Duck Creek's Agentic AI Platform Redefines the P&C Vendor Stack

Duck Creek used its Formation '26 conference on April 28 to launch a five-layer agentic platform, anchored by a Model Context Repository that combines carrier-specific rules, knowledge graphs and neuro-symbolic reasoning with generative models.

The architectural claim worth testing is the constraint layer. Rather than training one model to run underwriting end to end, LLM outputs pass through deterministic rule checks before any decision reaches production, so each one traces back to a policy clause, a rating table entry and the rule that fired.

Key Takeaways

• Every output is checked against filed rates and policy forms before it lands, which produces a replayable audit trail showing what evidence was used, which clause version applied, and which rule fired.
• BCG's $80 billion of annual U.S. P&C impact is roughly 9.2% of the industry's approximately $870 billion of 2025 net premiums written, which is why the number needs decomposing before it is used as a planning figure.
• Only 22% of insurers plan to have an agentic solution in production by year-end 2026, against 82% reporting some AI adoption and 7% at full operational scale.
• The Agentic FNOL application runs on Gemini through Google Cloud, so carrier loss data traverses Google's infrastructure just as the EU AI Act's high-risk requirements for claims systems take effect August 2, 2026.
• 370-plus carrier customers, including 33 of the top 50 North American insurers, carry more than $150 billion of annual premium through the platform, which is distribution no standalone AI vendor has.

What the Constraint Layer Is For

The design addresses a specific failure mode rather than a general capability gap, and the failure mode is what makes insurance AI hard to govern.

Pure LLM agents produce plausible underwriting decisions that can violate filed rates, contradict endorsement language or ignore jurisdiction-specific coverage requirements. The Model Context Repository sits underneath every agent as domain memory, combining fine-tuned generative models, traditional machine learning and neuro-symbolic reasoning grounded in the carrier's own rules and knowledge graphs, so generative output is constrained by actual policy forms and rating algorithms rather than by the model's parametric knowledge.

Above it, an orchestration layer coordinates agents across one workflow and supports three execution modes: full automation inside predefined parameters, semi-autonomous with human checkpoints above a complexity threshold, and manual override on an AI recommendation. An assurance layer embeds traceability and compliance controls in the execution path rather than alongside it, and an AI Gateway exposes an agent registry supporting the Model Context Protocol and Agent-to-Agent protocol.

The two launch applications are the Agentic Underwriting Workbench, which parses submissions, flags missing fields and scores risk against appetite, and Agentic FNOL, which routes claims across digital, voice and mobile with coverage verification and fraud pattern matching at intake.

Where the $80 Billion Goes

The projection that will drive budget conversations deserves the same scrutiny as any other assumption, and it decomposes into four levers with very different evidence behind them.

Boston Consulting Group projects up to $80 billion of annual U.S. P&C impact from agentic AI, with three-to-five-point loss ratio improvement and up to 36% efficiency gain in complex commercial lines. Against approximately $870 billion of U.S. net premiums written in 2025 and a combined ratio near 100%, $80 billion is about 9.2% of premium.

Value Lever	BCG Implied Impact	Stress-Test Assessment
Loss ratio improvement (3-5 pts)	$26B-$44B	Plausible at scale, but assumes full adoption across all lines. Personal auto and homeowners loss ratios are already optimized by incumbents like Progressive. Achievable range: $15B-$25B by 2030.
Expense ratio reduction (1.5-2 pts)	$13B-$17B	Consistent with Morgan Stanley's $9.3B projection for AI-driven expense savings by 2030. Aligns with Chubb's disclosed 1.5-point expense savings target from AI transformation. Achievable range: $8B-$13B.
Premium growth from faster quoting	$10B-$15B	Incremental growth from reduced submission-to-quote cycle times and improved hit ratios. Depends heavily on market conditions; less relevant in a softening market where capacity exceeds demand. Achievable range: $4B-$8B.
Fraud reduction	$5B-$8B	Consistent with the lower end of Deloitte's $80B-$160B savings projection, which we stress-tested in our earlier analysis. Achievable range: $3B-$6B.

Stress-tested lever by lever, the achievable range lands at $30 billion to $52 billion by 2030, roughly 40% to 65% of the headline. The gap is adoption: only 22% of insurers plan an agentic solution in production by year-end 2026, so the $80 billion describes a theoretical maximum rather than a forecast.

The pricing consequence of the architecture is more concrete than the projection. Where an agentic workbench generates a risk score or a price, the neuro-symbolic layer maps that output to specific rating factors and class codes, which is what makes it checkable against a filed rating plan. That check is a real piece of work rather than a claim to accept: territories, class codes and schedule rating modifications vary by state, and the constraint logic either encodes the filed algorithm correctly or silently does not.

The submission-side arithmetic is where the efficiency case sits. Data assembly consumes up to 50% of underwriting staff time on Cytora's research, so compressing it changes the denominator in hit-ratio calculations rather than the numerator.

Cytora launched Autopilot in March 2026 as a core-agnostic orchestration layer, reaching carriers on Guidewire, Majesco or a mainframe at the cost of API latency where real-time endorsement or billing data is needed. Verisk went the other way, telling its Q1 2026 call it had won a competitive RFP as co-development partner for a global insurer building a digitally native underwriting entity, betting that the data feeds are the durable position.

The Dependencies Underneath

Two things the platform changes are not on the feature list, and both land on people who did not choose the vendor.

The first is the model dependency. Agentic FNOL was co-developed with Google Cloud and runs on Gemini, so carrier loss data traverses Google's infrastructure. For carriers operating in the EU, where the AI Act's high-risk requirements for claims processing take effect August 2, 2026, that is a cross-border flow inside a system already classified high-risk. The AI Gateway's protocol support is designed to make the model swappable, and swapping a foundation model under a live claims workload is a different exercise from swapping it in an evaluation.

The second is the triangle. Claims entering through Agentic FNOL arrive with coverage verification and fraud indicators attached at first notice, which is exactly the intended improvement and also a change in the data generating process behind the reserving triangle. Claims processed through the new intake and claims processed through traditional channels may develop differently, and for as long as both run in parallel the paid and incurred triangles blend two populations under one set of development factors. The improvement and the discontinuity arrive together.

Differentiation is the quieter cost. Duck Creek spreads platform development across 370-plus customers, which is the buy-side argument, and it means every customer deploying the same Agentic Underwriting Workbench gets the same submission triage. What separates them afterward is appetite calibration, risk selection and pricing sophistication, which the platform explicitly does not supply. A midsize carrier writing $500 million to $5 billion cannot match the engineering capacity behind AIG's Palantir build at Lloyd's Syndicate 2479, and buying the orchestration moves the competitive question onto ground the vendor does not cover.