Starting Point: Your Pricing Reality

• Tell me about your current pricing process for the specific line you want to improve, from data ingestion to filing.
• How many distinct rate plans or rating cells do you maintain for this line today? Options: Fewer than 10, 10–50, 51–200, 201–1,000, More than 1,000, Unsure
• Name the systems that hold your master rate tables, policy records, and claims history. Options: Primary rating engine, Policy administration system, Enterprise data warehouse, Claims system, Actuarial toolset, Other (please describe)
• Who on your team prepares pricing analyses and who signs rate change approvals? Options: Chief Actuary, Head of Pricing, CUO, SVP Underwriting, CIO/CTO, Other (please list)
• Walk me through the last rate change you executed, from hypothesis to filing to market outcome.
• If you had to shorten analysis-to-filing time, which missing control or dependency would make you stop the effort?

Where Rates Are Quietly Losing Margin

• Where does your current model most often fail to separate high- and low-risk accounts in ways that matter for pricing?
• Point to the top three customer or risk segments where predictions and actual outcomes diverge, and explain why for each.
• Estimate the annual premium leakage you attribute to under-segmentation or mispriced segments. Options: Less than 1% of premium, 1%–3% of premium, 4%–7% of premium, More than 7% of premium, Do not know
• Provide an example of a recent pricing decision that produced unexpected adverse selection, and describe the sequence of events.
• What single modeling failure would cause you to halt a rollout immediately? Options: Severe calibration drift, Clear adverse selection in a pilot, Regulatory rejection risk, Data integrity failure, Other (please specify)

The Other Paths You're Considering

• Choose the option you are most likely to follow next: keep the incumbent, build internally, engage another partner, or stay with no change. Options: Keep incumbent vendor, Build internally, Hire another third party, Adopt open-source/tools internally, No change
• Outline the alternatives you have evaluated so far and the stage each is at in your decision process.
• Under what conditions would you remain with your current approach instead of switching?
• Has anyone inside your company proposed solving this using only internal resources? If so, who and what is the proposed scope? Options: Yes, small internal team, Yes, large internal program, No internal proposal, Unsure
• If one of the alternatives is chosen, would that stop you from continuing conversations with external partners right now? Options: Yes, we'll stop immediately, No, we'll continue evaluating, Depends on technical fit, Unsure

Stakeholders, Politics, and Decision Triggers

• Who inside executive leadership is most likely to push back on model-driven rate changes, and why?
• List the stakeholders who must approve model logic, rate moves, and filings, and their typical concerns.
• Describe how underwriters and pricing teams currently consume or override model outputs in new business and renewal workflows.
• When a regulator or board questions a rate change, who typically becomes the primary responder and what evidence do they expect?
• If a key stakeholder refuses to endorse the pilot, can the project continue without them? Options: Yes, with a delegated approver, Yes, with mitigation steps, No, their signoff is required, Unsure

Data Readiness: Can We Build on Your Foundation?

• How ready is your historical exposure, policy, and claims data to support frequency and severity modeling without major reconstruction? Options: Ready with minimal cleaning, Requires moderate ETL and linkage work, Requires major reconstruction, Not ready
• Which transaction-level fields do you currently have consistently available: policy effective dates, endorsements, exposures, claim severities, loss development identifiers, or other? Options: Policy effective dates, Endorsements, Exposure units, Claim severities, Loss development identifiers, Other (specify)
• How many policy years of clean, linked data do you have that include both exposures and closed claims? Options: Under 2 years, 2–4 years, 5–7 years, 8+ years, Unsure
• Identify the person or role who holds ETL access and who can approve extracts for an external modeling engagement.
• Would a timeline longer than 8 weeks for delivering a production-quality extract stop the engagement? Options: Yes, we need <8 weeks, No, we can wait 8–16 weeks, No, we can accept longer timelines, Unsure

Operational Constraints and Integrations

• Which integration dependency is most likely to delay deployment: API availability, rating engine mapping, or hosting and model scoring format? Options: API availability, Rating engine mapping, Model hosting/scoring format, Data handoff processes, Other (explain)
• List the internal owners for API, rating engine, and infrastructure work and indicate their typical response time for technical requests.
• Describe the model scoring formats your environment accepts today, for example model ID plus score, inline probabilities, or score bands. Options: Model ID + score, Inline probability vector, Score bands/buckets, Feature-level scores, Other (specify)
• How quickly can your engineering team implement and test a new model mapping in a staging environment? Options: Less than 2 weeks, 2–4 weeks, 1–2 months, Longer than 2 months, Unsure
• Would the inability to accept an external score feed force a stop to the project, or do you have a feasible workaround? Options: Stop the project, Feasible workaround available, Temporary manual process possible, Unsure

Regulatory and Compliance Gateways

• What regulatory objection from a state filing would be most likely to block approval for a model-driven rate change?
• Provide the states or jurisdictions where prior approval is required for this line.
• Have you had a rate filing returned or rejected in the last 36 months? If so, describe the reason. Options: Yes, technical documentation, Yes, actuarial methodology, Yes, public policy/outcry, No, Unsure
• Identify the role responsible for preparing actuarial narratives and regulator responses, and confirm their availability.
• Would a regulator's demand to simplify the model enough to materially reduce lift cause you to pause the project? Options: Yes, pause required, No, accept simplified model, We would negotiate alternatives, Unsure

Success Metrics That Move the Needle

• Name the one pilot outcome that would make you sign a production agreement within 30 days.
• Select the KPIs you will use to accept the model, in order of importance. Options: Relative uplift in pure premium, Improved Gini / rank ordering, Calibration by decile, Reduction in adverse selection, Regulatory explainability score, Time to file / deployment speed
• Estimate the minimum relative uplift or calibration improvement you require to accept the model. Options: 0%–2%, 3%–5%, 6%–10%, More than 10%, Will decide after baseline analysis
• Provide the approval chain for pilot acceptance, by role and title.
• Would underwriter refusal to operationalize the pilot stop you from accepting the results even if the metrics are met? Options: Yes, acceptance requires operationalization, No, accept results and plan rollout, Accept with phased rollout and mitigations, Unsure

Pilot Design: Test, Measure, and Protect

• Choose the pilot design you believe will surface adverse selection fastest: holdout backtest, live geographic A/B, policy-level score caps, or another approach. Options: Holdout backtest, Live geographic A/B, Policy-level score caps, Hybrid (backtest + live), Other (describe)
• Estimate the monthly policy volume required for a meaningful pilot to reach statistical confidence. Options: Under 1,000 policies/month, 1,000–5,000, 5,000–25,000, More than 25,000, Unsure
• State the number of months of out-of-time holdout you expect for backtesting and why. Options: 6 months, 12 months, 24 months, Other (specify)
• Provide the name or role of the pilot decision owner and list the artifacts they will require at pilot close.
• Specify the numeric adverse-selection threshold that would require pausing the project, and how you would measure it.

Commercial Signals and Next Steps

• Outline the procurement or budget step that must happen to enable signature within 30 days if the pilot proves the numbers.
• Name the role that signs the SOW and the role that countersigns the data sharing agreement.
• Do you have a standard legal template for data sharing and how long does negotiation typically take? Options: Yes, standard template and <2 weeks, Yes, template and 2–4 weeks, No template, negotiation varies, Unsure
• Would an upfront pilot pricing concession accelerate your internal approval, and who would authorize such a concession? Options: Yes, procurement, Yes, budget owner, No, Unsure
• What single administrative requirement would stop procurement from approving a purchase in the next quarter?

• Solution Experience, Predictive Pricing and Validation

• Confirm the current state and its cost
• You confirm the diagnosed current state and agree the quantified cost is accurate enough to prioritize a pilot.
• Deliver a back-test and lift analysis on the provided 12 months of holdout data comparing current rates to the proposed model before the follow-up meeting.
• Provide a sample rating plan, recent filings, and a 12-month holdout dataset in the agreed schema within seven business days.
• Define success criteria for models and filings
• You validate that the presented validation workflow would detect adverse selection and measure calibration against the agreed thresholds.
• Proof, segmentation and rate adequacy walkthrough
• You agree that the shown explainability artifacts meet your regulatory filing needs and that integration touchpoints are feasible for your rating engine.
• Confirm the decision criteria and gate timeline for the controlled pilot, including target metrics and acceptable calibration limits.
• Proof, validation workflow and adverse-selection checks
• Integration and regulator explainability artifacts
• Validate alignment

• Solution Experience Session
• Solution Experience Deck
• Solution Brief — Predictive Pricing & Validation

• meeting
• slides
• document

• Data Ingestion and Normalization
• Data Quality Cleansing and Reconciliation
• Feature Engineering and Variable Creation
• Claims‑to‑Policy Exposure Matching
• Loss Frequency Model Development (GLM/ML)
• Loss Severity Model Development (GLM/ML)
• Pure Premium Model Assembly and Calibration
• Model Validation and Back‑Testing
• Interpretable Model Documentation for Review
• Segment‑Level Lift Analysis and Competitive Comparison
• Policy‑Level Premium Simulation and Impact Analysis
• Rating Engine Deployment and Rule Mapping
• Underwriting Workflow Integration
• Post‑Deployment Monitoring and Performance Alerts

Data Ingestion and Normalization

• Which data extracts will you provide for modeling (select all that apply): policy master, endorsement history, earned exposures, premium transactions, claims payments, claim reserves, underwriting rules? Options: Policy master, Endorsement history, Earned exposures, Premium transactions, Claims payments, Claim reserves, Underwriting rules
• List the file formats and delivery methods available for those extracts (for example: CSV SFTP, parquet via API, database read-only replica). Options: CSV via SFTP, Parquet via API, Database read-only replica, Fixed-width files, Other
• Provide the primary keys we should use to join across datasets (for example: policy_number, policy_version_id, claim_id, transaction_id).
• Specify the historical window available for each extract in years (policy effective date to present, claim payment runout), and indicate any known gaps by year. Options: 0-2 years, 3-5 years, 6-10 years, 10+ years
• Confirm whether exposure units are available at the policy-period level (exposure hours, vehicle-years, payroll) and which column contains that measure. Options: Yes - policy-period exposure column provided, No - exposure needs to be derived, Partial coverage
• Identify any regulatory or PII masking requirements for the extracts (for example: redact named insured, partial VIN, state-specific identifiers). Options: Redact named insured, Partial VIN masking, State identifier restrictions, No masking required, Other

Data Quality Cleansing and Reconciliation

• Describe the existing reconciliation artifacts you have (for example: monthly premium roll-forward, earned premium triangles, claims paid by development month).
• Estimate the acceptable data completeness threshold for model inputs (for example: >=95% required for key fields such as coverage code, exposure, premium). Options: >=99%, >=97%, >=95%, Custom threshold
• Attach or reference the current master data definitions (column names and business meaning) for policy, exposure, and claim extracts or note if none exist. Options: Master data definitions available, Definitions incomplete, No definitions available
• Indicate the frequency you want data reconciled during engineering and pilot (weekly, monthly, one-time initial), and which balances must match statutory or general ledger totals. Options: Weekly, Monthly, One-time initial only
• Identify the owner in your organization who will resolve data mismatches (name, role, email preferred).
• Are there known data lineage issues we should plan for (for example: policies re-keyed across systems, multiple policy_number scopes, or merged legacy systems)? Options: Yes - multiple legacy systems, Yes - inconsistent policy_number usage, No known issues

Feature Engineering and Variable Creation

• Which domain features must be included or preserved in the modeling dataset (for example: ISO symbol, territory code, vehicle model year, building construction class)? Options: ISO symbol / Class code, Territory code, Vehicle model year, Building construction, Other
• Specify whether you require creation of time-varying features (for example: rolling 12-month claim counts, months-since-last-claim, tenure at current address). Options: Yes - rolling features required, No - static snapshot only, Some features only
• Provide the business rules for categorical grouping you prefer (for example: territory aggregation to 3 tiers, occupancy mapping).
• Identify any external datasets you want joined into features (for example: ZIP-level flood score, census income band, vehicle theft rate) and their expected delivery format.
• Confirm whether you require interaction terms or automated feature selection documented for actuarial peer review. Options: Yes - interactions documented, No - keep linear terms only, Prefer a mix
• Describe any governance limits on features (for example: exclude credit-score proxies, exclude certain telematics variables for regulatory reasons).

Claims‑to‑Policy Exposure Matching

• Provide the matching rules we should use to link claims to policies (for example: match on policy_number and loss_date within policy effective period, or use member_id with cross-policy matching).
• Indicate the maximum allowed runout window for matching payments to an origin period (for example: 24 months, 36 months). Options: 12 months, 24 months, 36 months, Custom
• List the priority fields to use when claims lack a policy_number (for example: insured_tax_id, vehicle_vin, address plus loss date).
• Estimate the expected match rate between claims and policy records and indicate the minimum match rate acceptable to proceed. Options: >99%, 95-99%, 90-94%, <90%
• Identify how you handle subrogation and recoveries in your claims ledger and whether recoveries should be netted in severity measures. Options: Net recoveries from severity, Keep recoveries separate, Mixed handling
• Describe any jurisdictional rules affecting matching (for example: claims reported across different state entities, multi-state policies).

Loss Frequency Model Development (GLM/ML)

• State the target frequency measure for modeling (for example: claim count per exposure unit, frequency per policy-year). Options: Claim count per exposure unit, Frequency per policy-year, Other
• Specify which link functions and distributions you prefer for GLM baselines (for example: log-link Poisson, negative binomial). Options: Log-link Poisson, Negative binomial, Zero-inflated models, No preference
• Provide the acceptable model performance thresholds for frequency models (for example: improvement in AIC, lift at deciles, or percent reduction in out-of-sample deviance).
• Indicate whether you require tree-based models (for example: gradient-boosted machines) alongside GLMs for comparison and whether variable importance must be produced. Options: Compare GLM and ML, GLM only, ML only
• Identify any regulatory constraints on model complexity (for example: maximum number of non-linear transforms or disallowed automated interactions).
• List the production input column name you expect the model score to be written to in the rating feed (for example: score_frequency_v1).

Loss Severity Model Development (GLM/ML)

• Define the severity target you want modeled (for example: incurred loss per claim, paid loss within 12 months, ultimate severity with reserve estimates). Options: Paid loss within X months, Incurred loss per claim, Ultimate severity
• Specify if you require modeling of claim-level development patterns (for example: link to chain-ladder reserving factors or explicit development modeling). Options: Yes - explicit development, No - single-period severity
• Provide the minimum claims-per-segment threshold you consider statistically reliable for separate severity modeling. Options: <100, 100-500, 500-1,000, 1,000+
• State whether expense load or claim handling costs should be included in the severity outputs for premium calculation. Options: Include expense load, Exclude expense load, Separate expense schedule
• Identify expected transformations for heavy-tailed losses you permit (for example: log-transform, winsorization at 99th percentile). Options: Log-transform, Winsorize, Pareto tail modeling, No transform
• Which claim features must remain out of model inputs for regulatory fairness concerns (for example: exclude credit proxies, exclude protected class indicators)?

Pure Premium Model Assembly and Calibration

• Confirm whether you require a single-step pure premium model or separate frequency and severity aggregation into pure premium. Options: Single-step pure premium, Separate frequency and severity then assemble, Either - recommend best practice
• Provide the calibration targets for pure premium (for example: match to aggregate earned loss ratio by line, match to statutory net loss).
• Estimate the allowed deviation from current average rate on a portfolio-level during pilot (for example: +/- 5% allowed before manual review). Options: +/- 2%, +/- 5%, +/- 10%, No bound defined
• Specify if credibility weighting to manual relativities or experience-based rate classes is required and provide the current credibility schedule if available. Options: Yes - provide schedule, No - not required, Partial
• List any statutory or filing-level adjustments that must be applied post-calibration (for example: residual trend, catastrophe load, expense constant).
• Identify the column name and format you expect for calibrated relativities to be exported for rating engine ingestion.

Model Validation and Back‑Testing

• Provide the holdout strategy you want for validation (for example: temporal holdout by policy effective date, geographic holdout by state, random k-fold) and the exact holdout window. Options: Temporal holdout, Geographic holdout, Random k-fold, Custom
• List the quantitative validation metrics you require in the delivery (for example: Gini, KS, calibration slope, lift by decile, Brier score). Options: Gini, KS, Calibration slope, Decile lift, Brier score, Other
• Specify the adverse-selection checks you want during back-testing (for example: retention by score band, conversion changes by tier, prior-notice lapse patterns).
• Provide the minimum out-of-sample performance gate that must be met for pilot acceptance (for example: at least X% improvement in decile lift vs baseline or no deterioration in calibration).
• Describe required validation artifacts for regulatory review (for example: model codebook, holdout datasets, reconciliation tables to statutory premium/claims).
• What evidence will validate the back-test and pilot as acceptable to your actuarial sign-off (for example: signed validation checklist, reconciliation report, sample scoring results on production-like data)? Options: Signed validation checklist, Reconciliation report, Sample scoring results, Formal peer review memo

Interpretable Model Documentation for Review

• Which documentation artifacts do you require for actuarial and DOI review (for example: model spec, variable engineering logs, partial dependence plots, model governance checklist)? Options: Model spec, Variable engineering log, Partial dependence plots, Governance checklist, Other
• Provide the required level of explainability per feature (for example: coefficient table for GLM, SHAP summary for ML, human-readable business description). Options: GLM coefficient table, SHAP summary, Human-readable descriptions, All of the above
• Identify the audience for documentation and their expected deliverable format (for example: state DOI PDF rate filing exhibit, internal actuarial slide deck, technical appendix with code). Options: DOI filing exhibit, Internal slide deck, Technical appendix with code, All
• Specify whether you require reproducible model code and environment artifacts (for example: notebook plus environment.yml or container image) to be delivered. Options: Notebook and environment file, Container image, No code delivery, documentation only
• State any redlines or data sensitivity rules for documentation (for example: omit raw policyholder PII from exhibits, provide masked sample rows). Options: Mask PII in exhibits, Provide masked samples, Full data allowed under NDA
• List items explicitly out of scope for documentation under a fixed-fee engagement (for example: on-site regulator presentations, detailed review custom visualizations).

Segment‑Level Lift Analysis and Competitive Comparison

• Identify the segmentation granularity you want lift reported at (for example: rating territory × ISO class, risk band decile, underwriting tier). Options: Territory × class, Decile bands, Underwriting tier, Custom
• Provide your baseline comparator for lift analysis (for example: current manual relativities, incumbent GLM, competitor rate indications). Options: Current relativities, Incumbent GLM, External comparator, Other
• Specify competitive benchmarks to include if available (for example: market average relativities by ISO symbol, publicly filed rate changes by state).
• Which lift visualizations are required for product teams and regulators (for example: decile lift tables, segment bar charts, rate change heatmaps)? Options: Decile lift tables, Segment bar charts, Rate change heatmaps, All
• Estimate the minimum segment size you consider reliable for lift claims (for example: at least N policies or M exposure-units). Options: <100 policies, 100-500 policies, 500-1,000 policies, 1,000+ policies
• Describe how competitive comparison results should be exported (for example: CSV of relativities, excel with pivot-ready tables, interactive dashboard link). Options: CSV, Excel pivot, Interactive dashboard, Other

• current_state
• success_criteria
• gaps
• desired_state
• decision_readiness
• stakeholders
• current_state
• desired_state
• decision_readiness
• success_criteria
• gaps
• stakeholders
• gaps
• stakeholders
• decision_readiness
• desired_state
• success_criteria
• current_state
• current_state
• decision_readiness
• decision_readiness
• decision_readiness
• decision_readiness

• Master Services Agreement (MSA)
• Statement of Work (SOW)
• Pricing Schedule and Order Form
• Data Access and Processing Agreement (DPA)
• Acceptance Test Plan
• Governance, Roles, and Ongoing Monitoring Agreement
• Regulatory Filing Support Addendum
• Security and Compliance Addendum
• Change Order Agreement
• Confidentiality and Non-Disclosure Agreement (NDA)
• Termination and Transition Plan

• Go-live Health Check (weeks 1-4)
• First Measurement Review (weeks 4-10)
• 90-day Performance Assessment (around day 90)
• Monthly Operational Triage
• Quarterly Business Review

• Apply agreed monitoring threshold updates in the alerting system.
• Confirm production scoring is running and model outputs are reaching the rating engine.
• List and prioritize all open operational issues with completion target dates.
• Re-confirm success criteria and owners
• If any criteria are not met, record explicit remediation actions and resolution dates.
• Confirm the incumbent model has been retired or is formally retained read-only and that fallback processes are closed.
• Publish the 90-day assessment packet showing per-criterion status and remediation timelines.
• Initiate remediation tasks for unmet criteria with target completion dates and verification checks.
• Execute the incumbent decommissioning or archive steps and provide evidence of completion.
• Monitoring dashboard and alert review
• Keep all high-severity incidents assigned with clear remediation deadlines.
• Ensure calibration error and new-business loss ratio remain within the acceptable bands defined in Model Pilot Evaluation.
• Agree which backlog items will be addressed in the next operational window.
• Create remediation tickets for each high-severity alert with target resolution dates.
• Establish the cadence and owner for the first measurement data pull.
• Schedule a targeted A/B analysis for any segment showing adverse-selection signals.
• Quarter performance summary
• Confirm whether Gini improvement and average premium change are delivering the business value expected per Model Pilot Evaluation targets.
• Agree the prioritized enhancement roadmap and monitoring improvements for the next quarter.
• Ensure regulatory documentation and explainability requests have assigned owners and delivery dates.
• Publish the quarterly performance deck with linkage to Model Pilot Evaluation targets and segment-level impact.
• Add top-priority enhancements to the next quarter engineering sprint with target delivery dates.
• Produce the regulatory explainability packet for any price-change clusters queried during the quarter.
• Publish the go-live health summary with log snippets showing successful scoring runs.
• Open remediation tickets for each high-priority blocker with target completion dates.
• Schedule the First Measurement Review and confirm the data extracts required.
• Present first measurement data
• Determine whether Gini improvement and new-business loss ratio are on track toward the targets recorded in Model Pilot Evaluation.
• Agree a prioritized remediation plan with concrete milestones to close any gaps.
• Lock the monitoring rules and data extracts for the 90-day assessment.
• Deliver the segment-level performance report used in the diagnostic within 3 business days.
• Implement agreed monitoring threshold changes in the production alerting system.
• Schedule the 90-day Performance Assessment and circulate the required data packet template.
• Restate numeric targets from Model Pilot Evaluation
• Produce a documented status for each acceptance criterion recorded in Model Pilot Evaluation (met, partially met, or unmet).
• Deployment and integration validation
• Present outcome data against each criterion
• Diagnose gaps
• Incident triage and status updates
• Regulatory and stakeholder issues
• Document criterion status and next steps
• Update monitoring thresholds and alerts
• Prioritize enhancement and bug backlog
• Early adoption and usage signals
• Risk review and adverse-selection trends
• Agree corrective actions and timeline
• Blockers and open issues
• Adjust monitoring or rollback parameters
• Incumbent model wind-down status
• Enhancement backlog and prioritization
• Confirm path to the 90-day assessment
• Governance and monitoring health
• Agree immediate remediation actions
• Agree follow-up verification steps