Tokens are the unit of cost, and agents spend them like a fleet, not a person. Every call pays for context in and text out; workflows make many calls per task, and fan-out multiplies them. Engineer cost with the same discipline as quality.
Why specialty underwriting has unusually high leverage
Personal lines and standardized small-commercial products increasingly use structured data, rules, and straight-through processing. Specialty and other large-complex accounts remain bespoke, document-heavy, and expert-driven. Manuscript wording, layered exposures, broker negotiation, loss runs, engineering reports, and exposure schedules require account-level judgment (Capco).
That means a specialty underwriter generally spends far more expert time per account. Even before risk judgment begins, a submission may wait hours or days while emails and attachments are read, keyed, validated, and made decision-ready. Hiscox described manual extraction for one London Market workflow as taking up to three days (Hiscox); Arch reported a two-to-three-day intake-to-decision cycle before automating submission intake (Ivans/Arch).
| Personal / standardized small commercial | Specialty / large-complex accounts | |
|---|---|---|
| Account shape | High volume, repeatable questions, structured data, filed products | Lower volume, bespoke terms, layered or unusual exposures, unstructured broker packs |
| Human role | Algorithms handle routine cases; people manage exceptions | Experienced underwriters interpret risk, negotiate terms, and retain pricing/bind authority |
| Best automation target | Straight-through decisions on simple risks | Make the account decision-ready: ingest, extract, enrich, check appetite, surface precedents, draft the memo |
| Why agents pay | Reduce transaction cost at scale | Return scarce expert hours on every account and let the same team evaluate more of the submission flow |
This is why triage agents and multi-agent underwriting stacks have disproportionate value in specialty. A triage agent prevents expert time from being spent on incomplete or out-of-appetite submissions. A multi-agent stack can run independent work in parallel: one worker extracts exposures, another checks guidelines, another enriches external data, another compares pricing and portfolio constraints, and an orchestrator assembles the decision-ready view. The underwriter still makes the consequential judgment, but no longer spends the first part of every account moving information between systems.
The measured results reflect that leverage: AIG reports 2–5x faster underwriting; Markel reports a 113% productivity uplift; Arch cut intake-to-decisioning by 70%; and Hiscox reduced one bounded specialty quote workflow from three days to three minutes. These are carrier- or vendor-reported results, not universal benchmarks; the source passages are collected on the evidence page.
1 · Token economics: where the money actually goes
Four drivers set the cost of an agentic workflow, and none of them is the model's sticker price:
- Context per call. An agent that reads a 40-page submission packet per step spends those input tokens on every step. Context discipline (send what the step needs, not the file room) is the biggest lever most teams miss.
- Calls per task. A five-stage pipeline makes five-plus calls; a retry loop makes more. Orchestrated agent teams have been measured at roughly 15x the token cost of a single agent (Anthropic's multi-agent research write-up); fan-out multiplies spend roughly linearly.
- Model tier per step. Frontier models cost several times more than mid-tier ones. Extraction and classification rarely need the frontier; judgment steps sometimes do. Route by step, not by habit.
- Rework loops. Failed output is regenerated, and regeneration is spend. Evaluation gates stop repeated low-quality attempts before they reach a person.
Budget anchors. Mid-size enterprise LLM agreements run roughly $250k to $1M+ a year; volume discounts are material but must be verified in procurement (benchmark guide). Insurer IT spend averages ~4.5% of GWP (Datos Insights), and two-thirds of insurance CEOs plan to allocate 10–20% of budget to AI (KPMG CEO Outlook, PDF). Vendor purchases succeed about 67% of the time versus 33% for internal builds (MIT, PDF), supporting a buy-commodity, build-differentiation strategy.
Measure cost per completed task and dollars per underwriter hour returned, not price per token. A use case that cannot clear both bars is a demo. The winning comparison is machine cost per submission versus loaded human cost per submission reviewed, and it usually is not close.
2 · The operational realities the demo never shows
| Consideration | What it looks like in production | The practical rule |
|---|---|---|
| Latency and broker SLAs | Multi-step agents take seconds to minutes; brokers notice turnaround, not your architecture | Put speed where the broker sees it (acknowledgment, triage, appetite answer) and depth behind it |
| Enterprise terms | Zero data retention and no-training clauses, audit logging, regional processing | No enterprise agreement, no company data. This is the first control in the governance section |
| Model deprecation and drift | Vendors retire and upgrade models; behavior shifts silently under a pipeline that passed its evals in March | Re-run the golden dataset on every version change, and pin versions where the provider allows it |
| Rate limits and surge | A CAT event is a volume spike; provider rate limits are a hard ceiling | Capacity-plan for surge events, and keep a queue-and-degrade path that keeps humans working when the cap is hit |
| PII and residency | Claimant and policyholder data in prompts and logs; state and partner rules on where it may be processed | Minimize and de-identify by default; log retention is a compliance surface, not an IT detail |
| Vendor viability | 95% of H1 2026 insurtech funding went to AI startups (funding data), which means consolidation is coming | Due diligence on funding and runway; an exit plan for any vendor whose output feeds a regulated decision |
| Lock-in and portability | A harness written to one provider's quirks is a migration project later; both core vendors shipped agentic frameworks in 2026 (Guidewire, Duck Creek) | Keep context specs, evals, and hooks model-agnostic; they are your portable assets |
3 · AI agents for pricing and costing model development
The right mental model is an agent-supported model-development lifecycle, not an autonomous pricing actuary. The predictive model can still be classical and deterministic. Agents accelerate the work around it: gathering evidence, producing reproducible code, running independent challengers, testing, documenting, and monitoring.
Keep costing separate from transaction pricing. The costing layer estimates expected loss and risk costs. The pricing/rater layer then adds expenses, commission, reinsurance, capital and profit provisions, rules, and permitted underwriting judgment. An agent may trace and reconcile those components; it should not silently mix them or choose assumptions.
| Agent role | Work it can accelerate | Control that remains human |
|---|---|---|
| Data steward | Profile data, reconcile premium/exposure/loss totals, identify missingness, leakage, duplicates, and policy-year contamination; generate repeatable QA artifacts | Data suitability, limitations, on-leveling, trend periods, large-loss and catastrophe treatment |
| Experience analyst | Refresh experience studies, loss-cost trends, mix shifts, retention/conversion, expense and commission summaries, benchmark research | Selection of assumptions, credibility, external benchmarks, and actuarial judgment |
| Model workers | Build independent GLM, GBM residual, frequency/severity, current-plan relativity, and scenario challengers from one approved specification | Candidate methods, permitted variables, constraints, and interpretation of signal |
| Validation worker | Run one frozen out-of-time holdout; compare deviance, calibration, stability, segment impacts, residuals, and sensitivity; reconcile predictions to dollars | Independent review, fairness/proxy assessment, materiality, and model selection |
| Rater engineer | Translate approved factors into code, schemas, parameter tables, UI inputs, test cases, and parallel-run comparisons against the current rater | Rate level, relativities, implementation approval, and production release |
| Documentation worker | Draft model documentation, filing checklists, change logs, test evidence, assumption inventories, monitoring packs, and underwriter guidance | Actuarial communication, regulatory representations, sign-off, and monitoring response |
The useful multi-agent pattern
Freeze the data snapshot, holdout, metric definitions, and model charter first. Then fan out independent workers by model formulation or task. One agent builds the baseline GLM, one tests a GBM residual, one builds frequency/severity, one researches peer filings, and one validates every output. The orchestrator compares artifacts; it does not average recommendations, and no builder grades its own model.
This pattern compresses iteration while preserving effective challenge. Deloitte describes agentic pricing analysis that reverse-engineers complex regulatory manuals and reduces competitor analysis from weeks toward minutes (Deloitte). Aviva is testing an Actuarial Agent to enhance pricing tools with new insights and optimize existing model code in Global, Corporate & Specialty business (Aviva/hx). Hyperexponential demonstrates agents generating pricing schemas, rating logic, interfaces, parameter tables, profiling suggestions, and documentation (hx product demonstration; vendor-reported).
Non-negotiable controls
- Agents call code; they do not perform actuarial arithmetic in prose. Every result must be reproducible from versioned code and data.
- One approved model charter. Target, exposure, offsets, weights, time split, loss treatment, constraints, and success criteria are fixed before fan-out.
- One frozen holdout. All challengers are compared on identical future-period data and converted to the same loss-dollar basis.
- No autonomous assumption or rate changes. Trend, credibility, expense, reinsurance, profit, and judgment selections require named actuarial approval.
- Independent validation. A separate reviewer challenges data lineage, leakage, variable justification, stability, fairness, and implementation.
- Production is gated. The agent may draft rater code and tests; deployment requires version control, approvals, parallel-run reconciliation, and rollback.
Thin specialty data raises the value of workflow automation but lowers the credibility of unconstrained model search. Agents should help the actuary test what the current plan misses, document uncertainty, and maintain repeatable diagnostics; they should not manufacture confidence from sparse claims. The SOA's 2026 agentic-workflow research scope includes rate development, assumption development and benchmarking, model governance, documentation maintenance, and regulatory reporting, with explicit attention to reliability, explainability, bias, and human oversight (SOA Research Institute).
4 · The most impactful use cases for insurance companies, ranked
Ranked by measured value divided by implementation risk, from the carrier and vendor results behind the roadmap (verbatim source passages for the top entries are on the evidence page).
| # | Use case | Why it pays (measured) | Where it sits |
|---|---|---|---|
| 1 | Submission intake and triage | 2–5x underwriting speed and 370k+ submissions a year at AIG; Markel's 113% productivity uplift; 50–97% faster processing and +15% hit ratios at Sixfold customers (evidence) | Underwriting; the proven first move |
| 2 | Document extraction and summarization | Loss runs, SOVs, claims files: routine 50–80% time cuts on document-heavy work; the foundation every other use case reads from | Everywhere; assistive, low scrutiny |
| 3 | Claims triage, severity and litigation prediction | Attorney-involved claims cost ~4.9x more (CLARA data); early triage moves both cycle time and indemnity | Claims; decision support with human authority |
| 4 | Bordereaux processing | 85–94% processing time savings (Verodat); the unglamorous pain point of program business | Program/delegated-authority operations |
| 5 | Fraud detection | 5x more fraud detected at Tokio Marine (case study, vendor-reported) | Claims; scoring with human review |
| 6 | Knowledge access (RAG over guidelines) | Appetite and procedure answers in seconds; multiplies every other use case by keeping context current | Enterprise-wide; internal only |
| 7 | Actuarial filing research and pricing workbench | Filing research from weeks to hours (Akur8); 13 pricing tools in 13 weeks at Allianz Commercial (hx) | Actuarial/pricing; assistive |
| 8 | Leakage and subrogation | AI pre-payment controls prevent 90–95% of detectable leakage (analysis); $15–20B of subrogation goes uncollected annually (industry estimate) | Claims finance; Phase 3 material |
| 9 | Bounded agentic quoting | 3 days to ~3 minutes at Hiscox London Market; CFC's agentic pilot (evidence). Real, but gated: human authority, governance gates, bounded segments only | Underwriting; last, under Phase 3 gates |
Sequencing rule: start assistive (ranks 1–2), move to decision support with human authority (3–5), and reach bounded automation last (9). That is the roadmap's phase logic applied to a single function; the harness section of the practice ladder is the build manual for each step.