 2025 rate notice challenges, causing the stock to decline 40-50% from peaks.

“Humana has been a Medicare Advantage winner over the last 15 years. Been a big compounder, been, you know, 15-bagger over the last 15 years. Um but recently, it's had, you know, some real issues on on the margins. Margins have been under pressure. Stock's off 40, 50%. Um you know, over the last 18 months.”

Building autonomous coding systems is receiving far more venture capital and research attention than building autonomous analyst systems, even though the analyst use case is larger; this creates opportunity for entrepreneurs to solve underserved AI-for-investing problems.

“And that's what I mean, if you look at some other domains, right? Like autonomous coding engines. There's been a lot of hype and so much more money going into that than you know, autonomous AI analysts.”

The May 2024 Humana BofA conference presentation contained breadcrumbs hinting at ongoing operating pressures and a more cautious outlook on utilization and margin recovery, which would have allowed a trained investor to anticipate the severe tone shift and deterioration that occurred on the July 31st earnings call.

“And in this case, it was a big thumbs up, right? The May conference didn't explicitly predict the full extent of the challenges, but they did hint at ongoing operating pressures and more cautious pressures and more cautious outlook acknowledging persistent utilization and they gave some comments they left some breadcrumbs that would would set the stage for a longer path to margin recovery.”

Claude AI projects at $20/month were effective at summarizing the narrative and tone changes in Humana's communications since Q4 2023, identifying the July tone shift that was previewed in the May conference, and analyzing compensation incentive structures via proxy data, but were limited by storage constraints (could not process full 10-K retrospectives) and poor user interface (slow, buggy document uploads).

“I I used ChatGPT 6 months ago to try this and I'm like, "This is Like this is not working. This is not effective. They're they're not really capturing the what's happening here, right?" You know, fast forward 6 months, these models have just gotten better, right? They're they're just pretty good now at doing this and I thought this was pretty good.”

Currently, treat AI tools like interns—they provide first-pass work and help new analysts ramp quickly. By end of 2025, expect ubiquitous solutions that create pixel-perfect work reliably. By 2030, expect true thought partners that you ask for strategic advice and meaningful answers.

“I mean, this is where I'm getting a little bit into the fluff before I show you where we actually are. Uh today, I really would treat these tools like interns. They're first pass at work. They can help you get started. They can help new folks ramp quickly. Uh and then they can help, you know, even if you're more senior, it's still helpful to have an intern. I think by the end of this next year, you're going to start to see ubiquitous solutions. Things that are able to routinely create sort of pixel perfect work the way that someone on your team would. Uh and that's just because there's these tools are spending a little bit more time in market and are getting better and better at serving very direct user needs. Finally, that sort of 2030 time horizon looks like a thought partner.”