Insurance AI Gateway Cost per Case | Moring
Cost per case measures what an insurer spends on AI for a single claim, application, or policy action, computed by attributing every model call and tool invocation for that case through the gateway. Because a claim triage can involve a dozen model calls across rented and owned models, per-case attribution is the only way to connect token economics to claims economics and price automation honestly.
Short answer
Cost per case measures what an insurer spends on AI for a single claim, application, or policy action, computed by attributing every model call and tool invocation for that case through the gateway. Because a claim triage can involve a dozen model calls across rented and owned models, per-case attribution is the only way to connect token economics to claims economics and price automation honestly.
How it works
Cost per case is the unit metric that connects AI spending to insurance operations. Total monthly spend answers finance; cost per answered claim, cost per processed first notice of loss, or cost per underwriting submission answers the operating question of whether automation pays. Computing it requires attribution at the level of the business object, and that is what an AI gateway provides, because every model call and tool invocation for a case passes through one control point.
The arithmetic is less obvious than it looks. An LLM request does not map to a predictable unit of work: cost depends on the selected model, prompt size, generated output, retries, tool calls, and provider-specific token accounting. A single claim triage can touch a dozen model calls across frontier models rented by the call and models the insurer hosts inside its own perimeter. Without a gateway joining those calls to a case identifier, per-case cost is assembled by hand from incompatible logs.
Case-level attribution works when the gateway records carry the case reference. Agents handling intake, document extraction, triage, fraud screening, and settlement drafting pass the claim or application identifier with each request, and the gateway logs identity, model, tokens, and cost against it. Aggregation then produces honest unit metrics, including where in the workflow cost concentrates and which cases consume disproportionate resources.
Per-case cost exposes decisions that aggregate spend hides. Document extraction on low-quality photos may cost multiples of the average, suggesting either a routing rule to better-suited models or a human path for poor inputs. Fraud triage with deep investigation may justify its cost per case explicitly, while routine status inquiries should cost nearly nothing, which argues for small models or cached responses on that path. Semantic caching in the gateway removes repeated identical calls from the bill entirely.
Unit economics also discipline vendor and model selection. When a carrier can state what a claim currently costs to process with AI, negotiated model pricing, routing changes, and caching gains become measurable rather than anecdotal. Procurement conversations change character: the discussion is about observed cost per case under real claims traffic, not about list prices per token that nobody can connect to outcomes.
There is a governance dividend as well. The NAIC AI Model Bulletin, adopted in roughly half of US states, expects carriers to govern their AI estate, and a workflow whose per-case cost is known and owned is by definition inventoried and supervised. Cost per case therefore functions simultaneously as a management metric, a procurement instrument, and evidence that the AI estate is under active control.
Frequently asked questions
Why is cost per case the right metric for insurance AI spending?
Because insurance operations are counted in cases, not tokens. Executives approve automation based on cost per processed claim or per underwriting submission, and aggregate monthly spend cannot support that decision. A gateway attributes every model call and tool invocation to the case it belongs to, producing a unit metric that connects AI consumption to claims economics and reveals where in the workflow cost actually concentrates.
How does a gateway compute cost per case for a single claim?
Agents pass the claim identifier with each request, and the gateway logs identity, model, token counts, and computed cost for every call, then aggregates against the case reference. This matters because one claim triage can touch a dozen model calls across rented frontier models and self-hosted models, with cost varying by prompt size, output length, retries, and tool calls. Hand-assembly from provider invoices cannot reconstruct that picture reliably.
What insurance-specific decisions does per-case cost data improve?
Routing and staffing decisions surface first: low-quality claim photos that drive extraction costs upward may need a different model or a human path, while routine status inquiries should shift to small or cached responses. Fraud investigation cost per case can be justified explicitly against recovered losses. Procurement improves too, because model negotiations can reference observed cost per case under real claims traffic instead of abstract token pricing.
Does cost per case have any governance value for an insurer?
Yes, as evidence of an owned estate. The NAIC AI Model Bulletin, adopted in roughly half of US states, expects carriers to govern AI resources, and a workflow with a known, attributed per-case cost is demonstrably inventoried, budgeted, and supervised. Unit metrics also make anomalies visible, since a sudden rise in cost per claim points to a specific agent or model change that governance can investigate immediately.