Multi-Agent Orchestration Becomes the Carrier AI Playbook: How AIG, Gen Re, and Open Protocols Are Defining the Architecture

From analyzing carrier earnings calls and vendor announcements across the past two quarters, a clear architectural pattern emerges: the winners in insurance AI are not deploying smarter individual agents but building coordination layers that manage agent fleets as unified systems. AIG’s Q1 2026 earnings call revealed an orchestration layer live across commercial and E&S lines, processing 370,000 submissions without adding headcount. Gen Re published a practical reinsurance orchestration blueprint in December 2025. Verisk shipped Model Context Protocol connectors for Anthropic’s Claude in Q1 2026, giving the protocol standard real traction in insurance. And Gartner, while predicting 40% of enterprise applications will embed AI agents by year-end 2026, simultaneously warns that over 40% of agentic AI projects may be canceled by 2027 due to governance gaps.

The trade press has covered these developments individually: an AIG earnings story here, a Verisk product launch there, a Gartner forecast somewhere else. This article connects them into a single architectural narrative. The emerging pattern is multi-agent orchestration, where carriers deploy not one AI tool but coordinated fleets of specialized agents governed by shared protocols and human escalation triggers. Understanding this architecture matters for actuaries because the orchestration layer, not the individual model, is where pricing signals are routed, underwriting decisions are challenged, and governance controls either hold or break down.

AIG’s Orchestration Layer Goes Live

AIG’s Q1 2026 results provided the strongest public evidence yet that multi-agent orchestration is producing measurable financial outcomes. General Insurance underwriting income more than tripled to $774 million, up from $243 million in Q1 2025. The calendar-year combined ratio improved to 87.3% from 95.8%, an 8.5-point swing. Net premiums written increased 24% year over year. Adjusted after-tax income per diluted share rose 80% to $2.11. While multiple factors contributed to these results, including lower catastrophe losses ($180 million versus $525 million) and favorable prior year development ($132 million versus $64 million), CEO Peter Zaffino drew a direct connection between AI deployment and the company’s ability to grow premium without proportional headcount increases.

The critical architectural detail emerged in Zaffino’s description of AIG’s agentic AI expansion. “For example, one agent may handle submission ingestion and data extraction, another may perform risk evaluation against our underwriting guidelines and another could benchmark pricing against our portfolio targets,” he explained on the earnings call. This was not a description of a single AI tool performing multiple tasks. It was a description of specialized agents, each with a defined role, coordinated by an orchestration layer that determines activation timing, information sharing protocols, and human escalation triggers.

The metrics from AIG Assist, the carrier’s primary underwriting AI platform, reinforce the scale: a 30% improvement in quoting more submissions, a 55% reduction in time to quote, and a roughly 40% increase in binding of submissions. Lexington Insurance, AIG’s primary E&S carrier, surpassed 370,000 submissions in 2025 and is targeting 500,000 by 2030, with at least $4 billion in new business premiums. AIG achieved these volumes without adding underwriting headcount, a structural productivity gain that would be difficult to replicate through single-agent deployment.

The Three-Agent Taxonomy: Knowledge, Adviser, Critic

AIG’s orchestration architecture, as described on earnings calls and analyzed by AI News and Coverager, deploys three distinct agent types within the orchestration layer.

Knowledge assistants provide real-time information retrieval, pulling relevant data from AIG’s Palantir-powered ontology, policy databases, regulatory filings, and external sources. When an underwriter opens a submission, the knowledge assistant ensures comprehensive context is immediately available: loss history, comparable accounts, regulatory requirements for the jurisdiction, and relevant portfolio exposure data. This agent type replaces the manual research that historically consumed a significant portion of underwriter time.

Adviser agents generate insights based on historical case patterns. These agents surface comparable historical accounts when an underwriter reviews a new submission, identify patterns in loss experience for similar risks, and recommend pricing or terms based on what has worked or failed in the past. The adviser agent transforms institutional knowledge, which previously resided in the heads of senior underwriters, into a systematically accessible resource available to every member of the team.

Critic agents challenge recommendations and decisions before they reach the underwriter for final approval. This adversarial layer is the most architecturally significant from a governance perspective. Rather than confirming AI-generated suggestions, critic agents probe for weaknesses, flag potential concerns, and stress-test decisions. As we explored in our analysis of AIG’s full AI underwriting stack, AIG holds patents describing a “response validator” component that checks every LLM extraction against expected parameters and a chunk-level verification system that detects hallucinations by comparing cited sources against actual content. The critic agent concept operationalizes this technical capability at the workflow level.

The orchestration layer sits above these three agent types and determines when each activates, how they share information, and when human oversight is required. Zaffino emphasized that the 2026 focus is “orchestrating multiple agents in an orderly way to achieve scale across the enterprise,” spanning front-office underwriting, mid-office operations, and back-office functions. In partnership with Palantir and Anthropic, this effort extends the architecture AIG built for individual use cases into a coordinated enterprise system. The Q1 2026 earnings call also revealed that Claude (Anthropic’s model) “aligned with the adjusters 88% of the time” on claim assessments, providing one of the first quantitative AI-human agreement benchmarks from a major carrier.

Gen Re Maps the Reinsurance Orchestration Blueprint

AIG is not the only organization thinking architecturally about agent coordination. Gen Re’s December 2025 paper, “AI Agent Potential: How Orchestration and Contextual Foundations Can Reshape (Re)Insurance Workflows,” authored by Matthew Montero, provides the most detailed public blueprint for multi-agent orchestration in reinsurance.

The paper identifies two critical orchestration components. The first is observability: system transparency achieved through monitoring of logs, metrics, traces, and user feedback. The second is the information domain, a contextual repository of business-critical information, supporting processes, IT systems, and infrastructure that enables information flow between agents. Together, these components create a unified foundation connecting applications, data, and AI agents while maintaining human oversight at critical decision points.

Gen Re’s practical contribution is a redesigned quote-to-bind workflow for reinsurance that replaces the traditional linear process with a coordinated multi-agent system. The paper describes six specialized agents: an Orchestration agent that routes tasks and manages state, a Submission agent that handles intake, a Parsing agent that extracts structured data from heterogeneous documents, a Quote agent that generates pricing recommendations, a Binding agent that manages execution, and Support agents that handle auxiliary functions. Each agent has a defined scope, and human-in-the-loop approval is required at critical stages.

Gen Re CTO Frank Schmid framed the trajectory: “adoption will shift from enhancing existing tasks to enabling new ones” through redesigned workflows. The paper’s central argument, that workflow redesign delivers greater economic value than task-level improvement, directly parallels what AIG is demonstrating in practice. A carrier that deploys a single AI tool to speed up data extraction captures incremental efficiency. A carrier that redesigns the entire submission-to-bind workflow around coordinated agents captures structural competitive advantage.

For reinsurance actuaries, the Gen Re blueprint has direct implications for treaty placement and claims coordination workflows. The multi-agent architecture enables real-time exposure analysis, automated treaty compliance checking, and dynamic capacity deployment, capabilities that could fundamentally alter the speed and granularity of reinsurance transactions.

The Protocol Backbone: MCP and A2A

Multi-agent orchestration requires a common language for agents to communicate with tools and with each other. Two protocols are emerging as the standard plumbing for this architecture, both now governed by the Linux Foundation’s Agentic AI Foundation (AAIF), which was announced in December 2025.

Model Context Protocol (MCP), originally developed by Anthropic, handles vertical connectivity: the interface between AI agents and the tools and data sources they need to access. MCP adoption has been explosive. Monthly SDK downloads reached 97 million by March 2026, up from roughly 2 million at the November 2024 launch, a 4,750% increase in 16 months. More than 10,000 active MCP servers are deployed across platforms including ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and VS Code. Platinum members of the AAIF include Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

Agent-to-Agent protocol (A2A), originally developed by Google and contributed to the Linux Foundation in June 2025, handles horizontal connectivity: the interface between peer agents that need to collaborate. A2A uses HTTP, Server-Sent Events, and JSON-RPC 2.0 for transport, with “Agent Cards” that advertise each agent’s capabilities to other agents in the network. As of April 2026, over 150 organizations support A2A, including Google, Microsoft, AWS, Salesforce, SAP, ServiceNow, Workday, and IBM. The protocol is natively integrated into Azure AI Foundry, Amazon Bedrock AgentCore, and Google Cloud.

The two protocols are complementary, not competing. MCP connects agents to passive capability providers (databases, APIs, analytics tools) while A2A enables communication between autonomous agents that have their own reasoning capabilities. In insurance terms, MCP is what allows an underwriting agent to query Verisk’s loss cost data. A2A is what allows the knowledge assistant, adviser agent, and critic agent in AIG’s taxonomy to coordinate with each other. Together, they form the interoperability layer that makes multi-carrier, multi-vendor agent ecosystems feasible.

For carriers evaluating orchestration architectures, the open governance of both protocols under the Linux Foundation reduces vendor lock-in risk. An insurer building on MCP and A2A can, in principle, swap agent vendors or model providers without rebuilding the coordination infrastructure. This is a meaningful consideration given the dual-vendor AI stacks that carriers like Travelers and AIG are already deploying to hedge model concentration risk.

Verisk Ships Insurance-Specific MCP Connectors

The MCP protocol became more than an abstraction for insurance when Verisk launched two MCP connectors on May 5, 2026, embedding its analytics directly inside Anthropic’s Claude. The first connector, Verisk Underwriting Intelligence (ISO Indications), provides conversational access to loss cost trends, experience insights, and filing signals from Insurance Services Office. Verisk estimates this saves “hundreds of hours per carrier per year” in manual data retrieval. The second connector, Verisk XactRestore, embeds Xactware pricing and estimating intelligence for restoration professionals, saving an estimated “30 minutes to two hours per estimate.”

Verisk CEO Lee Shavel framed the launch in governance terms: “Trust is the foundation of insurance... Data must be authoritative, decisions must be explainable.” The connectors operate within Verisk’s established data governance framework with embedded security controls, and they respect existing customer entitlements, meaning carriers only access the data they have licensed.

The Verisk launch matters for the orchestration narrative because it demonstrates how the protocol standard translates into production insurance tooling. An MCP-enabled agent can now pull ISO loss costs, run Xactware estimates, and incorporate these data points into an underwriting recommendation without requiring manual analyst intervention at each step. As more insurance data providers ship MCP connectors, the agent’s access to authoritative data sources expands, and the orchestration layer becomes progressively more capable without requiring changes to the agents themselves.

Microsoft and Cognizant Target End-to-End Underwriting

The vendor ecosystem is building orchestration-ready platforms specifically for insurance. In February 2026, Microsoft and Cognizant announced a joint agentic AI offering targeting end-to-end underwriting and claims automation for insurers on Azure. The partnership, formalized in December 2025 with a multi-year expansion agreement, pairs Cognizant’s Agent Foundry (prebuilt tools and frameworks for insurance agent deployment) with Microsoft’s Azure AI Foundry infrastructure.

Cognizant’s “Underwriter of the Future” demonstration illustrates what orchestrated underwriting looks like from the vendor side. The system features dual AI assistants: “Sam,” which supports underwriters with risk assessment and coverage recommendations, and “Alex,” which assists brokers with premium comparisons and negotiation guidance. Behind these front-end assistants, a network of specialized agents handles specific functions: a Quotes Agent benchmarks market pricing, a Pricing Agent runs calculations across property type, jurisdiction, and security features, and a Supervisor Agent orchestrates the overall workflow. The system runs on Azure AI Foundry with Microsoft Fabric for data integration and Azure Cognitive Services for real-time voice transcription and speaker diarization.

Cognizant’s Neuro AI Multi-Agent Accelerator (neuro-san) serves as the core orchestration platform, and the company strengthened its Azure capabilities by completing the acquisition of 3Cloud, one of the largest independent Azure services providers, on January 1, 2026. The initial target is London Market commercial property insurance renewal processes, where the multi-party, multi-document nature of subscription market transactions makes orchestrated agents particularly valuable.

This vendor-built orchestration stack represents the alternative to AIG’s build-your-own approach. Mid-market carriers that lack AIG’s engineering resources and Palantir partnership can potentially deploy orchestrated agent workflows through platform vendors. The tradeoff is customization versus speed: AIG’s proprietary system is tailored to its specific ontology and underwriting guidelines, while the Microsoft-Cognizant offering provides a more generic framework that requires less internal engineering but offers less differentiation.

The 40% Cancellation Prediction: Governance as the Failure Mode

Gartner’s dual prediction captures the central tension in the orchestration wave. On one hand, the firm projects that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. On the other hand, Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. Senior Director Analyst Anushree Verma characterized most current agentic AI projects as “early-stage experiments” driven by hype rather than strategic planning.

The governance gap is quantifiable. Only 21% of organizations have a mature governance model for managing autonomous agents, according to Gartner. A separate Gartner poll of 3,412 webinar attendees found that just 19% reported significant agentic AI investments, while 31% were still waiting or unsure. The firm also flagged “agent washing,” estimating that only about 130 of the thousands of self-described agentic AI vendors are genuine, with widespread rebranding of chatbots, RPA bots, and basic AI assistants as “agentic.”

Insurance-specific data tells a similar story. Adoption is broad but maturity is shallow. Industry surveys show 22% of insurers plan to have an agentic AI solution in production by year-end 2026, with the market projected to grow from $5.76 billion to $7.26 billion (a 26% increase). But as the Grant Thornton AI proof gap survey found, only 24% of insurance executives could pass a governance audit within 90 days, even though 52% reported AI-driven revenue growth. The audit layer, not the technology, is where projects fail.

For orchestration specifically, governance complexity multiplies with the number of agents. A single-agent deployment requires one set of monitoring, validation, and escalation controls. A three-agent system with an orchestration layer requires controls for each agent, controls for the orchestration logic that determines agent activation and information routing, and controls for the interactions between agents. As we explored in our analysis of 30-hour autonomous agent cycles, the governance challenge grows nonlinearly as agent autonomy and coordination complexity increase. The Agent Charter framework offers one approach to managing this complexity, defining per-agent decision authority, approval thresholds, and human escalation triggers in auditable documents.

Why This Matters for Actuarial Practice

Multi-agent orchestration intersects with actuarial work at several critical points, and the intersection is becoming more direct as orchestration layers embed deeper into underwriting, pricing, and claims workflows.

Pricing signal routing. In an orchestrated system, pricing data flows through multiple agents before reaching the underwriter. The knowledge assistant retrieves loss cost trends, the adviser agent surfaces historical comparables, and the critic agent challenges the resulting recommendation. Each handoff introduces the potential for information loss, latency, or bias amplification. Pricing actuaries need to understand how the orchestration layer routes signals through the agent network, because the quality of pricing decisions depends not just on the data inputs or the individual models but on the coordination logic that determines which agent acts on which information at which stage.

Model validation scope expansion. ASOP No. 56 requires actuaries to understand the models on which their work relies. For a multi-agent orchestrated system, the “model” is no longer a single algorithm; it is a network of interacting agents with an orchestration layer governing their coordination. Validating this system requires assessing each agent’s individual performance, the orchestration logic that determines activation and information routing, the interaction effects between agents (does the critic agent consistently override adviser recommendations for certain risk classes?), and the human escalation triggers that define where automated decision-making ends. This is a fundamentally different validation exercise from testing a single GLM or gradient-boosted model.

Reserve implications of orchestrated underwriting. If orchestrated agent systems produce systematically different risk selection and pricing than traditional underwriting, the resulting books of business may exhibit different loss development patterns. Reserving actuaries should monitor whether portfolios underwritten through multi-agent orchestration show different frequency, severity, or development factor characteristics compared to traditionally underwritten business. AIG’s 88% AI-adjuster agreement rate on claims provides an early baseline, but the long-tail implications of orchestrated underwriting on reserve adequacy will take several accident years to emerge.

Regulatory compliance architecture. The NAIC’s evolving AI governance framework, including the agentic AI governance guidelines and the 12-state evaluation tool pilot, will need to address multi-agent systems explicitly. A rate filing that relies on AI-generated pricing signals must demonstrate governance over the full agent pipeline, not just the final pricing model. The orchestration layer’s audit trail, which tracks which agents were activated, what information they shared, and where human oversight intervened, becomes essential documentation for regulatory submissions. Carriers that build this auditability into the orchestration architecture from the start, as AIG has done with its Palantir-powered ontology, will have a compliance advantage over those that attempt to retrofit governance onto existing agent deployments.

Build versus buy at the orchestration level. The emerging split between AIG’s proprietary orchestration stack and the Microsoft-Cognizant platform offering creates a strategic question for carrier actuaries and CIOs. Building custom orchestration, as AIG has done with Palantir, provides maximum control over agent behavior and deep integration with the carrier’s specific underwriting philosophy. Buying platform orchestration provides faster deployment but less differentiation. For actuaries advising on AI strategy, the key question is whether the carrier’s competitive advantage comes from proprietary risk selection logic (favoring custom orchestration) or from speed to market (favoring platform adoption). The answer likely varies by line of business and carrier scale, which is precisely the kind of analysis actuarial teams are well positioned to provide.