What Gennaro Brooks-Church Did Differently When Building Underwriting AI

00:00:00: Picture this, you're sitting in a mandatory all-hands meeting.

00:00:02: Right And the executive team is up there and they are practically glowing.

00:00:08: Oh yeah we've all been in that room

00:00:09: right?

00:00:10: They announced the rollout of this brand new supposedly game changing AI system.

00:00:15: Yeah...and they promise it's gonna completely revolutionize your department.

00:00:19: Huge promises

00:00:20: always Always.

00:00:22: But then Monday morning rolls around You log in The harsh reality sets in this multi-million dollar piece of software just added like, twelve agonizing confusing steps to a process that used to take you maybe five minutes.

00:00:38: Yeah

00:00:39: it is uh...it's honestly the universal tragedy of enterprise software.

00:00:43: It really is.

00:00:43: And y'know..It happens because there was just massive persistent disconnect A gap between people signing checks in their boardroom and actual people sitting at desks trying to well Just get their work done.

00:00:54: Exactly

00:00:55: So today's deep dive is all about a founder who looked at that exact disconnect and he looked it specifically in the you know highly complex world of Insurance underwriting which

00:01:05: is incredibly complex.

00:01:06: Yeah,

00:01:06: oh absolutely.

00:01:07: And he basically completely rewrote the code on how AI should be built.

00:01:13: We're diving into the work of Gennaro Brooks Church.

00:01:16: He's the CEO and Founder of Kazamir and Brooks Energy and he looks at this totally predictable almost I don't know, stubborn pattern of how Enterprise AI is deployed and he just flipped every single core assumption on its head.

00:01:29: which is what makes this such a crucial case study for us to look at, because I mean whether you work in insurance or tech finance really any industry that's currently staring down an AI revolution.

00:01:39: Which was

00:01:40: pretty much all of them right now?

00:01:41: Right exactly!

00:01:42: This is a blueprint for how artificial intelligence can actually be integrated into a human workforce as a functional tool rather than just becoming this massive disruptive burden.

00:01:52: Okay let's unpack this Because to understand what Brooks Church did differently we have We have to look at the trap.

00:01:59: The rest of the tech industry usually falls into

00:02:01: yeah, the classic playbook.

00:02:03: right.

00:02:03: for years AI and insurance underwriting followed this incredibly rigid Playbook.

00:02:09: a Tech company builds This massive language model or some new extraction algorithm.

00:02:14: They automate A few you know repetitive data entry steps And they immediately push it To market without

00:02:20: testing in the real world

00:02:21: exactly.

00:02:22: they start with the shiny new tech and then they go hunting for a problem that it can theoretically solve.

00:02:27: It's just the completely backward engineering process, I mean you end up with these software developers in Silicon Valley designing workflows for say commercial underwriters in New York or London and these developers have never actually seen what an underwriter desk looks like during renewal season.

00:02:44: they just imagine friction points rather than observing them.

00:02:47: which is buying this wildly expensive, voice-activated smart toaster.

00:02:54: And only after you plug it in on your counter do try and figure out what household problem that actually solves?

00:02:59: That's

00:02:59: a perfect analogy yeah

00:03:01: But Brooks Church reversed this.

00:03:03: instead of starting with like A blank coding screen or some pre packaged algorithm he starts by physically putting himself inside real underwriting teams.

00:03:12: Yeah she watches the process.

00:03:13: He observes the actual daily grind before single line back end architecture is even finalized.

00:03:19: But

00:03:21: okay, I do want to push back on this a bit.

00:03:22: Sure because...I mean i hear this whole user-centric design pitch all the time of The Tech World.

00:03:29: it's kind of buzzword.

00:03:30: Does sitting behind an underwriter for a week actually fundamentally change the hard code base?

00:03:36: Or is this just you know clever UX window dressing To make the staff feel heard?

00:03:42: Oh no!

00:03:42: What's fascinating here Is that It fundamentally alters the entire architecture Really

00:03:47: The architecture itself.

00:03:48: Absolutely, because the reality of enterprise data is just entirely different from the theory of Enterprise Data.

00:03:53: What do you mean by that?

00:03:54: Well if you build an AI in a vacuum You assume data flows logically right.

00:03:57: Clean spreadsheets perfect forms

00:03:59: Right...the ideal scenario.

00:04:01: But when actually sit with underwriting team You realize A broker might send them massive submission packet.

00:04:07: That it's just absolute mess.

00:04:09: Oh

00:04:09: yeah Like hundreds of pages.

00:04:11: Exactly, it's a two hundred page PDF.

00:04:13: It's got loss runs schedules value random emails and literally scanned faxes with like Coffee stains on them

00:04:22: right?

00:04:22: Right its unstructured chaos.

00:04:24: That's

00:04:24: exactly what it is.

00:04:25: So if you start at the tech You end up building an AI that tries to force all that Chaos into a rigid predefined template

00:04:32: which never works.

00:04:33: Never.

00:04:33: The AI inevitably fails on the edge cases.

00:04:36: And then what happens?

00:04:37: The human underwriter has to step in, open three different screens manually re-enter the data and basically just do the AI's job for it.

00:04:44: Ugh!

00:04:44: Which defeats the whole purpose of having an AI.

00:04:46: Precisely.

00:04:48: But when you start by observing a team first You build the data pipeline differently.

00:04:53: You realize the AI's primary job isn't to be this perfect, rigid extractor.

00:04:58: Its job is to untangle the unstructured mess in a way that aligns with the underwriter specific messy reality.

00:05:07: So you build a system That is fundamentally designed To handle exceptions not just standard rules.

00:05:12: Okay!

00:05:13: That makes total sense.

00:05:13: Your building tool with workers Not just, you know for some theoretical version of it.

00:05:19: Exactly.

00:05:19: And that observation phase leads right into how these systems handle the inevitable moment when AI makes a mistake Because in traditional enterprise tech You always run to this wall Of what they call frozen automation.

00:05:31: Oh

00:05:31: yeah Frozen Automation is essentially The silent killer of productivity.

00:05:35: Right

00:05:35: because In older AI models Engineers built This one time automation layer.

00:05:40: So system extracts data Based on static set rules.

00:05:46: It doesn't matter if the process is ten insurance submissions or ten thousand.

00:05:49: Right, it never evolves

00:05:50: completely static.

00:05:51: Yeah, its like having a highly enthusiastic but just incredibly stubborn intern.

00:05:56: Yes They make the exact same mistake.

00:05:58: every single day you tell them Hey The property addresses on page two not page four and then the very next day they pull the address from page for again.

00:06:08: I mean it's exhausting for the user right?

00:06:10: But Brooks Church broke that cycle.

00:06:11: He built systems where every single correction an underwriter makes is actually treated as new training data.

00:06:18: Yeah, the system actually learns from its daily use which is huge it Is.

00:06:22: and to really understand the mechanics of this you have to think about how traditional software is built.

00:06:28: It's often like um Like pouring concrete.

00:06:31: okay once the rules are set they harden.

00:06:34: Changing them requires this massive development cycle.

00:06:37: You know, a new software patch scheduled downtime

00:06:41: whole IT headache.

00:06:42: exactly what Brooks Church is doing Is more like a self updating map?

00:06:45: Like ways.

00:06:46: Oh I love that comparison.

00:06:48: right when a driver encounters traffic or a pothole and takes a detour the way system registers That Delta And it updates The route for everyone else in real time.

00:06:57: okay But how does that actually work For an underwriter In practice?

00:07:00: like if i'm reviewing A commercial property submission and the AI pulls the wrong square footage.

00:07:07: And I just take in the right one, does the AIs underlying code change right then or there?

00:07:12: Well not the core code base itself but the model's localized weighting and retrieval mechanisms change.

00:07:19: Got it.

00:07:19: When the underwriter highlights the correct value on the original document, The system captures that specific interaction.

00:07:26: It maps the visual layout you know... ...the context of surrounding text even the specific broker's formatting

00:07:32: quirks.".

00:07:33: Oh wow!

00:07:33: Yeah

00:07:34: it essentially writes a new micro rule for That specific type of document On-The-Fly.

00:07:39: So over time the AI is literally capturing the institutional knowledge Of the senior underwriters.

00:07:45: so really as an attentive apprentice instead of a stubborn intern?

00:07:48: Exactly you know, capturing all that institutional knowledge implies the underwriter actually trusts this system enough to use it in first place.

00:07:54: Which is the hardest part?

00:07:55: Right and high stakes fields like insurance trust just as massive hurdle because of black box problem.

00:08:02: Yeah!

00:08:03: The Black Box is a critical failure point in enterprise AI particularly heavily regulated industry like Insurance.

00:08:11: We're not talking about asking chatbot read marketing email here.

00:08:14: No Not at All.

00:08:16: we are talking assessing millions of dollars in risk, where regulatory compliance and audit trails are completely non-negotiable.

00:08:24: Right!

00:08:25: And yet so many these tools operate as these opaque black boxes... you know the system.

00:08:31: It ingests a hundred-page policy document and words for second, then it just spits out risk assessment score or data summary.

00:08:37: Without

00:08:38: explaining anything?

00:08:39: Exactly!

00:08:39: Doesn't show its work.

00:08:40: And an underwriter whose professional reputation is literally on the line.

00:08:43: that's terrifying

00:08:44: Absolutely terrifying.

00:08:45: But Brooks Church made transparency day one requirement In his software.

00:08:50: every single extracted datapoint must trace back to exact source documents in the exact page like it literally gives you receipts

00:08:59: brilliantly solves the audit problem mechanically.

00:09:02: I mean, if a reinsurer comes back a year later and asks why a specific premium was priced the way it was... Right The underwriter isn't forced to just shrug And say well-the algorithm told me too

00:09:14: Which would never fly?

00:09:15: Never!

00:09:16: Now they have direct cryptographic level link Back To The Source Truth.

00:09:21: But here's where It gets really interesting though because i mean you could argue that Providing all those receipts like a citation for every single data point could just create more visual noise.

00:09:33: For an underwriter, right?

00:09:35: Like they're already buried in information.

00:09:37: so why not just give them the clean extracted data and keep all those citations hidden on the back end Just for the auditors?

00:09:44: I mean is this just a technical feature to prevent AI hallucinations or there's something else going on here?

00:09:48: psychologically

00:09:49: Oh it's definitely A fundamental psychological shift.

00:09:52: It completely changes The human computer dynamic.

00:09:55: How so?

00:09:55: Think about it.

00:09:56: If an AI just hands you a clean sheet of data with no context, It is presenting itself as an oracle

00:10:02: Right like in all-knowing entity

00:10:04: right...it's basically saying I am the authority trust my output blindly.

00:10:09: But experienced professionals know that machines hallucinate and they know that real world context matters immensely.

00:10:15: Give me an example.

00:10:16: Sure, so if the AI says a building's roof was updated in twenty-twenty The underwriter might actually need to see the original instruction note To know well Was it a full replacement or just a quick patch job?

00:10:27: Ah okay That's a huge difference and risk huge

00:10:30: difference.

00:10:31: So by forcing the AI to show its work It basically demotes the software

00:10:36: Precisely.

00:10:37: It demotes the AI from this unquestionable oracle to just a highly capable, verifiable assistant.

00:10:43: By putting those bounding boxes around the exact text on the original PDF The AI is saying hey here's the data and where I found it.

00:10:51: Now you apply your human judgment.

00:10:53: Love that!

00:10:53: It completely changes how the underwriter feels about It

00:10:57: really does.

00:10:57: They're no longer fighting this machine that feels like it's trying to do their job, instead they are managing a system that just does all the heavy lifting of data triage which leaves the complex risk analysis-like the actual craft of underwriting to them.

00:11:13: and A lot AI companies get this so wrong in there messaging

00:11:17: constantly these

00:11:18: is Really vague language about over supporting workers.

00:11:21: but you look at their eventual roadmap And it is clearly aimed at automating the human out of a loop entirely.

00:11:28: Yeah, The ultimate goal is replacement

00:11:29: exactly.

00:11:30: but Brooks Church's systems explicitly draw line Like, the machine does high volume classification and human makes judgment calls.

00:11:39: Right!

00:11:39: But okay even if you have this perfectly transparent dynamically learning team centric system I mean it is completely useless to actually install it.

00:11:48: Oh

00:11:48: absolutely The implementation phase.

00:11:50: honestly where most enterprise software just goes to die.

00:11:52: The track record for Enterprise AI rollouts is brutal.

00:11:57: I mean, we are talking about months of setup.

00:12:00: Massive IT disruptions endless cross-debar mental committees and millions of dollars spent on integration consultants.

00:12:06: It's a nightmare for everyone involved.

00:12:08: but apparently Brooks Church configures and deploys these new systems For clients in days yeah not months days And this often involves him working directly with the client to map The software to their specific brokers and data formats.

00:12:22: But logically Wait, rolling out enterprise AI in days instead of months.

00:12:26: How is that even possible?

00:12:28: I know it sounds crazy!

00:12:29: How modular does this architecture have to be... ...to actually pull that off without completely breaking a client's legacy systems?

00:12:37: Well this raises an important question about how monolithic software compares to modular software.

00:12:42: Okay break that down for

00:12:43: me If a system takes eight months to deploy It usually because the software as a monolith The tech vendor literally has to spend months writing custom API integrations just to force the client's old legacy mainframe data into their rigid AI product.

00:12:58: Like trying to fit a square peg in an round hole.

00:13:01: Exactly!

00:13:01: They're basically rewiring entire house, just plug-in and new lamp

00:13:06: Which is incredibly risky

00:13:08: And expensive Highly Risky.

00:13:11: But To deploy In Just Days, the AI must be built as modular.

00:13:15: overlay and

00:13:16: overlay.

00:13:17: Yes the document ingestion engine, language models ,and user interface.

00:13:22: they are all completely decoupled.

00:13:24: so instead of ripping out clients old database this system acts as a flexible translation layer right on top.

00:13:31: it is smart because brooks church focuses heavily that human centric UI.

00:13:37: we talked about.

00:13:38: heavy lifting isn't done in deep back end code integration.

00:13:42: where's

00:13:43: It's done?

00:13:43: by pre-training the models to recognize standard insurance formats first, and then allowing actual underwriters to tailor that final mile of data mapping right there in the interface during week one.

00:13:54: So users themselves adjust it into the interface?

00:13:57: Exactly!

00:13:57: The system molds instantly rather than demanding a whole corporate environment change to fit this system.

00:14:03: Okay.

00:14:04: That proves that the whole team first philosophy isn't just a slick marketing slogan, it's quite literally baked into The Engineering from

00:14:12: the ground up.

00:14:13: and that need for speed ties directly into how Brooks Church measures the ultimate success of the technology doesn't?

00:14:21: It

00:14:21: does because the standard metrics For software success in the tech industry are usually completely detached From actual utility.

00:14:29: yeah

00:14:30: Usually at Tech startup measure success by the demo

00:14:33: The infamous boardroom demo.

00:14:35: Right,

00:14:35: can the sales team go into a board room run a highly sanitized perfectly scripted demonstration and get the executives to sign a five-year contract?

00:14:44: Yeah And if yes then boom it's a success.

00:14:47: It doesn't matter If every day employees hate using at six months later Not

00:14:50: to the salesteam anyway

00:14:51: Exactly.

00:14:53: But Brooks Church threw that metric completely out.

00:14:55: He doesn't care how flashy the demo is.

00:14:57: The only metric That matters To him Is everyday adoption

00:15:00: Which is the only metric that should matter.

00:15:02: Right, like.

00:15:03: do the underwriting teams actually keep logging into this system on a random Tuesday long after initial rollout?

00:15:09: excitement has faded?

00:15:11: He had his standard of adoption over appearance

00:15:14: which is remarkably unforgiving.

00:15:17: to hold yourself too.

00:15:18: Why'd you say

00:15:19: so?

00:15:19: Because a boardroom demo is just theater, right?

00:15:22: Yeah totally.

00:15:22: You control all the variables you control the pristine data being fed into the system and you are pitching to executives who frankly don't ever actually have To use this software themselves.

00:15:32: Right they're just looking at theoretical ROI

00:15:34: Exactly but judging tool by whether an underwriter willingly chooses to Use it on a rainy Tuesday afternoon When They Have Fifty Complex Renewals Sitting On Their Desk And A Stressed Out Broker Is Screaming At Them On The Phone.

00:15:48: I mean, that is the ultimate crucible of software utility.

00:15:51: Because if it adds even one ounce of friction in that moment they will completely abandon it and

00:15:56: a heartbeat

00:15:56: It'll minimize the AI window.

00:15:58: pull up their trusty old Excel spreadsheets get out a physical neon highlighter And just do at the hard way because it's predictable.

00:16:06: Exactly by focusing entirely on everyday adoption Brooks Church forced a standard of absolute uncompromising usefulness Yeah?

00:16:16: Or if it doesn't learn from human corrections or if it hides its work in a black box, adoption simply drops to zero.

00:16:24: Every day usage is literally the only metric that proves the AI as actually solving friction rather than creating more of it.

00:16:31: and That philosophy thankfully Is quietly starting to influence how major firms evaluate all their AI procurement.

00:16:39: So what does this all mean?

00:16:40: I mean when we zoom out and look at his entire stack of research We're really looking at a massive paradigm shift in how we think about automation

00:16:47: necessary shift.

00:16:48: Yeah, you've seen the mechanics of moving away from that top-down tech.

00:16:51: first engineering.

00:16:52: You've seen How continuous learning, you know treating every human correction as a real time architectural update saw us out That frozen automation problem.

00:17:00: right we explored why building verifiable receipts Right into the UI shatters The black box turning it threatening Oracle Into A trusted assistant.

00:17:09: And finally, we saw how modular architecture allows for rollouts in days proving that adoption on a rainy Tuesday really is the only true metric of success.

00:17:18: It Really Is A masterclass In Technological Empathy.

00:17:21: Technological

00:17:22: empathy I like That

00:17:23: it is.

00:17:24: it demonstrates that The Ultimate Goal Of Enterprise AI Shouldn't be to build a Machine that Works independently of Humans Right but Rather To Build a System Perfectly Calibrated to Amplify Human Expertise.

00:17:35: so For you listening right now The next-time management calls an all hands meeting to announce a massive game changing new software platform.

00:17:42: Get ready!

00:17:43: Ask yourself, was this built to look good in a sanitized boardroom demo?

00:17:48: Or is it build from the ground up with modular architecture and verifiable workflows... ...to be genuinely useful on your ordinary day?

00:17:55: because as we've learned….

00:17:56: …the gap between those two things is difference of tool you have to fight against….

00:18:00: And tools that actually want use.

00:18:02: Though, you know it does leave us with one final kind of structural question to consider.

00:18:08: What's that?

00:18:08: Well we've established at the true value of this specific approach to AI relies entirely on it continuously learning from the nuanced corrections of expert humans.

00:18:19: Right right The Waze analogy.

00:18:21: It needs drivers to report the potholes

00:18:22: Exactly.

00:18:24: But what happens to the capability of that AI ten or twenty years down the line If the current generation of senior experts eventually retires, and we've successfully outsourced all the foundational repetitive data gathering work to... ...the AI.

00:18:40: How do we actually train next-generation human expert?

00:18:43: Oh wow!

00:18:44: Because eventually someone has be knowledgeable enough know when machine is wrong And make that critical judgement call correct it.

00:18:50: Well yeah That's definitely something to chew on while you're staring at grid of faces.