Engineering Blog · Post #110

The Clearance Agent: How InsightUW Replaces 2 Days of Manual Screening With 9 Seconds of Agentic Pipeline

Every commercial carrier runs the same gauntlet on every new submission: OFAC, duplicates, broker conflicts, capacity, territory, licensing, large limits. Done manually, it eats two days, lives in screenshots, and is unevenly applied. InsightUW's Clearance Agent runs all 11 checks in parallel, drafts a recommendation per issue, enforces role-verified sign-off, and writes a machine-readable attestation — the moment the submission lands.

TL;DR

	Manual flow	InsightUW Clearance Agent
Clearance cycle time	1–3 days	9 seconds + UW review
Screening consistency	Analyst-by-analyst	Uniform across every submission
Sanctions evidence	PDF screenshots in folders	Structured rows, 7-year retention
Disposition artifact	Free-text note	Role-verified attestation
Threshold change	Memo + retraining	Edit one row, live everywhere
DOI exam retrieval	Manual file pulls	One query
Recommendation per issue	Improvised	Calibrated narrative + ordered next steps
Authority enforcement	Honor system	Server-side role check

Part 1 — The Manual Reality

Walk into a CBP (corporate business processing) or new-business team and watch what happens when a submission lands. This is a real-world timeline for a single Excess Casualty submission worth $25M:

Time	Analyst action	Tool	Outcome
8:40 AM	Submission pulled from FIFO queue	Email + workstation	Picked up
8:42 AM	OFAC SDN search — insured	OFAC portal	14 results, eyeballed
8:48 AM	EU + UK lists — insured	Two more portals	9 more results, eyeballed
8:55 AM	Sanctions — broker (3 lists)	Same portals	12 results, eyeballed
9:00 AM	Sanctions — three principals (3 lists each)	Same portals	27 results, eyeballed
9:05 AM	Duplicate check	Queue spreadsheet	Found similar; uncertain
9:20 AM	Broker conflict check	Prior-policy table	Clear
9:25 AM	Coverage overlap	Prior-policy table	Likely renewal; email broker
9:30 AM	Broker standing	NIPR + accounting spreadsheet	Current
9:40 AM	Capacity check	Email to LOB manager	Wait for reply…
10:30 AM	Territory ambiguity (Bermuda sub)	Teams ping to compliance	Wait for reply…
11:15 AM	Large-limit review	Drop on Senior UW's desk	Wait for tomorrow…
2:15 PM (Tue)	Capacity reply arrives	Email	93% utilized — clear
Wed PM	Senior UW reviews everything	Workstation	Asks for re-verification
Thu AM	Hand-written disposition note	Free-text Notes field	"Cleared 5/27. OFAC clean. Renewal. Apex current. …"

Two days. Six tools. Four conversations. Zero machine-readable artifact.

Now multiply by 200 submissions a week.

Why the manual flow breaks at scale

Failure mode	What it looks like
Inconsistency	Maya screens principals; the analyst next to her doesn't. Tuesday's threshold is "obvious match"; Friday's is "anything plausible."
Cycle time	Capacity check waits a day. Compliance ping waits half a day. Senior-UW review waits a day. Two-day clearance is the bottleneck on a five-day quoting SLA.
Audit fragility	"Show me the sanctions evidence for March" = walking 800 folders pulling screenshots.
No recommendation layer	Every issue gets "found a possible match — escalating." Resolution is improvised. The 87th sanctions hit gets a different next step than the 86th.
Appetite drift	Threshold change lives in a memo. Half the team reads it. Six weeks later, nobody knows the rule.
Silent gaps	The skipped principal. The Bermuda subsidiary that got skimmed past. The risk in a manual flow isn't speed — it's the check that should have happened and didn't.

The manual flow isn't wrong. It works. It's just expensive, slow, structurally unauditable, and unevenly applied. It can't be the system of record for a carrier that wants to underwrite at machine speed.

Part 2 — The Agentic Shift

The naïve fix is "automate the manual flow" — wire up scripts that type into OFAC's search box, scrape the queue, send the capacity email. That's brittle. It shifts the bottleneck from human time to script maintenance, and it doesn't change the structure: still one rule at a time, still no recommendation layer, still no role-verified attestation.

The agentic framing is structurally different. The Clearance Agent doesn't replace the analyst typing into OFAC; it replaces the structure of the work.

Manual reality	Agentic shift in InsightUW
15 sequential OFAC/EU/UK searches across browser tabs	All sanctions screenings run in parallel, results structured + audited
Analyst eyeballs match plausibility	Configurable fuzzy threshold; score recorded; threshold is policy, not gut feel
Free-text disposition notes	Structured issue per finding — severity, match score, supporting data, generated narrative
Next step improvised per analyst	Ordered next-step list calibrated to issue type + match strength
Capacity email waits a day	LOB aggregate computed instantly against in-flight + bound limits
Audit lives in screenshots	Every screening, finding, resolution, acknowledgment is a row
Disposition in a Notes field	Role-verified, immutable acknowledgment
Threshold changes in a memo	Configurable rule — admin-editable, instantly enforced

Two ideas matter most:

Parallel checks + structured outputs collapse cycle time from days to seconds. But the real win is consistency. Every submission gets the same 11 checks at the same thresholds. The audit team can finally answer "are we screening uniformly?" with a query.

The recommendation layer is what makes the Agent worth using. Pure automation that says "found 3 issues, go figure it out" is no better than manual at the resolution step. An Agent that says "issue 1 is a 92% duplicate, recommend merge, here are 6 ordered next steps" is a different kind of artifact. The UW remains in control; the cognitive load drops from "what do I do" to "is this what I'd do."

Part 3 — Inside the InsightUW Clearance Agent

The Clearance Agent sits between intake (email, broker portal, policy admin systems) and the underwriter's workbench. Every submission passes through it. Every result is structured. Every acknowledgment is role-verified.

graph TD SUB[Submission arrives email / portal / API] PSN[Policy admin systems policy admin systems] Agent[InsightUW Clearance Agent] R1[11 active rules across 6 categories] Screen[Run 11 checks in parallel] AI[AI recommendation per issue] Summ[Overall risk + disposition] State{Clearance status} Clear[Clear] Issue[Issue Found] Block[Blocked] UW[Underwriter review] Comp[Compliance review] ACK[Role-verified acknowledgment] Flag[clearance cleared flag set] Down[Routing / quoting / pricing unlocked] SUB --> Agent PSN -.cross-reference.-> SUB Agent --> R1 R1 --> Screen Screen --> AI AI --> Summ Summ --> State State --> Clear State --> Issue State --> Block Clear --> ACK Issue --> UW Block --> Comp UW --> ACK Comp --> ACK ACK --> Flag Flag --> Down

The 11 default screening rules

Code	Category	Severity	Action	Fires when
SANC-001	Sanctions — insured	critical	block	Insured name ≥ 70% match on OFAC / EU / UK
SANC-002	Sanctions — broker	critical	block	Broker name ≥ 70% match on OFAC / EU / UK
SANC-003	Sanctions — principals	critical	block	Principal ≥ 70% match on OFAC / EU / UK
TERR-001	Territory restriction	critical	block	Risk location in comprehensively restricted country
LOB-001	LOB moratorium	high	block	LOB in carrier's restricted list
DUP-001	Duplicate detection	high	flag	Insured name ≥ 85% match against active queue
BRK-001	Broker conflict	high	flag	Same insured, different broker, active prior policy
COV-001	Coverage overlap	medium	flag	Same insured + same LOB already bound
BRK-002	Broker standing	high	flag	Broker not licensed / appointed / current
CAP-001	Aggregate capacity	high	escalate	LOB total exceeds carrier cap (e.g., $100M)
LIM-001	Large limit review	medium	flag	Requested limit ≥ senior-UW threshold (e.g., $25M)

Six categories. Four failure actions. Four severities. Every column admin-editable. The carrier's appetite lives in the Clearance Rules tab — versioned, enforceable, uniformly applied.

Architecture

graph TB subgraph UI[Presentation] UI1[Clearance panel] UI2[Clearance Rules admin] UI3[UW workbench] UI4[Compliance queue] end subgraph Agent[InsightUW Clearance Agent] Orch[Pipeline orchestrator] Rules[Rules engine] Screen[Screening services] Aisvc[AI recommendation] Acksvc[Acknowledgment + role check] end subgraph Data[InsightUW Data] D1[Clearance rules] D2[Per-rule results] D3[Issues + AI fields] D4[Acknowledgments] D5[Sanctions audit trail] D6[Submission queue clearance cleared flag] end subgraph EXT[External Systems] E1[Ofac / EU / UK lists] E2[Policy admin systems] E3[Broker licensing Nipr / internal] E4[Audit + reporting data integration hub] end UI1 --> Orch UI1 --> Acksvc UI2 --> Rules UI3 --> Orch UI4 --> Acksvc Orch --> Rules Orch --> Screen Orch --> Aisvc Orch --> Acksvc Rules --- D1 Screen --- D2 Screen --- D5 Aisvc --- D3 Acksvc --- D4 Orch --- D6 Screen --> E1 Screen --> E3 E2 -.submission number.-> D6 D2 --> E4 D5 --> E4 D4 --> E4

Layer	What it does
Presentation	Clearance panel, admin tab for tunable rules, workbench column + filter, compliance queue
Pipeline orchestrator	One entry point per submission; runs the pipeline; manages state transitions
Rules engine	Loads active rules, scopes to LOB, evaluates pass / fail / flag
Screening services	Actually contacts OFAC, scans the queue, computes aggregates, checks standing
AI recommendation	Per-issue narrative + ordered next steps + confidence + risk level
Acknowledgment + role check	Enforces minimum role against resolved issue set; writes immutable attestation
Data layer	Rules, per-rule results, issues, acknowledgments, sanctions trail, the cleared flag
External systems	OFAC/EU/UK feeds, policy admin systems (policy admin systems), NIPR, audit consumers

Runtime sequence

sequenceDiagram autonumber participant Intake participant Agent as Clearance Agent participant Rules as Rules Engine participant Screen as Screening Services participant Ext as External Lists participant AI as AI Recommendation participant DB as InsightUW Data participant UW as Underwriter participant Comp as Compliance participant Down as Downstream Intake->>Agent: Submission lands Agent->>DB: Status = screening Agent->>Rules: Load active rules (scoped to LOB) Rules-->>Agent: 11 rules par 11 checks run in parallel Agent->>Screen: Run all 11 screenings Screen->>Ext: Sanctions / standing lookups Ext-->>Screen: Results + scores Screen-->>Agent: Structured outputs end Agent->>DB: Persist per-rule results Agent->>DB: Materialize issues Agent->>DB: Status = clear / issue found / blocked loop For each open issue Agent->>AI: Generate recommendation AI-->>Agent: action + narrative + next steps + confidence Agent->>DB: Update issue end Agent-->>UW: Clearance panel ready alt Critical sanctions / territory hit Agent->>Comp: Route to compliance queue Comp->>Agent: Resolve + acknowledge else Non-critical issues UW->>Agent: Resolve each issue UW->>Agent: Acknowledge & sign off Agent->>Agent: Authority check vs resolved issues end Agent->>DB: Write immutable attestation Agent->>DB: clearance cleared flag set Agent->>Down: Submission visible to routing / quoting

The parallel checks (step 4) are the structural win — 11 screenings in seconds, where manual took two days. The recommendation loop (steps 9–12) turns rule results into actionable next steps. The authority check (step 17) is what makes the acknowledgment a real attestation, not a UI gesture — a UW cannot clear a sanctions match no matter how many times they click the button.

Part 4 — Worked Example: Sarah Clears Vanguard Industries

A real submission, walked through both flows side-by-side.

The submission: Vanguard Industries Corp · Excess Casualty · $25M requested · broker Apex Wholesale.

The Clearance Agent's 9 seconds

Rule	Result	Detail
SANC-001 (insured)	Pass	38% top match — well below 70% threshold
SANC-002 (broker)	Pass	41% top match
SANC-003 (principals)	Pass	29% top match across three named officers
TERR-001 (territory)	Pass	No restricted-territory mentions
LOB-001 (LOB moratorium)	Pass	Excess Casualty not restricted
DUP-001 (duplicate)	Flag	92% match against in-queue "Vanguard Industries Inc"
BRK-001 (broker conflict)	Pass	Same broker on both submissions
COV-001 (coverage overlap)	Flag	Existing bound policy expiring 2026-07-01
BRK-002 (broker standing)	Pass	Apex Wholesale: licensed, appointed, current
CAP-001 (capacity)	Pass	93% LOB utilization recorded
LIM-001 (large limit)	Flag	$25M at senior-UW threshold

Result: 8 pass · 3 flag · 0 block · Moderate risk · Proceed with caution.

The AI Recommendation layer

Issue	Recommendation	Confidence	Why
DUP-001	Merge	88%	92% match against a withdrawn submission from the same broker 2 weeks ago — likely alternate corporate alias
COV-001	Proceed	85%	Bound policy expires 2026-07-01; effective date matches — this is the renewal
LIM-001	Proceed	85%	Standard senior-UW review threshold

Each recommendation includes 5–6 ordered next steps and a contextualized narrative. The UW sees both the action and the reasoning, with the underlying evidence one click away.

Sarah's review

Time	Action
9:05 AM	Opens submission — clearance panel pre-loaded with rule checklist, AI assessment, 3 issue cards
9:08 AM	Clicks Merge on DUP-001 after one-click pivot confirms the withdrawn match
9:09 AM	Clicks Proceed on COV-001 with note: "Renewal confirmed via Apex Wholesale call"
9:10 AM	Clicks Proceed on LIM-001 with note: "Senior UW review per protocol"
9:11 AM	Clicks Acknowledge & Sign Off — server-side role check passes — immutable attestation written, clearance cleared flag set
9:12 AM	Downstream routing picks up submission — appears in Sarah's quoting tab

Manual flow vs Clearance Agent — same submission, same decision

Dimension	Manual flow	Clearance Agent
Wall-clock time	Tue 8:40 AM → Thu AM (~48 h)	8:40 AM → 9:12 AM (32 min)
Agent / system time	n/a	9 seconds
UW review time	spread across 3 days, fragmented	~7 minutes contiguous
Tools opened	6+ browser tabs / apps	One panel in InsightUW
Conversations	Email manager, Teams compliance, walk to senior UW	One broker call
Disposition artifact	Free-text Notes field	Role-verified immutable attestation
Audit retrieval	Pull screenshots from folder	One query

Part 5 — When the Agent Blocks Instead

A different submission, same morning — Sentinel Maritime Holdings · Marine · $5M · broker "Coastal Wholesale Brokerage" (never seen before).

Step	What happens
SANC-001	82% match against "Sentinel Maritime Holding" on OFAC SDN — Cuba-linked vessel registration. Severity: critical.
Status	Jumps to BLOCKED.
Routing	Diverts to Compliance Officer queue. Sarah doesn't see it on her workbench.
AI recommendation	Decline · critical risk · 96% confidence. 7 ordered next steps: EDD package, compliance review, FEIN cross-ref, OFAC license consultation, SAR consideration, retain documentation, draft formal decline if EDD fails.
Tom (Compliance)	Works the EDD. Confirms name collision — Delaware C-corp, no Cuba nexus. Resolves + acknowledges.
Authority check	Required role for sanctions match = Compliance Officer (level 4). Tom is level 4. Passes.
Status	blocked → issue found → pending acknowledgment → cleared.
Sarah	Picks up submission only after the cleared flag flips.

The structured handoff to compliance is automatic. In the manual flow, this is the kind of thing that gets skipped under volume pressure — Sarah might have started pricing before realizing there was a sanctions issue. The Clearance Agent doesn't rely on the analyst remembering to escalate.

Part 6 — Outcomes

Metric	Before InsightUW	With Clearance Agent
Average clearance cycle time	1–3 days	< 1 hour (UW review only)
Agent / system time	n/a	9 seconds
Submissions screened per analyst per day	15–25	Throughput capped by UW review, not screening
Screening consistency across analysts	Variable	Uniform
OFAC / sanctions screening coverage	Insured always; broker / principals sometimes	All three entities, every submission
DOI exam evidence retrieval	Days of folder pulls	Single query, ranged by date
Sanctions retention policy compliance	Best-effort folder retention	Automatic 7-year structured retention
Threshold change deployment	Memo + retraining + drift	One row edit, live everywhere
Audit completeness for cleared submissions	~60–80% (varies by analyst)	100% (machine-enforced)
Authority-violation incidents	Recurring (UW clears sanctions issues)	Zero (server-side role enforcement)

Part 7 — What's Deferred

Capability	Status
Streaming sanctions list updates (real-time list deltas)	Deferred — daily refresh today
Per-carrier rule overrides on a multi-tenant instance	Schema supports it; admin UI deferred
AI-extracted principals from submission documents	Stub in place; integration deferred
Threshold drift simulator ("replay last 90 days at 65%")	Replay supported; UI wrapping deferred
Cross-LOB combined aggregate capacity	New rule type; deferred
PEP screening (Dow Jones / WorldCheck)	Separate rule type; deferred until feed contract

Part 8 — What This Means for Carriers

Clearance is the gate every submission must pass. Doing it manually doesn't make it more rigorous — it makes it slower, less consistent, and harder to audit.

	Takeaway
1	The manual flow is the bottleneck, not the screening logic. Carriers running clearance manually aren't doing it wrong; they're doing it slow and inconsistently because nothing stitches the tools together.
2	Two layers stay separated — deterministic rules answer "did this hit a 70%+ OFAC match?", the AI layer answers "what should the UW do about it?". The AI never acts autonomously.
3	Rules are admin-tunable without a deploy. Severity, threshold, LOB scope, failure action, required role — all editable. Appetite changes ship in minutes.
4	Role-verified acknowledgment is server-enforced. A UW cannot clear a sanctions match no matter how many times they click the button.
5	clearance cleared at is the contract. Every downstream service — routing, quoting, pricing, binder — gates on this single flag.
6	Policy-system numbers thread through everything. policy admin systems submission numbers appear in the badge, the rationale text, the attestation, the audit trail. One submission, two systems, traceable both ways.
7	Audit is a side effect, not a project. Every screening, recommendation, acknowledgment is a row. 7-year retention is automatic.
8	Confidence is calibrated, not asserted. An 82% sanctions hit reports higher AI confidence than a 71% one. No false certainty.
9	Severity drives action, not the rule code. Critical sanctions blocks. High broker conflict flags. Medium overlap flags. The disposition disambiguates.
10	Two days → 9 seconds — and the audit gets stronger, not weaker. The Agent doesn't just speed up clearance; it produces an artifact the manual flow can't.

What's Next

The Clearance Agent is one of several AI Agents inside InsightUW. The next blog walks through the Routing Agent — which picks up where Clearance leaves off (clearance_cleared_at IS NOT NULL) and decides whose desk the submission lands on, using the same admin-tunable rules pattern applied to a different decision space.

Want to see how InsightUW's Clearance Agent runs 11 screening checks the moment a submission lands — and produces a role-verified attestation in seconds instead of days? Request a demo.

A "Day in the life" infographic — built off the Part 1 timeline — works on all three channels.