When agentic AI executes a multi-step actuarial pricing workflow, from exposure ingestion through triangle development, loss development factor selection and final rate indication, a flawed assumption at step one propagates through every downstream node with no automatic correction.

Only 23% of organizations have scaled agentic systems within a single function, which means most agentic pricing pipelines now running are operating without a quality-control baseline. The unit of validation and the unit of risk have come apart.

Key Takeaways

  • Five agents, one signature. An exposure-base error at ingestion reaches the filing as a fluent narrative about "continued adverse development", with each agent behaving correctly on the inputs it received.
  • 85.1% to 86.5% agentic decision accuracy against a 73.4% to 77.6% single-LLM baseline across 635 synthetic small-commercial applications, with the actuarially useful finding in the failure-mode decomposition rather than the headline.
  • 66.1% versus a 70.1% baseline. Naive retrieval-augmented generation underperformed a plain single model on multi-step cases. More source data without structured intermediate reconciliation made the system worse.
  • 58% of eligible life applications auto-approved within two minutes at Manulife, on an engine architecturally unable to deny coverage. Every negative outcome routes to a human by design rather than by discretion.

How One Bad Assumption Becomes a Filed Rate

Trace a commercial auto pricing pipeline end to end. A data-ingestion agent pulls exposure counts and claims history from a policy administration system. A second builds and ages loss triangles. A third selects loss development factors from the triangle patterns. A fourth blends those into an a-priori loss cost. A fifth drafts the rate indication narrative accompanying the filing.

Suppose the ingestion agent misreads a schema change and pulls earned car-years instead of written car-years for a subset of policies. The exposure base is wrong at source. The triangle agent does not know that; it ages the resulting loss ratios, which now run systematically high because the denominator understated exposure. The LDF-selection agent, seeing elevated and volatile development, may reasonably select higher tail factors to compensate.

By the time the indication agent drafts language describing continued adverse development in the most recent accident years, the narrative is internally consistent, well written, and explaining an artifact rather than a loss trend. Each agent behaved correctly given its inputs.

That is the distinction the SOA's Emerging Topics Newsletter draws between agentic AI and the two prior waves of actuarial automation. Rule-bound batch scripts and pattern-recognition machine learning both operate on a defined, checked input, while agentic systems "perceive, plan, act, and learn autonomously to achieve defined goals" across a chain where no agent owns the correctness of what came before it.

Collapsing that into a single category of AI risk runs together three failure modes with different mitigations.

Risk surface Failure mode Where it originates
Data ingestion error Wrong exposure counts, stale triangles, or a mismatched schema silently feed every downstream agent The connection between source systems and the pipeline's first agent
Orchestration error The wrong agent chain is invoked for the product or peril, so a workers' compensation logic path processes a commercial auto submission, or a state-specific rule is skipped The routing and control logic that decides which agents run and in what order
Output formatting error The narrative-generation agent describes the quantitative indication inaccurately, smoothing over a discontinuity or mischaracterizing a driver, even though the underlying numbers are correct The final agent translating structured output into filing language

Data ingestion error is the familiar one, a descendant of data-quality problems pricing teams have always managed, now propagated at machine speed with no human reading each intermediate output. Orchestration error is a routing failure rather than a data or model failure, invisible to any test checking whether each agent performed correctly on the inputs it happened to receive.

Output formatting error carries the professional-liability exposure. The quantitative indication may be sound while the generated narrative overstates confidence or omits a caveat, and a regulator reading the filing cannot tell it was machine-drafted.

Node Validation Against Pipeline Risk

Actuarial model validation is built around a node-by-node discipline: back-test the LDF-selection logic against historical triangles, check GLM relativities for stability across refits, confirm the rating algorithm reproduces the filed table. Every one of those checks can pass while the pipeline as a whole produces a distorted indication, because the checks are scoped to nodes and the error lives in what connects them.

A July 2026 preprint from Brigham Young University researchers makes the adjacent point empirically. Across 635 synthetic small-commercial applications, an agentic retrieval-and-reasoning pipeline reached 85.1% to 86.5% overall decision accuracy against a 73.4% to 77.6% single-LLM baseline. The more useful finding sits in the decomposed failure modes.

On cases involving irrecoverable missing information, the agentic pipeline recognized and escalated 84.3% against 56.7% for the single-LLM approach. On multi-step reasoning, where a conclusion depended on combining facts across documents, orchestration scored 78.0% against 70.1%, a gap the researchers attribute to holding intermediate state across steps rather than collapsing reasoning into one pass.

The result that matters most for pricing runs the other way. Naive retrieval-augmented generation, meaning a model with document search bolted on but no orchestration layer forcing structured intermediate checks, scored 66.1% against the 70.1% plain baseline on multi-step cases. Access to more source data without a structure reconciling it at each step made the system worse. Adding an agent does not reduce compounding error and can increase it.

The pricing consequence is multiplicative rather than additive. An indication distorted by an exposure-ingestion error does not misstate one number; it distorts every relativity and trend factor derived from the corrupted base, because pricing models build on top of the aggregate indication. The reserving parallel sits in agentic claims triage feeding IBNR: an orchestration error routing a subrogation-eligible claim through the wrong severity model distorts a reserve estimate that a review scoped to the reserving model, rather than the intake pipeline upstream, will not surface.

The Checkpoint Sits at the Wrong End

Agentic pipelines are frequently designed to present a complete filing draft for review rather than surfacing each intermediate output as it is produced, and that design choice is where the governance problem actually lives.

The reasoning behind it is sound. Five separate intermediate approvals slow the workflow and can train reviewers to rubber-stamp routine steps. But it also compresses the actuary's effective review to a single moment, at the end, when the narrative is fluent and the numbers reconcile internally. Catching an exposure-base error there requires independently re-deriving a chain the pipeline has already smoothed into a coherent story, which is much harder than reviewing an intermediate triangle that looks visibly off.

Two live deployments show the design space. Manulife's MAUDE engine auto-approves more than 58% of eligible advisor-submitted life applications within two minutes but is architecturally restricted to approvals, so every negative outcome reaches a human by design. It has captured roughly $300 million of a $1 billion AI value target set for 2025 through 2027, using smaller purpose-built models that keep individual agent scope narrow enough to validate.

Sixfold's AI Underwriter, launched in June 2026 across six carriers representing $270 billion in gross written premium after a $30 million Series B, offers a choice between an augmentation configuration preserving a documented human decision point at each stage and a straight-through automation one.

Guidance has not caught up to the pipeline as the unit of risk. The NAIC Model Bulletin requires carriers to designate a person responsible for each AI system and document validation and testing sufficient to answer a regulatory inquiry, a structure assuming one model, one owner, one accountability chain.

Twelve states were piloting the AI Systems Evaluation Tool as of March 2026, with adoption expected at the Fall National Meeting, and its exhibit structure was built for single-model documentation. The preprint's own architectural recommendation, that missing information requires an explicit escalation pathway, is narrower and more testable than maintaining human oversight. Neither the SOA newsletter nor the bulletin has translated it into tolerance bands for a pricing pipeline. Our earlier work on the NAIC's governance gap tracked the same absence across underwriting and claims.

Further Reading

Sources