◇ Zen Insurance · Nova AI

← queue

LEAD-2026-0422

99% confidence ? · rec: refer▶ View workflow run
Runbook play: General intake— Nova recommends Review & approve
judgment: hard gate
Swimming PoolsRoof ClassElectrical SystemsPlumbingFire Protection & Defensible Space
broker tier-2judgment: hard gate53 provided · 4 service-filled · 0 to ask
playbook:
🔒 Approval locked9 items still to resolve — 5 PlanDesign fields · 4 open items. Fill each with a source + reasoning (Nova learns from every fix) to unlock approval.
Property Address
5 Birch Ln
City
Portland
State
OR
Coverage A (Dwelling)
510000
Year Built
2008
Nova AI underwriting copilotProvisions — decide, and teach me as you go

Every provision is a decision I learn from. Approve, change the value, waive, or request it from the customer — with the source and your reasoning. I capture each one to my vector memory attributed to you (a full audit trail), and I use it to make a sharper per-provision recommendation next time.

5 clean · folded
Decision-tree outcomesthe DB rule engine’s per-point verdicts (clear / refer / decline)5 points · all clear

UW Decisions · 5 clear · 0 need attention · from decision_rules (DB)

All decision points clear — nothing risky to review. Show all to see them.

Item-level approvals & request-info4 open · 0 approved · 0 unselected — bulk approve or compose the applicant ask4 items · 2 to ask
4 open · 0 approved · 0 unselected
Insured has a primary policy with Zen Insurance97%rule-safe
UW AI Agent
has_primary_policy_with_carrier = Yes
Source text
“prefill (internal)”
— autofill
P(F) - Probability of Failure (fire simulation)97%reasoned
UW AI Agent
p_f = 0.790
Source text
“prefill (catmodel)”
— autofill
Property currently listed for sale92%rule-safe
UW AI Agent
listed_for_sale = No
Source text
“prefill (mls)”
— autofill
Fire Department Response Time97%rule-safe
UW AI Agent
fire_dept_response_time = 18
Source text
“prefill (hazard)”
— autofill
Act on this lead

Nova pre-wrote a question for each missing item. For each ask, tell Nova why you need it — it learns what gets requested (and where some of it could be pulled instead of emailing the customer).

2 selected · a reason is required for each — Nova learns every ask
Auto-filled PlanDesign formthe full binding form Nova pre-populated — item-approved fields are already confirmed; edit only to correct (edits still teach Nova)9 sections · 5 missing
Auto-filled PlanDesignThe customer submitted this instance; you are editing it. Each section folds — open one, edit a field, and every change is captured with a source + your reasoning, attributed to you and versioned (audit).
3 sections shown · 5 fields missing overall
Insured3 fields · 3 missing
Insured Emailmissing
Insured Phonemissing
Insured Occupationmissing
Property1 field · 1 missing
In gated community or multi-acre propertymissing
Protection1 field · 1 missing
Opening Protectionmissing
Risk tree — first iterationour initial risk-appetite design, kept to show the iteration — superseded by provision-level approvals aligned with the runbook (▦)9 categories · all clear
Risk hierarchy
Locationclear
Accountclear
Insuredclear
Primary Coveragesclear
Propertyclear
Constructionclear
Plumbingclear
Protectionclear
Wildfireclear
Terminal state

Follow-up — one consolidated email

A guardrail capping the agent at N asks (default 3) before handing to the UW — no spamming the customer/broker.

Nova pre-wrote 2 questions for the ONE consolidated email — review, trim, and send. Sending moves this lead to Awaiting reply and the email lands in Outbound.

inbound classification: New customer