Generative AI reads like a brilliant intern with no memory of your rulebook. Neurosymbolic AI pairs that intern with a structured mind — an ontology and a rule engine — so every decision is auditable, consistent with your guidelines, and defensible to a regulator. This is a working explainer, with a live commercial property submission at the end.
Neurosymbolic AI is a hybrid architecture that combines the pattern recognition of neural networks with the logical rigor of symbolic reasoning.
Large language models are neural networks — statistical engines that learn from vast amounts of text. They are brilliant at reading unstructured documents, understanding intent, and producing fluent language. But they reason probabilistically, not logically. Two runs of the same question can produce two different answers, and the model cannot show its work in a way a regulator would accept.
Symbolic AI is the older tradition — knowledge graphs, ontologies, rule engines, and formal logic. It is rigid, explicit, and fully auditable, but it cannot read a broker email or interpret a loss-run PDF on its own.
Neurosymbolic systems use the neural side to perceive — extracting structured facts from messy inputs — and the symbolic side to decide — applying your underwriting guidelines, regulatory requirements, and portfolio rules with deterministic precision. The neural model proposes. The symbolic engine disposes.
Extract entities, relationships, and intent from unstructured submissions — emails, ACORD forms, loss runs, inspection reports, news, geospatial imagery.
Map every extracted fact into a formal domain model: what a Location is, what a COPE characteristic is, how Occupancy relates to Protection and Exposure.
Apply underwriting guidelines, capacity limits, regulatory constraints, and reinsurance treaty rules as executable logic — with every inference traceable to a rule and a fact.
Generate human-readable narratives, referral memos, and broker responses grounded in the reasoner's output — never hallucinated, always traceable.
The difference isn't about accuracy on any single task — it's about the kind of cognition each system performs. An LLM generates the most plausible next token. A symbolic reasoner evaluates the truth value of propositions against a formal knowledge base. Understanding this gap is the key to knowing where to apply each.
"Based on what I've read, this pattern usually leads to that outcome."
"Given these facts and these rules, this conclusion follows — here is the proof."
A submission comes in as text. The model tokenizes it and passes the tokens through dozens of transformer layers, each applying attention across the sequence. At each position, the model produces a probability distribution over its vocabulary and samples the next token.
There is no separate "reasoning step." What looks like reasoning — chain-of-thought, for example — is the model generating text that resembles reasoning it has seen in training data. It often gets the answer right because that pattern is well-represented in the corpus. It sometimes gets it confidently wrong for the same reason.
The model has no persistent representation of your underwriting manual. Even if you paste the manual into context, the model's attention to it is probabilistic; there is no guarantee rule CP-TIV-03 will be applied when it should be.
The same submission arrives as text. A neural extractor (often an LLM itself, constrained) parses it into structured facts: Location(id=L1, address=..., construction=JoistedMasonry, occupancy=LightMfg, tiv=USD 38M).
These facts are asserted into a knowledge graph built on an ontology. A reasoner — a rules engine or a description-logic inference engine — then evaluates every applicable rule. Rule CP-TIV-03 is represented as executable logic: IF tiv > authority_limit(underwriter) THEN require_referral.
The conclusion isn't generated — it's derived. Every derivation produces a proof: the fact, the rule, and the binding. That proof is what makes the output auditable, consistent, and defensible.
System 1 is pattern recognition at speed. It's the brain's autopilot — fluent, confident, and excellent for familiar problems. It's also where cognitive bias lives. When the pattern matches something common, System 1 is fast and accurate. When the situation is ambiguous, novel, or hinges on precise rules, it produces a confident answer that can be materially wrong.
System 2 is effortful thinking — the kind you use to apply regulations, work through a multi-step calculation, or weigh alternatives under constraints. It's slower and more expensive, but it's also the only kind of thinking that produces an auditable trail of why a decision was made. Underwriting is a System 2 discipline.
Because a neurosymbolic system has an explicit model of the world, it can apply the reasoning mode that fits the question — not just probabilistic association. Each of these shows up somewhere in the underwriting pipeline. The rule engine does deductive work. Classifying a risk is abductive. Pricing involves causal reasoning. Structuring a placement is constraint-satisfaction. Stress-testing is counterfactual.
Think of an ontology as the vocabulary and grammar of your business, written in a form a computer can reason over.
When an underwriter says "a joisted-masonry building in a protected Class 4 town with a sprinkler system in a light manufacturing occupancy," every word is loaded with meaning — and every word connects to other concepts. Construction class implies vulnerability to wind and fire. Protection class modifies that vulnerability. Occupancy shapes exposure. The underwriter's brain holds this web of concepts.
An ontology makes that web explicit. It names the classes (Location, Building, Occupancy, Peril), the properties (TIV, year built, construction type), and the relationships (a Building hasProtection, is exposedTo Perils, isLocatedIn a CatZone).
Once the world is modeled this way, rules become composable and inheritable. A rule that applies to "all combustible-construction buildings" automatically applies to anything the ontology classifies as combustible — frame, heavy timber, ordinary. You don't write the rule four times. You write it once, against the concept.
Most importantly, ontologies compose across domains. A location ontology plugs into a peril ontology plugs into a reinsurance ontology. The same structure that supports underwriting supports claims, portfolio management, and regulatory reporting.
An LLM might see "frame construction" and "wood frame" as similar strings. An ontology knows they are the same concept, and knows that concept is classified as combustible under ISO CC-1.
A rule attached to Combustible Construction automatically applies to every subtype — without re-authoring. Your underwriting manual becomes a small, maintainable tree instead of a sprawling document.
The same Location concept anchors underwriting, claims, portfolio, and reinsurance. One broker address resolves to one entity everywhere — the foundation of clean accumulation and PML analysis.
A broker submits a mid-market commercial property risk. Below is the actual submission. Then: how the process traditionally runs, how a pure LLM would handle it, and how a neurosymbolic system handles it — not just gating the referral but reasoning across the enterprise ontology to derive a priced, structured quote. The NSAI pipeline runs in five stages: extract facts from the submission, enrich with ontology-driven joins, evaluate underwriting rules, compose an explanation, and finally reason across pricing, reinsurance, portfolio, claims, and regulatory ontologies to produce the actual terms and premium. Every number traceable to its source.
Before introducing any AI, here's the traditional workflow a commercial property underwriter runs on a submission of this size. The process is sound — but it's slow, uneven across underwriters, and heavily dependent on manual interpretation.
Check for conflicts, prior submissions, sanctioned entities. Verify appetite fit.
Review Construction, Occupancy, Protection, Exposure for each location.
Evaluate loss runs, CAT exposure, accumulation vs portfolio limits.
Run rating model, compare to authority grid, determine referral need.
Accept, decline, or quote with subjectivities. Issue terms to broker.
A submission this size typically takes 5–8 business days across multiple handoffs. Two underwriters looking at the same risk routinely produce different terms. When a decision is challenged — by a broker, a regulator, or internal audit — the rationale often lives in one underwriter's email thread or handwritten notes. Accumulation exposure is checked manually against stale portfolio extracts.
A pure LLM approach: paste the submission into a prompt along with the underwriting manual, and ask for a recommendation. Fast, fluent, and dangerous to rely on. Here is what the model produced.
The output reads well, but it is materially incorrect on multiple counts that an experienced underwriter would catch — and a regulator would later find. These aren't subtle gaps; they are the kind of errors that become errors-and-omissions claims.
The neurosymbolic system runs the submission through four stages: neural extraction into the ontology, symbolic enrichment from external sources, rule-engine evaluation of underwriting guidelines, and neural explanation of the derived decision. Every step is inspectable.
| Rule | Condition | Inputs | Outcome |
|---|---|---|---|
| CP-APT-01 | Named insured must pass clearance & appetite screen | Metal Fab · NAICS 332710 | In Appetite |
| CP-TIV-03 | If TIV > USD 50M, route to senior authority | TIV = USD 84.5M | Refer |
| CP-CAT-12 | If location in Tier 1 named-wind zone, require 5% NWS deductible | L1 ∈ Tier1_NamedWindstorm | Subject To |
| CP-CAT-18 | If FEMA flood zone ∈ {A*, V*}, require flood sublimit or exclusion | L1 ∈ FEMA_AE | Subject To |
| CP-HAZ-07 | HighHazard_Welding occupancies require hot-work permit warranty | B1 ⊑ HighHazard_Welding | Warranty |
| CP-ACC-02 | Gulf Tier 1 accumulation must remain < 80% of treaty capacity | Post-bind: 79.0% → 80.4% | Breach |
| CP-LOSS-04 | Loss ratio < 40% over 5 yr — no load required | LR ≈ 11% (est.) | Pass |
| CP-DISC-09 | Prior non-renewal requires written disclosure from broker | Non-renewal flagged | Subject To |
Unlike the LLM, the NSAI system didn't stop at a fluent answer. It caught every compliance flag, diagnosed the accumulation breach with real math, then reasoned across six enterprise ontologies — treaty, portfolio, pricing, claims, regulatory filings, and commission — to find a cession structure that resolves the breach and arrive at a defensible net premium. Every value here has a named source.
Press Run. Both systems receive the Meridian Fabrication submission at the same moment. Watch the LLM sprint to a fluent answer while the neurosymbolic reasoner extracts, enriches, evaluates rules, composes an explanation, and then reasons across the enterprise ontology — deductively, causally, and by constraint-satisfaction — to structure and price the actual quote. Every number traced to a source.