Underwriting intake triage: idea to MVP

THE DELIVERY PATH

Nine decisions, one path.

The case follows the work from business objective to reusable delivery system. Each section owns a different release decision.

A hand-drawn nine-part case-study map: opportunity, discovery, shaping, responsible delivery, build log, running MVP, rollout, value, and what scales. — Business value begins the path; standards and lessons are the final deliverable.

First slice: intake specialists and underwriters; eight fixed cases; two routes; rationale and audit metadata. Ambiguity goes to a human. Open the current demo →. Reviewed-reference and failure-path evidence is verified locally, and a separate capped-key capture records eight audited live Claude Opus 5 runs.

The delivery map

§	Section	Delivery responsibility mirrored
1	The opportunity brief	Business objectives and measurable value
2	Discovery and prioritization	Opportunity comparison and portfolio choice
3	Shaping the delivery	Scope, stories, acceptance, increments, backlog
4	Responsible delivery	Testing, controls, data, security, lifecycle
5	The build log	Technical partnership, dependencies, trade-offs
6	The running MVP	Observable behavior and evidence
7	Rollout and adoption	User testing, training, feedback, support
8	Value realization	Process outcomes and post-launch optimization
9	What scales	Reusable standards, templates, and lessons

1. The opportunity brief

Takeaway: automate the routing work, preserve the judgment, and measure speed and leakage together.

Intake staff check structured applications and choose a straight-through or underwriter route. Queueing a clean case creates delay; routing a material disclosure past a human creates risk.

When a complete application enters intake, recommend one of two routes and explain it in under 30 seconds, so low-complexity cases move faster and underwriters focus on cases that need judgment.

STRAIGHT_THROUGH means “continue through automated intake,” not “coverage approved.” REFER_TO_UNDERWRITER preserves human review.

Value hypothesis and success measures

Controlled routing can move capacity to complex work without missing referrals. Gross hours are eligible volume × minutes removed ÷ 60; the owner must record their use.

Targets are pilot hypotheses. Shadow mode establishes the real baseline.

Outcome	Baseline	Four-week target	Guardrail
Straight-through rate	Existing route on complete in-scope cases	At least 10 percentage points higher	Zero mandatory referrals routed straight through
Intake-to-route time	Median and 90th percentile in shadow week	Median at least 30% lower; tail no worse	Incomplete cases remain visible
Capacity	Task-level time sample	At least 40 gross hours per 1,000 cases	Record where released time goes; no double counting
Quality	Dual review of stratified sample	At least 95% rationale fidelity	One missed mandatory referral pauses assisted use
Usability	Timed UAT	At least 90% audited in under 30 seconds	Direct human route always available

In: complete structured applications, two routes, rationale, requirements, and audit metadata. Out: pricing, diagnosis, final eligibility, arbitrary documents, and free text.

Gate: process and policy owners approve route definitions, baseline method, mandatory referrals, and benefits ownership.

Use the readiness checklist →

2. Discovery and prioritization

Takeaway: the best first use case has visible value, accessible evidence, a narrow boundary, and a credible adoption path.

Use-case canvas

Lens	Current hypothesis	Evidence before pilot
Desirability	Repetitive routing delays simple cases	Observation, queue data, user interviews
Viability	Time, route, rework, and queue age are measurable	Baseline query and outcome-owner sign-off
Feasibility	Structured inputs and reviewed labels can support a narrow slice	Source sample, quality profile, integration spike
Controllable risk	Two routes, conservative referral, human gate, rollback	Policy and control-owner review
Adoption	A recommendation can fit inside the intake queue	Workflow prototype with both pilot teams

Stakeholder decisions

Stakeholder	Decision owned	Evidence or sign-off	Gate
Sponsor	Portfolio slot, outcome, scale	Opportunity and scale decision	Choose, scale
Operations owner	Workflow, baseline, capacity action	Process map, baseline, benefits sheet	Choose, release, realize
Underwriting policy delegate	Route rules, labels, quality tolerance	Signed policy, reviewed labels, release report	Prove, release
Pilot users and champions	Fit, clarity, training readiness	UAT record, training check, feedback	Shadow, assisted use
Product and delivery lead	Scope, backlog, dependencies	Plan, acceptance evidence, decision log	Every gate
Data, engineering, architecture	Source, quality, integration, service	Mapping, profile, tests, design sign-off	Shadow, production
Privacy, security, model risk	Permitted use, threat controls, lifecycle	Control evidence and sign-offs	Shadow, production
Service owner and support	Monitoring, runbook, rollback	Alert and recovery rehearsal	Assisted use, production

Portfolio choice

Scores are a discovery exercise from 1 to 5, higher being better. They expose assumptions; they are not insurer facts.

Candidate	User pain	Measurable value	Feasible slice	Controllable risk	Adoption	Total / 25
Structured intake routing	5	5	4	4	4	22
Draft a referral summary	4	3	4	3	4	18
Free-form assistant	4	3	2	2	3	14

Routing wins because its action is narrow, result observable, and safe failure clear. Stop for unavailable fields, invalid labels, duplicate work, no label owner, low volume, or a simpler deterministic rule.

Gate: the process, policy, data, and pilot-user owners validate the canvas before this use case takes the portfolio slot.

Open the canvas and scorecard →

3. Shaping the delivery

Takeaway: every story states observable behavior, a threshold, and what happens when the threshold fails.

The scope line

Now	Enterprise pilot	Never in this use case
Eight fixed synthetic records	Read-only feed from one approved queue	Coverage approval or pricing
Two route labels	Authorized override and retained evidence	Medical diagnosis
Known case IDs only	SSO, DLP, model gateway, monitoring	Arbitrary text or documents
Validated JSON and recorded fallback	More segments after evidence	Replacement of underwriter judgment

Twelve user stories with acceptance criteria

ID	User story	Testable acceptance criteria
US-01	As an intake specialist, I inspect every routing fact.	Every schema field renders; empty arrays say “None disclosed”; changing case clears results; cards work by keyboard.
US-02	As a security owner, I restrict requests to known cases.	API accepts one valid `caseId`; unknown, extra, malformed, or wrong-method requests fail before a model call; server loads data.
US-03	As a product owner, I receive one result contract.	Only the five defined fields pass enum and type validation; invalid output uses safe fallback.
US-04	As policy owner, I protect mandatory referrals.	Zero approved referral cases route straight through; at least 95% of eligible golden cases do; condition, medication, hazardous pursuit, pending test, missing field, or confidence below 0.80 refers; a miss blocks release.
US-05	As an underwriter, I audit rationale quickly.	Every bullet maps to a visible fact or versioned rule; no absent fact appears; 90% of UAT checks take under 30 seconds and rate at least 4/5.
US-06	As an underwriter, I see minimum next evidence.	Straight-through has no requirements; 95% of referral requests are domain-rated relevant; invented or irrelevant requests block that case type.
US-07	As a pilot user, I keep final accountability.	UI says “recommendation”; authorized override requires a reason and retains user, time, original and final route; public demo is non-editable.
US-08	As a model-risk reviewer, I compare versions.	Cases carry expected route and reasons; report records data, policy, prompt, model, and code; smoke set runs on every change; pilot set is larger and approved.
US-09	As a user, I get a usable failure state.	Missing key, error, timeout, invalid response, or limit returns labeled fallback within two seconds after detection; tests force each path.
US-10	As a control reviewer, I reconstruct a result.	Audit shows unique decision and case IDs, UTC time, versions, and mode; 1,000 test IDs are unique; secrets stay server-side.
US-11	As a privacy owner, I minimize data.	Records are synthetic; logs exclude applicant and rationale; secrets stay server-side; review covers egress, injection, access, dependencies, and retention.
US-12	As service owner, I stop degradation.	Monitor schema, fallback, latency, routes, overrides, quality; alert below 99.5% validity, above 2% hourly fallback, or on a missed referral; test human-only switch.

Backlog and increments

Order	Increment	Stories	Exit evidence
0	Evidence contract	US-04, US-08	Policy boundary, labels, thresholds, release report
1	Walking skeleton	US-01–03	Select case, safe server request, valid route
2	Governed demo	US-05–06, US-09–11	Traceable result, evidence, audit, privacy, fallback
3	Pilot-ready workflow	US-07, US-12	Override, monitoring, support, human-only switch
4	Scale	Expanded US-08 and enterprise controls	Approved data, integration, lifecycle evidence

Expected behavior precedes interface polish; each increment retires its riskiest assumption.

Milestone plan

Timing	Milestone	Owner	Dependencies	Exit evidence
Day 0	Opportunity	Delivery lead + operations owner	Sponsor, problem evidence, data sample	Canvas, alternative, measures, owners
Days 1–2	Governed demo	Product + engineering	Locked scope, policy, eight records, site shell	Build, contract test, fallback and smoke evidence
Two weeks before pilot	Shadow-ready	Operations + data + control owners	Source mapping, approved labels, access design	Threat review, workflow UAT, support path
End of pilot week 1	Assisted-use gate	Policy + operations owners	Clean shadow report, trained users	Thresholds met; stop switch tested
End of pilot week 4	Production gate	Sponsor + service + policy owners	Pilot quality, value, adoption, support	Approvals, monitoring, runbook, rollback, baseline
Four weeks after production	Scale gate	Sponsor + benefits owner	Stable service and realized-outcome evidence	Expand, optimize, hold, narrow, or retire

Done means the named test passes, evidence is attached, failure and accessibility are checked, and ownership is clear.

Open the story, backlog, and definition-of-done templates →

4. Responsible delivery

Takeaway: each risk needs a control, evidence, owner, threshold, and stop action.

Model-risk and data checklist

Area	Control and evidence	Owner and stop action
Boundary	Two route labels; UI and UAT confirm no coverage decision	Policy owner returns work to human-only routing
Data and privacy	Approved fields, purpose, quality, lineage, minimization, redacted logs	Data and privacy owners block the source
Referral leakage	Conservative rules, confidence gate, domain-labeled evaluation	Policy owner pauses on one confirmed miss
Rationale	Grounded prompt, fact-level review, user issue flag	Model owner blocks a version with invented facts
Security	Known IDs, server data, secret isolation, threat and dependency tests	Security revokes access and disables endpoint
Lifecycle	Code, policy, data, prompt, and model versioned as one release	Model owner rolls back to the last approved bundle
Operations	Timeout, fallback, monitoring, support runbook, human-only switch	Service owner contains, preserves evidence, recovers

The public smoke set has four straight-through, three clear referrals, and one borderline referral. Pilot data adds segments, boundaries, missing fields, interactions, and source failures. Domain reviewers label before seeing output.

The locked eight-card set stays 4/3/1. WP-6 owns separate contract fixtures for smoker=true, BMI 18.4/18.5/30.0/30.1, and sum assured CHF 1,000,000/1,000,001. They test v1 boundaries without changing the golden mix.

Release needs 100% mandatory-referral sensitivity, 95% eligible-case agreement, 100% schema validity, 95% rationale fidelity with no invented medical facts, and passed fallback, rollback, and human-only tests. Eight cases prove only the harness.

Test layers

Layer	Evidence
Unit and contract	ID validation, mapping, schema, rate limits, approved payload only
Behavioral	Route, reason themes, unsupported facts, relevant requirements by case
Adversarial	Unknown IDs, instruction-like data, missing and conflicting facts
Failure	Missing key, timeout, invalid JSON, provider error, each rate limit
UAT and operations	Audit speed, clarity, accessibility, support, alert, rollback
Security and privacy	Data flow, logs, secrets, dependencies, access, retention

Any code, policy, data, prompt, or model change reruns the gate. Monitor completeness, routes, overrides, misses, rationale, schema, latency, and fallback by version. Rollback restores human work and the last approved bundle, then needs clean evaluation and dual sign-off. Production also needs approved data use, audit evidence, training, support, baseline, and owners.

Open the risk, data, golden-set, and production-gate templates →

5. The build log

Takeaway: AI agents compress build time only when a human sets boundaries, verifies evidence, and owns release.

Date	Status	Work	Evidence
16 Aug 2026	Done	Set use case, truth constraints, scope, measures	Route intake only; no coverage decision
16 Aug 2026	Done	Created eight synthetic records and expected labels	Four straight-through, three clear referrals, one borderline referral
16 Aug 2026	Done	Shaped stories, controls, rollout, and benefits	Evidence contract before interface polish
16 Aug 2026	Done	Split content, interface, and integration into bounded paths	Shared plan; explicit integration owner
16 Aug 2026	Done locally	Connect UI, server route, canonical prompt, deterministic validation, streaming, limits, and fallback	24/24 tests; lint, production build, and dependency audit green
16 Aug 2026	Done locally	Run all eight reviewed-reference cases and exercise contract, provider-state, stream, emergency, and limit failures	8/8 missing-key HTTP SSE runs; invalid payload/MIME/method checks; 11th-request limit fallback; provider failure tests
16 Aug 2026	Done on fixed synthetic set	Capture all eight cases through the configured live Claude Opus 5 contract	8/8 exact live-provenance records; four straight-through and four referrals; model-sourced decisions and confidence

Evidence boundary: the reviewed-reference and failure-path results are local. A separate capped-key artifact records eight live-model runs from the public deployment, but neither artifact establishes production accuracy, performance on unseen cases, or generalization. This is a production-hosted demonstration, not a production underwriting service.

I direct bounded agents with acceptance criteria, inspect outputs, run checks, and integrate. Agents accelerate implementation; I own use case, policy, risk, evidence, trade-offs, and release.

Choice	Benefit	Cost
Fixed cases, no free text	Inspectable evaluation; small privacy and injection surface	No document extraction
Two route labels	Clear human boundary	No pricing or terms
Claude Opus 5	Strong structured reasoning with an explicit, versioned contract	Higher cost and latency; production comparison still required
Live-capable route plus immutable reviewed fallback	Missing-key, invalid-output, stream, emergency, and limit behavior is locally verified; a separate eight-case live capture is audited	Durable distributed limits and production evaluation remain future gates
Small synthetic set	Fast smoke test	No performance claim

Enterprise delivery adds SSO and roles, approved integration and lineage, DLP, model gateway, deterministic rules, retained audit, monitoring, continuity, control approvals, and stratified evaluation. These are future gates, not demo claims.

6. The running MVP

Takeaway: the MVP is credible only when behavior, boundary, failure mode, and audit evidence are inspectable.

Open the demo →

All applicants are synthetic. This is an educational delivery demonstration, not underwriting advice and not affiliated with any insurer.

Implementation: choose a fixed applicant; validate its ID; load data server-side; strip labels; optionally call the configured model; validate response shape and deterministic policy; replace unchecked narrative with reviewed evidence; revalidate the final result; return the live route or an immutable reference fallback.

Screenshot evidence: the capture below is a verified local reference-mode run. It shows Executed model: Not executed and must not be presented as live-model evidence.

Reference-mode demo result for synthetic case app-001, showing a straight-through recommendation, an explicit recorded-fallback notice, and session-only audit metadata.

Verified reviewed fallback and audited live capture

Case	Boundary	Expected route	Reviewed fallback	Live Opus 5 route	Live confidence
app-001 · Elena	Clear eligible	Straight-through	Straight-through	Straight-through	0.95
app-002 · Marc	Clear eligible	Straight-through	Straight-through	Straight-through	0.94
app-003 · Aisha	Clear eligible	Straight-through	Straight-through	Straight-through	0.94
app-004 · Lukas	Clear eligible	Straight-through	Straight-through	Straight-through	0.93
app-005 · Nora	Diabetes, medication, high sum	Refer	Refer	Refer	0.95
app-006 · Jonas	Private aviation	Refer	Refer	Refer	0.90
app-007 · Sofia	Pending cardiac investigation	Refer	Refer	Refer	0.94
app-008 · Maya	Borderline controlled asthma	Refer	Refer	Refer	0.90

Observed reviewed-reference contract: eight matches, four eligible straight-through routes, zero clear-referral leaks, the borderline referral preserved, and eight valid outputs. This confirms policy and fallback consistency, not model accuracy.

The separate capped-key capture records eight of eight results with exact live provenance: configured model, executed model, and result source are all claude-opus-5; mode is live; provenance is live_model; and both live flags are true. The split is four straight-through and four referrals, matching the locked expectations. Decision and confidence are model-sourced; rationale, risk factors, and requirements are canonicalized server-side from reviewed policy evidence.

The audit box shows decision and case IDs, UTC time, configured and executed model, result source, app/prompt/policy versions, and mode. It is session-only, not an underwriting record.

Evidence boundary: this is a smoke test over eight fixed synthetic cases. It does not establish production accuracy, performance on unseen applicants, subgroup behavior, calibration, or generalization. The locked labels were not changed to create a match.

7. Rollout and adoption

Takeaway: rollout is a sequence of reversible behavior changes, not an announcement.

Two teams, four weeks

Stage	Mode and behavior	Exit gate
Week 0	Name owner, two champions per team, users, support; baseline and train	Policy, data, support, and human-only switch ready
Week 1	Shadow recommendations hidden; humans route normally	Compare quality, source data, segments, and latency
Week 2	Trained users see, accept, or override; humans remain accountable	Daily safety review; zero missed mandatory referrals
Week 3	All trained pilot users; champions coach in workflow	Stable quality, rising use, no unresolved severe issue
Week 4	Bounded normal cadence	Scale, optimize, hold, narrow, or retire with evidence

Training covers boundaries, three cases, rationale checks, override, reporting, and human-only work. An observed check grants access. Communications explain change, accountability, and stop paths.

Feedback runs daily for safety, twice weekly with champions, weekly across owners, and fortnightly with sponsor and benefits owner. Overrides are learning signals.

Route policy and labels become owned artifacts. Business and control partners share evidence; each change names a metric, failure response, and owner.

L1 handles access and workflow; L2 handles service defects; L3 handles data, model, policy, and security. Exit needs quality, adoption, support, outcomes, rollback, and no critical control gap.

Open the rollout and adoption template →

8. Value realization

Takeaway: model metrics matter only when they protect a process outcome an accountable owner can measure and use.

KPI tree: baseline to target

Branch and KPI	Baseline	Pilot target	Owner
Faster service · median intake-to-route	Shadow timestamps	At least 30% lower	Operations
Released capacity · minutes per 1,000 cases	Task time sample	At least 40 gross hours	Benefits owner
Decision quality · missed mandatory referrals	Dual-reviewed sample	Zero	Policy owner
Decision quality · rationale fidelity	Reviewed bullets	At least 95%; zero invented medical facts	Model owner
Adoption · weekly active trained users	Eligible pilot roster	At least 80% by week four	Adoption lead
Usability · rationale audit time	Timed UAT	At least 90% under 30 seconds	Product owner
Service health · schema and fallback	Live events	At least 99.5% valid; under 2% hourly fallback	Service owner

The benefits sheet records definition, cohort, source, baseline, target, value, owner, confidence, capacity action, dependencies, evidence, and decision date. It does not monetize time without an action or double count it.

Post-launch: observe; diagnose policy, data, model, UI, training, or measurement; choose a fix, narrower scope, hold, or retirement; rerun gates; release reversibly; confirm process results.

Open the KPI tree and benefits-tracking template →

9. What scales

Takeaway: the next five use cases inherit a delivery system, not a copied demo.

They inherit eight editable templates: a use-case canvas that includes the opportunity brief, a prioritization scorecard, a testable AI user story, a model-risk and data checklist, a golden-set testing plan, a rollout and adoption plan, a benefits KPI tree, and delivery cadence with definition of done.

Gate	Question	Minimum evidence
Choose	Is this the right use case now?	Canvas, alternatives, score, outcome owner
Shape	Can the first slice be tested safely?	Scope, stories, data sample, controls
Prove	Does behavior meet its boundary?	Working increment, evaluation, UAT, failure tests
Release	Can users adopt, support, monitor, and stop it?	Training, runbook, rollback, baseline
Realize	Did the process improve safely?	KPI, quality, adoption, cost, capacity action

Choose the route before the model and expected cases before the prompt. Keep deterministic policy out of model judgment. Test fallback and rollback. Put human accountability in UI, tests, training, and metrics. Agents accelerate code; people own scope, policy, evidence, and release.

Get the templates in The AI Delivery Playbook →, start with the readiness checklist →, or run the eight-case demo →.