Gabriel Stengel

The through-line

Stengel's starting point is an absurdity he witnessed directly: the largest transactions in the world run at the speed of exhausted junior staff using ancient software. "The world's largest most important business transactions M&A between hundred billion dollar companies was bottlenecked and rate limited by 22y olds and 50-year-old to tools at 3:00 a.m." [1]. He sets that against the rest of markets, where "you can go on to Robin Hood and buy a bunch of stock and Citadel can make market make and, you know, do hundreds of millions of transactions instantaneously, but if you're Bob Iger and you want to sell Disney, you're going to be rate limited by a kid who just started at Goldman Sachs and then you're going to have to pay $20 million for it" [1]. The path from those nights as a junior banker at Lazard to building an AI company for finance is the narrative he keeps returning to [2][6], and it has become a broader argument about where AI belongs in financial services [4][5][8].

Everything else follows from that: the goal is to be "the intelligence layer for finance," which he defines operationally rather than grandly, as making users "smarter" in the context of the job to be done, meaning "better decisions faster," plus time returned from what he calls "the accidental complexity of finance of normalizing data pulling numbers research putting logos on a PowerPoint page doing Excel modeling" [1].

On why finance deserves its own AI platform

Asked why build for one vertical instead of everyone, Stengel argues that generality is what destroys quality: "the building in a generalizable way obviates all the small details, the long the context management, the integrations that are specific to your vertical" [1]. At Rogo, "the prompts are purpose-built, the context management is purpose-built, the UIUX, the features, the integrations, the data, everything is specific to financial workflows" [1]. The proof case in his demo is mundane on purpose: a comps table, or a query over Crunchbase and PitchBook data to estimate what Thrive owns of Cursor, work that "a chatbt would maybe not be capable of because it doesn't have the long tale of integrations with a faxet at a cap IQ a Bloomberg and so on" [1].

The second and third reasons are organisational rather than technical. Large banks require "a very white glove approach," which is why Rogo runs a deployed intelligence team that goes into a client such as Wells Fargo or Lazard "and teach them how to use the product, scope the use cases to what they're doing, connect their data," sometimes with custom code or custom prompts [1]. And finance sets "very high bars to entry around security, compliance, and regulatory issues," from flagging MNPI on data upload to deploying in a way "that a bulge bracket bank is comfortable with" [1].

On what an agent is actually missing

Stengel's mental model for building agents is deliberately narrow. He is untroubled by raw model intelligence, saying GPT-5 "is probably smarter than all of us here combined. Um, it just doesn't have the right tools, right? It literally cannot go out in the world and do anything" [1]. His job, as he frames it, "is give it the same tools that a human analyst has" [1]. For any finance workflow, only three things can be absent: the tool, whether that is Excel, PowerPoint, screening, updating a CRM or a VDR; the data, meaning a FactSet or PitchBook licence, transcripts, private market or private credit data; or the knowhow, where "the reasoning isn't there" [1]. The third is the interesting one and the least tractable. He puts it concretely: "If you've never scrubbed comps before for, you know, an MD who's very granular about the IBIDA adjustments for TMT companies, how are you going to know how to do it?" That may or may not be in domain for the model providers, and "it's sort of up to you to solve those reasoning challenges" [1].

On accuracy, auditability and the banking hierarchy

He does not promise perfection, and he grounds that in how banks already work: "we expect we're not going to be fully accurate the same way an investment bank expects their analyst is not going to be fully accurate that's why there's an associate and a VP and a director and an exec director and an D who all you know tear apart that work before it goes to a client" [1]. The design consequence is that usability equals auditability. Every figure links back to a source, so a claim about helping billions of people leads straight to "exactly where Sundar said that," and the same holds whether the underlying material is a transcript, internal data, a PDF or an Excel model [1].

On model training versus great engineering

Stengel sorts AI application companies into three camps and is explicit about which one he thinks most founders should be in. There are those who "treat RL and model training like a big hammer and every problem's a nail," which he associates with the large labs [1]. There are companies like Manus and GenSpark, which he calls best-in-class consumer agents from a product perspective and whose stated position is that "they should never ever have to train a model," relying instead on context window management, KV caching for low latency, tool access and memory [1]. And there are companies that engineer their way to differentiated data and only then train, the example being Cursor, which collected "a boatload of data mainly in the form of diffs," whether users accept or reject an edit from the tab model, and used it to ship an autocomplete model "far superior to what the labs have been able to do" [1]. His conclusion is a discipline rather than a doctrine: "you really need to know why you're doing what you're doing. And if you're trying to train a model, you should have a very good idea of why you're training a model and not just relying on great engineering unless you're open AAI" [1].

On domain expertise and short feedback loops

He believes "the value of expertise is just going through the roof," visible both in companies like Merkore whose business is sourcing expertise and in the pattern of app-layer companies being started by people with experience in the domain [1]. His argument is about the feedback loop rather than credentials. Building these products should not run through the usual cycle of interviewing a customer and returning to the system later: "If you can't look at an output yourself and kind of tell is that 90th percentile, is that 50th, is that 99th, it's going to be very hard for you to actually build the product for your end user" [1]. That same judgement is what lets someone act as a forward deployed engineer, explain the product to a customer and walk them through the use cases [1]. Expertise also has to sit inside engineering, not just product management, because someone must design "the systems that take place and need to happen for you to actually get the context the agent needs" [1].

On not getting bitter lessoned

The rule he repeats is to avoid over-investing in narrow fixes that the next model generation will erase: "if you try and solve a very niche small problem, you're probably going to get blown out the water by someone who trains a model to solve a much bigger problem" [1]. His example is Bloomberg's NLP models for classifying article sentiment, which GPT-3 outperformed while also doing far more [1]. Rogo has lived this. Started roughly three years ago, "we've had to rebuild everything, you know, every 6 months, you know, maybe maybe more often than that" [1]. A year ago the system had to explain EBITDA calculation, PowerPoint formatting and the fact that "bankers tend not to use the Oxford comma," and "all of that work was wasted. The the base models know that now" [1]. Those small models and components became "a complexity tax on our system," steadily peeled out and thrown away, which is why he now lets the agent start from scratch [1]. The judgement call, as he frames it, is deciding "when am I going to invest in something specific to make up for the gap in the intelligence of model right now and when am I going to let kind of the rising tide of quality, you know, lift me up with it," since "if you don't get to high quality today, you're not useful" [1]. He extends the same open question to memory, noting the thesis that memory is just a tool, writing notes to a file system and retrieving them later, something models are not yet good at "but but that might get solved" [1].

On why high finance is worth the work

His reason for choosing this domain has two parts. The commercial one is blunt: "if you want to work on a problem set in AI where there's an unbounded price people are willing to pay to be smarter that's in the financial services domain" [1]. The other is about who gets served. If the intelligence layer makes "capital allocation just a little bit more efficient," gets capital to the right businesses and teams, and democratises access so that "emerging markets companies you know companies in middle America that can't pay to work with Goldman Sachs or JP Morgan to sell themselves or raise capital can," the result is "a vastly healthier better global economy" [1]. The product surface reflects the ambition: comps tables, PowerPoint generation, table analysis, scheduled agents, and heavier workflows he calls pro and deep research, where output quality "is only increasing" [1]. Rogo is used by tens of thousands of bankers across some of the world's largest banks, investment firms, asset managers and family offices, and the sixty-person New York team is hiring across infrastructure, platform, data and product edge engineering [1].

Takeaways

• The founding observation: the world's biggest M&A deals are "bottlenecked and rate limited by 22y olds and 50-year-old to tools at 3:00 a.m." while retail trading and market making happen instantaneously [1].
• Vertical AI wins because generality erases detail; Rogo's prompts, context management, UX, integrations and data are all purpose-built for financial workflows, with a deployed intelligence team doing white-glove rollout inside banks [1].
• An agent fails a finance task for only three reasons: missing tools, missing data licences, or missing knowhow and reasoning [1].
• Don't promise perfect accuracy, promise auditability; banks already assume analysts are wrong, which is why there is an associate, VP, director and MD reviewing the work [1].
• Know why you are training a model rather than engineering; Cursor earned the right by harvesting accept/reject diffs, while Manus argues you should never need to train at all [1].
• Expect to throw work away: Rogo rebuilt everything roughly every six months, and hand-built scaffolding for EBITDA, slide formatting and banker style conventions became a complexity tax once base models absorbed it [1].
• Founders need enough domain judgement to grade an output as 50th versus 99th percentile themselves, because the customer interview loop is too slow [1].
• Finance is the domain with "an unbounded price people are willing to pay to be smarter," and the upside is broadening access to capital-raising services beyond firms that can afford Goldman Sachs or JP Morgan [1].