Groundbreak, decoded for compliance teams

Fresh from Groundbreak 2025, our Founder CEO Omri Stern reflected that the Construction world is at an AI inflection point. His core insight: ‘AI doesn’t understand your domain—it understands language.’

Procore’s product announcements hint at that, although there’s more to what meets the eye: AI Agents automate workflows at scale—but only when teams encode what matters into clear rules. Without expertise, automation just accelerates toward the wrong outcome.

Specifically, insurance compliance AI Agents must validate across three layers: what the COI claims, what the policy actually says, and what the contract requires. Without that cross-check, you get the worst kind of risk. Compliance reports appear airtight while silently carrying the gaps that sink coverage in a dispute.

As Procore introduces Agentic capabilities, what Jones learned building Agents for insurance auditing comes down to a dual control model: AI executes the logic, auditors own the decision. Build that right, and you get accuracy that scales beyond pilots—and holds up when claims hit.

Expert-Guided AI for Compliance: The Operating Principles

Forget “clever prompts” that you experiment with on ChatGPT. In the world of insurance compliance, prompt engineering is a discipline for subject-matter experts. It’s about building systems where expert judgment scales through AI.

  1. Validate evidence, not just COIs. Use iterative prompts with explicit context and rules to validate what matters across the COI, policy language, and contractual requirements.

  2. Reason in steps you can fix. Turn compliance logic into sequential decision trees that AI can execute—and debug when outputs miss critical details or misapply rules.

  3. Protect the decision. Treat privacy, auditability, and expert oversight as system requirements. Define clear escalation paths when confidence drops or documents contradict.

This is risk consulting as infrastructure: codifying insurance expertise into rules that scale safely.

→ See how Jones simplifies risk management—request a risk consulting demo.

Inside the Jones Prompt Engineering Lab

Training an insurance AI Agent is like onboarding a brilliant new hire who can read 10,000 documents per hour but doesn’t know the difference between Additional Insured and Named Insured—until you teach them.

We treated LLMs like junior analysts and our auditors like staff engineers. The challenge wasn’t getting AI to read certificates; it was teaching it what matters.

Initial test: In December 2024, we fed real insurance certificates into ChatGPT, Gemini, and Llama. They read text fluently—but couldn’t reason through coverage. They misread limit structure—e.g., Umbrella limits applying to Each Occurrence in defined cases—fell short on basic insurance logic, and were inconsistent in relating COIs to their endorsements.

The shift: Auditors first tightened how instructions and context were written. When the model’s accuracy was at risk, they switched to coded, deterministic logic—for example, telling the system exactly which page or clause to check, or chaining prompts so the model followed the same steps an auditor would. In practice, they worked as AI operators: translating rules into structured prompts, debugging bad outputs, and validating results with rule engines.

Twelve months later: Audits ran 25% faster with accuracy at 99.9%. But speed wasn’t enough—auditors needed to trust the AI. So we designed the system to ‘show its work’: every decision includes side-by-side comparisons of requirements versus results alongside AI-drafted gap comments. The AI handles the tedious 90%; auditors validate the critical 10% that requires judgment.

We’re accelerating compliance across 150,000+ vendors and 31,483 projects with logic that scales with depth and accuracy.

Turn automation into defensible decisions

AI without domain expertise is like Google Maps without real-time traffic data: technically correct, dangerously incomplete. Assume every compliance call will be reviewed by someone who wasn’t in the room. Build for that: rules tied to contracts, evidence tied to rules, routed exceptions.

At Jones, we’re both the insurance experts and the AI builders. The result: an AI system running 50,000+ insurance rules to validate COIs, endorsements, and policies—expertise that scales.

🔍Transparent by default.The rule applied is shown by name/ID with the exact source excerpt, the decision path is logged and low-confidence or high-impact calls route to an auditor with the evidence attached.

🧱Durable by design.To ensure updates don’t compromise existing checks, rules are versioned and rerun against known edge cases—covering the full spectrum of contractual risk, from entity naming to exclusion carve-outs.

 

→ Schedule a walkthrough to see what insurance auditing looks like when expertise architects the AI.

→ Read Omri’s full analysis: AI Ground Truth in The Built World