- ep 21
- 7 min read
- July 8, 2026
The Engine Behind the Platform: Inside the World Insurance Model
Hosted by Katie Dowson and David Haddad
On this episode of The Advocate Insurance Desk, host Katie Dawson goes one layer deeper than the usual carrier-by-carrier market breakdown. Instead of a specific market, the conversation is about the engine underneath the platform: the World Insurance Model, or WIM. Joining her is David, head of product engineering at Advocate, who has spent roughly five years building it. The through line is simple. Nearly every commercial property carrying debt has to prove its insurance meets the lender's requirements, at origination and again at every renewal. That verification can run to hundreds of checks per policy, and much of it is still done by hand. Miss one check and the property can be under-insured. WIM is Advocate's answer to doing that work automatically, consistently, and at scale.
Key takeaways
- WIM (the World Insurance Model) is a deterministic rule engine: the same inputs always produce the same outputs, so a compliance result is testable rather than a judgment call.
- The market it addresses is large. David cites roughly 30 million commercial insurance policies a year and about 500 billion dollars in annual premium, with nearly every debt-secured property subject to a compliance process. [VERIFY: confirm 30M policies and 500B premium figures]
- In Advocate's own testing, WIM lifted a frontier model's gap-detection rate substantially at roughly the same cost, and let smaller-context models finish jobs they otherwise could not.
- Review time in one internal test dropped from about 90 minutes to a couple of minutes, moving humans toward judgment calls and away from rote data extraction.
- A human reviewer still outperforms the engine on accuracy, which is why citation and human review stay built into the workflow.
What is the World Insurance Model?
David frames WIM less as a compliance checker and more as a way to categorize the entire world of insurance: coverages, risks, perils, and causes of loss. A user creates a risk profile (their tolerance, the asset they care about, such as a multifamily home or a Manhattan office building), and WIM maps that against the coverages and requirements that apply.
Because the categorization is structured, the number of asset types is not fixed. David notes the system can expand the asset types it supports at any time, and that this flexibility is a direct result of how WIM is built. [VERIFY: current count of supported asset types, left unspecified in the recording]
What does "deterministic" actually mean here?
Deterministic is a word Advocate uses constantly, and David is careful to separate it from the "probabilistic" language attached to most AI models.
"When you had a math equation, if you put in the same inputs, you always got the same outputs. Five plus five is ten. There is no debate about it."
That is the point. A probabilistic model might tell you five plus five is ten today and something else tomorrow. WIM does not. If you check compliance on a multifamily home in Colorado, the same checks return the same output for the same input every time.
"It is not a thought, it is not thinking, it is not reasoning. It is just accurate."
A quick worked example from the episode: upload a flood zone determination certificate that says the asset sits in a flood zone, and the flood requirement activates automatically. The trigger is not a guess; it follows from the document.
How do the engine and the AI models work together?
This is the part that surprised me. Advocate does not replace AI models with WIM. It pairs them. The deterministic engine handles the core compliance and inventory work, while probabilistic models handle the reading and reasoning that engines are bad at.
Walk a policy and an appraisal report into the platform and the sequence looks like this. WIM tells the model which fields it needs to find. The model does the OCR and reasoning to pull those values out of the documents. WIM then decides which coverages matter for that asset type, runs the calculations, and returns a compliance report grounded in testable code.
"The model is not really doing insurance. It is using a partner, WIM, and using its reasoning abilities to do the tasks that WIM is instructing it to do."
Because the model only has to look for what WIM asks it to look for, token cost goes down and accuracy goes up. The calculations themselves never leave the deterministic side.
Does WIM actually improve the numbers?
David pulled up Advocate's labs page during the recording. The chart plots each model twice: once without WIM, once with it. The results he cited, all of which should be checked against the published labs page:
Claude Sonnet 4.6 found about 26 percent of gaps on its own at roughly 0.37 dollars per review (on what he estimated was around 1,000 pages of policy and asset content). With WIM, that rose to about 63 percent, at close to the same cost. [VERIFY: David flagged the page count as approximate and said to double-check]
GPT 5.5 moved from about 32 percent to about 58 percent. [VERIFY against labs page]
Some open-source and smaller-context models (he mentioned Grok and DeepSeek) could not finish the task at all without WIM, because 1,000 pages of extracted text overran their context windows. With WIM narrowing what they had to process, the job became feasible. [VERIFY: which specific models failed and their context limits]
At its most efficient, David put WIM's standalone accuracy on a compliance check at roughly 74 percent as listed on the labs page, against about 96 percent for a human reviewer. He added that since that publication, Advocate's "model harness" has pushed past 75 percent and is now closer to the 80s. [VERIFY: current harness accuracy, 74% and 96% figures, and the "closer to the 80s" claim]
What is the "harness," and why mix models at all?
The harness is Advocate's method for combining WIM with several frontier models rather than betting on one. David describes breaking a review into steps and assigning each step to whichever model is best at it: one model for reading an ACORD form, another for strong vision, another for identifying assets on a document, another for following methodical instructions. Mixing the right models per step is why the harness outperforms any single model paired with WIM.
What about hallucinations?
Since probabilistic models can invent answers, this is a fair worry in a setting where a single miss can hurt a lender. David's illustration was an anecdote about someone asking a chat model whether to drive or walk to a car wash one minute away, and getting an answer that missed the point. [VERIFY: this is an unattributed online anecdote, not an Advocate result]
Advocate's mitigation is structural rather than hopeful. The deterministic engine does the thinking that matters, so the model is not left to reason about compliance on its own. Everything a model pulls is cited, users can accept or reject each item, and the platform keeps traces of what the model was doing. The human stays the source of truth.
"Their time to review is compressing quite significantly, because the initial work is being completed by models."
Does this replace the reviewer's job?
David's answer is no, at least for now. The task that shrinks is reading documents and extracting data, which in one internal test went from about 90 minutes to a couple of minutes. What stays human is the judgment: what an endorsement really means, whether what the broker provided is sufficient, whether to reach back out. The repetitive extraction of limits and deductibles is what gets automated away.
What does this mean for brokers?
The broker angle was David's favorite takeaway. Brokers tend to specialize by asset class (multifamily, boats, and so on), and breaking into a new market usually takes hard-won expertise about which requirement sets apply and when.
"The hard part is not just reading the policy. It is knowing what rules apply and when."
WIM effectively democratizes that knowledge. A broker who normally writes boat policies can move into office buildings by browsing the coverage library, understanding each risk profile, and selecting the one that fits. The system takes over the requirement logic from there, which means a broker can grow their addressable market almost overnight.
Where can users see WIM in the Advocate app today?
According to the episode, getting started is deliberately low-friction. You can sign up for free with no credit card, and an onboarding tour walks you through it. To run a case, hit the plus button, create a case, and drop your documents in. The Advocate app handles the rest.
FAQ
Didn't find your answer?
If you couldn't find the answers you need, feel free to reach out to the host.
Listen to the episode
- Episode 21
- 24 min
The Engine Behind the Platform: Inside the World Insurance Model
What if the most accurate way to run an insurance compliance check is to stop the AI from thinking? Katie sits down with David, who built Advocate's World Insurance Model (WIM): a deterministic engine that roughly doubles coverage-gap accuracy while token cost stays flat.
Listen now