Start with the cost of completing a task to an agreed quality standard. Include model usage, software, integration, and the time people spend checking or correcting the output.
The examples below come from insurance workflows. They are individual reported results, not a cost model for your company. For the workflows and source context, see Insurance & specialty.
1 · Token economics: where the money actually goes
Measure four drivers of model usage, alongside the costs of running and supervising the workflow:
- Context per call. An agent that reads a 40-page submission packet per step spends those input tokens on every step. Send the material each step needs and measure the effect of caching or retrieval on cost and quality.
- Calls per task. A five-stage pipeline makes five-plus calls; a retry loop makes more. Anthropic reports about 15x the token usage of chat interactions for its multi-agent systems (Anthropic's multi-agent research write-up); measure the additional calls and coordination in your own design.
- Model tier per step. Compare candidate models on representative tasks. Choose the least expensive option that meets the quality threshold, and recheck its rate and service terms at procurement.
- Rework loops. Failed output is regenerated, and regeneration is spend. Evaluation gates stop repeated low-quality attempts before they reach a person.
Build a budget from the work. Obtain current quotes for seats, usage, vendor software, and support. Add implementation, security review, evaluation, training, and ongoing ownership. The roadmap's budget ranges are planning assumptions; a vendor case study or industry average cannot establish your required spend.
Measure cost per accepted task and net staff time released, including review and rework. Treat released hours as capacity until Finance can identify avoided expenditure or additional profitable work. Higher gross written premium is not itself a profit measure.
What sets model cost per completed task
Add the cost of every model call, including retries. Token counts and rates can differ at each step; these four factors help explain the total.
For each call: input tokens × applicable input rate + output tokens × output rate. Sum across calls, splitting token categories when rates differ. Add tools, hosting, and human review to measure whole-task cost.
Anthropic (June 2025) reports about 15× as many tokens as chat interactions for its multi-agent systems. That is token usage in its own data, not a universal cost multiplier.
A decision sheet for the pilot
| Record before testing | Report at the decision gate |
|---|---|
| Task volume, handling time, quality standard, and current cost | Completed tasks, end-to-end time including review, and material errors that escaped review |
| One-time implementation budget and recurring cost limit | Actual implementation spend and monthly operating cost, including human support |
| How released capacity would be used | Hours returned, cash expenditure avoided, and incremental contribution separately; avoid double-counting |
| Named owner, review date, and stop conditions | Scale, revise, or stop, with the evidence supporting that decision |
Simple calculation: net hours returned = completed tasks × (baseline minutes − pilot minutes, including review) ÷ 60. Test a downside case with lower volume, more exceptions, and higher service costs before approving expansion.
Explore the tradeoff · illustrative assumptions
A faster draft is only part of the time saved
Compare average staff effort per completed task at the same quality standard. These starting values are invented; replace them with measured pilot results. Count each person's time once, including unsuccessful attempts.
The example below uses fixed assumptions. Enable JavaScript to change them.
- Preparation
- Review
- Rework
Enter a whole task count between 1 and 100,000 to calculate staff hours.
16.7staff hours returnedacross 100 completed tasks
10 minutes released per task
100 × (30 − 20) ÷ 60 = 16.7 hours
This measures potential staff capacity, not elapsed turnaround time, cash savings, or ROI. Software, model usage, implementation, and other operating costs belong in the business case separately. An apparent time gain does not justify lower quality.
2 · Plan for production
| 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 | Training restrictions, approved retention settings, audit access, and processing locations | Use approved tools and data under reviewed agreements. See 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 | Service continuity, financial capacity, and support commitments may affect a critical workflow | 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 · A suggested order for evaluating use cases
This is an editorial starting order, based on workflow fit and reviewability. It is not a calculated value-to-risk score. Consider the carrier and vendor examples behind the integration phases (verbatim source passages for the top entries are on the evidence page).
| # | Use case | Reported example or pilot rationale | Where it sits |
|---|---|---|---|
| 1 | Submission intake and triage | AIG reports 370,000+ Lexington submissions processed in 2025; Cytora's 2023 Markel case reports 113% higher GWP per full-time equivalent (evidence) | Underwriting; candidate pilot with human review |
| 2 | Document extraction and summarization | Loss runs, schedules of values, and claims files provide repeatable extraction tasks. Measure field accuracy and net handling time; no general savings range is assumed | Across functions; control level depends on data and downstream use |
| 3 | Claims triage, severity and litigation prediction | CLARA reports average indemnity for attorney-involved casualty claims is about 4.9 times as much as for unrepresented claims (vendor data). This association does not establish savings caused by AI triage | Claims; decision support with human authority |
| 4 | Bordereaux processing | 85–94% processing time savings reported by Verodat; validate exception handling and reconciliation in your own bordereaux | 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) | Test whether staff find current, cited guidance faster. Track unsupported answers, stale material, and permission failures | 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 | Test detection of duplicate payments, missed recovery opportunities, and policy exceptions on reviewed historical files. Measure confirmed recoveries and false positives before attributing savings | Claims finance; Phase 3 material |
| 9 | Bounded agentic quoting | Hiscox reported 3 days to ~3 minutes in a 2023 terrorism-risk quote proof of concept; CFC reports an agentic pilot (evidence). These are not general production averages | Underwriting; last, under Phase 3 gates |
The same nine, grouped by the proposed operating model
Suggested sequence, grouped by how the output is used. Numbers refer to the editorial priority order in the table; they are not measured value or risk scores. All groups need governance, with controls matched to the decision.
Start here · assistive
AI drafts, a person decides. Match review to the sensitivity of the data and the effect of the output.
1Submission intake and triage 2Document extraction and summarization 6Knowledge access over guidelines 7Actuarial filing research and pricing workbenchThen · decision support
The model scores or recommends; a named human holds authority.
3Claims triage, severity and litigation prediction 4Bordereaux processing 5Fraud detectionLast · bounded automation
Gated work: bounded segments only, behind Phase 3 controls.
8Leakage and subrogation 9Bounded agentic quotingThese groups describe the proposed use of each workflow. A leakage report reviewed by a person can start as an assistive pilot; a system that blocks payments needs stronger controls. Determine the requirements from what the system actually does.
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 phase logic of the integration phases applied to a single function; the harness section of the practice ladder is the build manual for each step, and what each stage must have in place before it starts is on Governance, next.